diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 9bface50..3e547232 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -161,6 +161,11 @@ jobs:
     steps:
       - uses: actions/checkout@v5
 
+      - uses: actions/setup-node@v7
+        with:
+          node-version: "24"
+          package-manager-cache: false
+
       - uses: actions/setup-python@v6
         with:
           python-version: "3.11"
@@ -181,6 +186,9 @@ jobs:
       - name: Import-origin guard
         run: python3 tools/check_import_origin.py --strict
 
+      - name: Browser wizard regression tests
+        run: node --test tests/site/*.test.cjs
+
       - name: CLI smoke checks
         run: |
           forgelm --version
diff --git a/AGENTS.md b/AGENTS.md
index 629b12a7..d3b0470b 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -166,6 +166,7 @@ Default workflow for a non-trivial change:
    ```bash
    python3 tools/check_import_origin.py --strict && \
      ruff format . && ruff check . && pytest tests/ && \
+     node --test tests/site/*.test.cjs && \
      python3 -m forgelm --config config_template.yaml --dry-run && \
      python3 -m mypy --strict --follow-imports=silent forgelm/__init__.py forgelm/_version.py tests/typing/public_surface_probe.py && \
      python3 tools/check_field_descriptions.py --strict forgelm/config.py && \
diff --git a/CHANGELOG.md b/CHANGELOG.md
index febf460c..818f6667 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,8 +4,36 @@ All notable changes to ForgeLM are documented here.
 
 ## [Unreleased]
 
+### Breaking
+
+- Tighten `monitoring.endpoint_env` to trimmed ASCII identifiers; URLs, `$`/`env:` prefixes,
+  internal whitespace, and non-ASCII names now fail validation. Blank names remain absent.
+  This changes acceptance of previously valid YAML; use a variable name, such as `MONITORING_PLAN_URL`.
+- Require non-blank `monitoring.endpoint` literals to be absolute HTTP(S) URLs with a host
+  and valid port; schemeless addresses are rejected instead of silently becoming `<malformed-url>`
+  in evidence. Use `https://monitor.example.com:8443/plan` or an `endpoint_env` name (#93).
+
 ### Changed
 
+- Record monitoring default provenance in `defaulted_fields` and mark literal URL projection with
+  `endpoint_redacted` (also true when the URL already contains only scheme/host/port). Literal
+  endpoints retain scheme/host/port in JSON evidence, including IPv6 brackets; userinfo, path,
+  query and fragment are removed. Env references remain unresolved. Absent endpoints record `null`
+  instead of the legacy `"$None"`, and non-empty `endpoint_env` takes precedence.
+  Monitoring metadata is exported to `compliance_report.json` and, when emitted,
+  `annex_iv_metadata.json`; the flattened `training_manifest.yaml` summary does not contain it.
+  Monitoring declarations remain optional in ForgeLM at every risk tier, including strict tiers;
+  providers’ Article 72 obligations apply independently. Treat legacy monitoring blocks without
+  `declarative_only` as declarations; their default provenance was not recorded (#93).
+- Default CLI/browser wizard drift-alert answers to No / unchecked; only declare alerts when
+  the operator confirms them. Preserve the schema's `true` default with `defaulted_fields`
+  provenance. Both wizards explicitly record the selected `none` exporter (#93).
+- Accept case-insensitive `env:` references with surrounding whitespace, lowercase letters and
+  initial underscores in both CLI wizard monitoring and webhook inputs; reject malformed names.
+- Omit Pydantic `input_value` / `input_type` excerpts from all rendered `ForgeConfig` validation
+  errors, preserving field paths and reasons. `MonitoringConfig` also applies this protection
+  standalone; other standalone submodels retain their existing diagnostics (#93).
+
 - Redesign the root README around a GPU-free first run, a concise capability overview,
   optional evidence controls, and clear contribution and documentation paths. Reuse the
   site's ForgeLM mark and distinguish the Apache-licensed code from CC-BY-SA seed datasets.
@@ -15,6 +43,15 @@ All notable changes to ForgeLM are documented here.
 
 ### Fixed
 
+- Restore the monitoring-plan evidence contract: `monitoring.*` declares an external plan,
+  labelled `declarative_only: true`; ForgeLM runs no exporter, drift detector, alert, or scheduler.
+  Warn once per config-file load for an enabled exporter declaration, without repeating during
+  pipeline stage validation. Align CLI/web wizards and EN/TR guidance, correct Annex IV monitoring
+  point 9 / risk-management point 5 labels, and verify exported declarations end to end (#93).
+- Correct webhook troubleshooting in the site and EN/TR manuals: delivery failures appear as
+  `forgelm.webhook` warnings, never `webhook_failed` audit events; use the actual
+  `allow_private_destinations` setting (#231). Clarify that `--offline` sets HF offline flags;
+  reporting, API judges, and webhooks require separate disconnected-run configuration (refs #100).
 - Rework the homepage terminal into a first-run guide with separately copyable installation and
   config-generation commands, an optional training step, and readable code in both themes and on
   narrow screens, with interface copy in all six site languages. Keep the dated CPU recording and
diff --git a/CLAUDE.md b/CLAUDE.md
index e293ab9c..67197574 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -166,6 +166,7 @@ Default workflow for a non-trivial change:
    ```bash
    python3 tools/check_import_origin.py --strict && \
      ruff format . && ruff check . && pytest tests/ && \
+     node --test tests/site/*.test.cjs && \
      python3 -m forgelm --config config_template.yaml --dry-run && \
      python3 -m mypy --strict --follow-imports=silent forgelm/__init__.py forgelm/_version.py tests/typing/public_surface_probe.py && \
      python3 tools/check_field_descriptions.py --strict forgelm/config.py && \
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 900a46a0..dbf0da6a 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -165,6 +165,7 @@ supported Python matrix:
 ```bash
 python3 tools/check_import_origin.py --strict && \
   ruff format . && ruff check . && pytest tests/ && \
+  node --test tests/site/*.test.cjs && \
   python3 -m forgelm --config config_template.yaml --dry-run && \
   python3 -m mypy --strict --follow-imports=silent forgelm/__init__.py forgelm/_version.py tests/typing/public_surface_probe.py && \
   python3 tools/check_field_descriptions.py --strict forgelm/config.py && \
diff --git a/config_template.yaml b/config_template.yaml
index ce08c2d6..8aec1398 100644
--- a/config_template.yaml
+++ b/config_template.yaml
@@ -256,11 +256,15 @@ data:
 #     - "Human review required for policy-related responses"
 #   vulnerable_groups_considered: true
 
-# Post-Market Monitoring (Art. 12+17)
+# External post-market monitoring plan (Art. 72 / 17(1)(h)) — declarative metadata only.
+# Recorded in the compliance manifest / Annex IV; ForgeLM does not contact the endpoint,
+# export metrics, detect drift, send drift alerts, or schedule checks. Operate your own stack.
+# Uncommented values are explicit declarations about your external stack; review them before use.
 # monitoring:
 #   enabled: true
-#   endpoint_env: "MONITORING_WEBHOOK_URL"
-#   metrics_export: "prometheus"  # "none", "prometheus", "datadog", "custom_webhook"
+#   endpoint: ""  # optional literal URL; evidence keeps scheme/host/port only; must be http(s)://host
+#   endpoint_env: "MONITORING_PLAN_URL"  # recorded as "$NAME", never read; overrides endpoint
+#   metrics_export: "none"  # other exporters emit a metadata-only warning at config load
 #   alert_on_drift: true
 #   check_interval_hours: 24
 
diff --git a/docs/design/iso27001_soc2_alignment.md b/docs/design/iso27001_soc2_alignment.md
index 49483942..19003e2f 100644
--- a/docs/design/iso27001_soc2_alignment.md
+++ b/docs/design/iso27001_soc2_alignment.md
@@ -157,7 +157,7 @@ Each row classifies coverage as one of:
 | A.5.5 Contact with authorities | OOS | — | Establish regulator contact for serious-incident reporting (EU AI Act Art. 73) | ForgeLM cannot interface with regulators |
 | A.5.6 Contact with special interest groups | OOS | — | ML safety / red-team community engagement | Out of scope |
 | A.5.7 Threat intelligence | OOS | — | Subscribe to ML supply-chain CVE feeds, model-poisoning advisories | Out of scope |
-| A.5.8 Information security in project management | FL-helps | `risk_assessment` config block; F-compliance-110 strict gate; Annex IV §9 metadata | Embed ForgeLM dry-run + safety-eval gates in project sign-off checklist | Project methodology deployer-side |
+| A.5.8 Information security in project management | FL-helps | `risk_assessment` config block; F-compliance-110 strict gate; Annex IV point 5 metadata | Embed ForgeLM dry-run + safety-eval gates in project sign-off checklist | Project methodology deployer-side |
 | A.5.9 Inventory of information & associated assets | FL-helps | `data_provenance.json`, `model_integrity.json`, SBOM | Maintain corporate asset register linking model SHA + data fingerprint | Org asset registry deployer-side |
 | A.5.10 Acceptable use of information | OOS | — | Acceptable-use policy for trained-model outputs | Out of scope |
 | A.5.11 Return of assets | OOS | — | Decommissioning of deployed models | Out of scope |
@@ -244,7 +244,7 @@ ForgeLM contributes nothing here.
 | A.8.13 Information backup | FL-helps | `audit_log.jsonl` + `.manifest.json` form a backupable unit; Annex IV bundle (`forgelm --compliance-export`) | Off-site backup with encryption |
 | A.8.14 Redundancy of information processing facilities | OOS | — | Multi-AZ infra |
 | A.8.15 Logging | FL | `audit_log.jsonl` JSON Lines + HMAC chain + genesis manifest; `forgelm verify-audit` validates | Ship to SIEM |
-| A.8.16 Monitoring activities | FL-helps | Webhook lifecycle events; `safety_trend.jsonl` cross-run trend tracking | Define alert thresholds |
+| A.8.16 Monitoring activities | FL-helps | Webhook lifecycle notifications and `safety_trend.jsonl` evaluation trends; external monitoring-plan metadata only | Operate monitoring and define alert thresholds |
 | A.8.17 Clock synchronisation | FL-helps | All audit entries ISO-8601 UTC | NTP on training hosts |
 | A.8.18 Use of privileged utility programs | OOS | — | OS-level privilege management |
 | A.8.19 Installation of software on operational systems | FL-helps | `forgelm doctor` reports installed packages + versions; `pyproject.toml` pins; `pip install forgelm==X.Y.Z` | Package allowlist |
diff --git a/docs/guides/quickstart-tr.md b/docs/guides/quickstart-tr.md
index ed185a57..43b6ab2a 100644
--- a/docs/guides/quickstart-tr.md
+++ b/docs/guides/quickstart-tr.md
@@ -72,7 +72,7 @@ Bundled seed dataset'lerin lisansları için bkz. [LICENSES.md](https://github.c
 forgelm --wizard
 ```
 
-Sihirbaz önce curated quickstart-template kısayolu önerir; reddedilirse 9 adımlı etkileşimli akış açılır (welcome / use-case / model / strategy / trainer / dataset / training-params / compliance / operations) ve her `ForgeConfig` bloğunu kapsar — model, LoRA / DoRA / PiSSA / rsLoRA / GaLore stratejisi, trainer-spesifik hyperparam'lar (`dpo_beta` / `simpo_*` / `kto_beta` / `orpo_beta` / `grpo_*`), EU AI Act Madde 9 / 10 / 11 / 12+17 uyumluluk metadata, retention, monitoring, evaluation kapıları, webhook, sentetik veri — ve kullanıma-hazır bir YAML yazar. Geri dönmek için `back` / `b`, sıfırlamak için `reset` / `r`; state `~/.cache/forgelm/wizard_state.yaml`'da persistent, Ctrl-C / yeni oturum kaldığı yerden devam edebilir.
+Sihirbaz önce curated quickstart-template kısayolu önerir; reddedilirse 9 adımlı etkileşimli akış açılır (welcome / use-case / model / strategy / trainer / dataset / training-params / compliance / operations) ve her `ForgeConfig` bloğunu kapsar — model, LoRA / DoRA / PiSSA / rsLoRA / GaLore stratejisi, trainer-spesifik hyperparam'lar (`dpo_beta` / `simpo_*` / `kto_beta` / `orpo_beta` / `grpo_*`), EU AI Act Madde 9 / 10 / 11 / 12 / 17(1)(h) / 72 uyumluluk metadata, retention, monitoring, evaluation kapıları, webhook, sentetik veri — ve kullanıma-hazır bir YAML yazar. Geri dönmek için `back` / `b`, sıfırlamak için `reset` / `r`; state `~/.cache/forgelm/wizard_state.yaml`'da persistent, Ctrl-C / yeni oturum kaldığı yerden devam edebilir.
 
 Review-cycle 2 (2026-05-09) ile eklenen operatör guardrail'ları: sihirbaz çıkmadan önce kayıtlı YAML'ı `ForgeConfig.model_validate` ile doğrular (schema reddetmeleri 30 dakika sonra eğitim sırasında değil, anında yüzeylenir), mevcut config dosyasını üzerine yazmadan önce sorar (reddederseniz `_2.yaml` / `_3.yaml` ile auto-suffix'ler), non-tty stdin altında çalışmayı reddeder (scriptli koşumlar için `forgelm quickstart <template>` kullanın), pre-flight checklist (GPU/VRAM/dataset/risk-tier sinyalleri) basar ve Ctrl-C / cancel durumunda `EXIT_WIZARD_CANCELLED = 5` ile çıkar — böylece CI "sihirbaz tamamlandı" ile "sihirbaz hiçbir şey yazmadı"yı ayırt edebilir.
 
diff --git a/docs/guides/quickstart.md b/docs/guides/quickstart.md
index 3d2d23ac..d91cb21d 100644
--- a/docs/guides/quickstart.md
+++ b/docs/guides/quickstart.md
@@ -67,7 +67,7 @@ See [LICENSES.md](https://github.com/HodeTech/ForgeLM/blob/main/forgelm/template
 forgelm --wizard
 ```
 
-The wizard offers a curated quickstart-template shortcut first; declining opens a 9-step interactive flow (welcome / use-case / model / strategy / trainer / dataset / training-params / compliance / operations) that covers every `ForgeConfig` block — model, LoRA / DoRA / PiSSA / rsLoRA / GaLore strategy, per-trainer hyperparameters (`dpo_beta` / `simpo_*` / `kto_beta` / `orpo_beta` / `grpo_*`), EU AI Act Article 9 / 10 / 11 / 12+17 compliance metadata, retention, monitoring, evaluation gates, webhooks, synthetic data — and writes a ready-to-use YAML. Type `back` / `b` to navigate backwards, `reset` / `r` to start over; state is persisted to `~/.cache/forgelm/wizard_state.yaml` so a Ctrl-C / fresh session can resume.
+The wizard offers a curated quickstart-template shortcut first; declining opens a 9-step interactive flow (welcome / use-case / model / strategy / trainer / dataset / training-params / compliance / operations) that covers every `ForgeConfig` block — model, LoRA / DoRA / PiSSA / rsLoRA / GaLore strategy, per-trainer hyperparameters (`dpo_beta` / `simpo_*` / `kto_beta` / `orpo_beta` / `grpo_*`), EU AI Act Article 9 / 10 / 11 / 12 / 17(1)(h) / 72 compliance metadata, retention, monitoring, evaluation gates, webhooks, synthetic data — and writes a ready-to-use YAML. Type `back` / `b` to navigate backwards, `reset` / `r` to start over; state is persisted to `~/.cache/forgelm/wizard_state.yaml` so a Ctrl-C / fresh session can resume.
 
 Operator guardrails layered on by review-cycle 2 (2026-05-09): the wizard runs `ForgeConfig.model_validate` on the saved YAML before exit (so schema rejections surface inline, not 30 minutes into training), prompts before overwriting an existing config (auto-suffixes `_2.yaml` / `_3.yaml` if you decline), refuses to launch under non-tty stdin (use `forgelm quickstart <template>` for scripted runs), prints a pre-flight checklist (GPU/VRAM/dataset/risk-tier signals), and exits `EXIT_WIZARD_CANCELLED = 5` on Ctrl-C / cancel so CI can tell "wizard finished" from "wizard never wrote anything".
 
diff --git a/docs/guides/safety_compliance-tr.md b/docs/guides/safety_compliance-tr.md
index 9ea817fe..eae6ef2c 100644
--- a/docs/guides/safety_compliance-tr.md
+++ b/docs/guides/safety_compliance-tr.md
@@ -302,7 +302,7 @@ Dağıtımınıza uygulanacak kuralları yetkin bir hukuk danışmanıyla değer
 - Makine-okunabilir uyumluluk kanıtı
 - Eğitim verisi kaynağı (provenance) takibi
 - Risk sınıflandırması
-- Sürekli izleme
+- Pazar sonrası izleme sistemini işletin; ForgeLM yalnızca harici plan metadata’sını kaydeder
 
 ### Uyumluluk artefaktları
 
diff --git a/docs/guides/safety_compliance.md b/docs/guides/safety_compliance.md
index 4fc5dc92..5569369c 100644
--- a/docs/guides/safety_compliance.md
+++ b/docs/guides/safety_compliance.md
@@ -269,7 +269,7 @@ Have qualified legal counsel assess the rules for your deployment. Documentation
 - Machine-readable compliance evidence
 - Training data provenance tracking
 - Risk classification
-- Continuous monitoring
+- Operate a post-market monitoring system; ForgeLM records external-plan metadata only
 
 ### Compliance Artifacts
 
diff --git a/docs/guides/troubleshooting-tr.md b/docs/guides/troubleshooting-tr.md
index ac286edd..c4cb2408 100644
--- a/docs/guides/troubleshooting-tr.md
+++ b/docs/guides/troubleshooting-tr.md
@@ -238,13 +238,15 @@ forgelm --config my_config.yaml --dry-run
 
 ### Config doğrulama hatası: Field type uyumsuzluğu
 
-ForgeLM doğrulama için Pydantic v2 kullanır. Hata mesajları tam
-field'ı gösterir:
+ForgeLM doğrulama için Pydantic v2 kullanır. Tüm `ForgeConfig` doğrulama hatalarının metin çıktısında Pydantic'in
+`input_value` / `input_type` alıntıları gizlenir; alan yolları ve hata nedenleri korunur. `MonitoringConfig` dışındaki
+alt modellerin tek başına kurulumu (örneğin `WebhookConfig` veya `AuthConfig`) mevcut Pydantic hata çıktısını korur.
+Hata mesajları tam alanı gösterir:
 
 ```text
 Configuration validation failed: 1 validation error for ForgeConfig
-training -> learning_rate
-  Input should be a valid number [type=float_parsing, input_value='not_a_number']
+training.learning_rate
+  Input should be a valid number, unable to parse string as a number [type=float_parsing]
 ```
 
 YAML değerini beklenen tipe uyacak şekilde düzeltin.
@@ -258,7 +260,7 @@ Typo'lar ya da desteklenmeyen alanlar net bir hata atar:
 ```text
 ConfigError: Configuration validation failed: 1 validation error for ForgeConfig
 training.lerning_rate
-  Extra inputs are not permitted [type=extra_forbidden, input_value=2e-5]
+  Extra inputs are not permitted [type=extra_forbidden]
 ```
 
 Bu kasıtlıdır: sessiz typo'lar (örn. `learning_rate` yerine
diff --git a/docs/guides/troubleshooting.md b/docs/guides/troubleshooting.md
index dde3d0db..7533eedb 100644
--- a/docs/guides/troubleshooting.md
+++ b/docs/guides/troubleshooting.md
@@ -228,12 +228,15 @@ forgelm --config my_config.yaml --dry-run
 
 ### Config Validation Error: Field Type Mismatch
 
-ForgeLM uses Pydantic v2 for validation. Error messages show the exact field:
+ForgeLM uses Pydantic v2 for validation. All rendered `ForgeConfig` validation errors omit Pydantic's `input_value` /
+`input_type` excerpts, while preserving field paths and reasons. Standalone submodels other than `MonitoringConfig`
+(for example, `WebhookConfig` or `AuthConfig`) retain their existing Pydantic diagnostics. Error messages show the
+exact field:
 
 ```text
 Configuration validation failed: 1 validation error for ForgeConfig
-training -> learning_rate
-  Input should be a valid number [type=float_parsing, input_value='not_a_number']
+training.learning_rate
+  Input should be a valid number, unable to parse string as a number [type=float_parsing]
 ```
 
 Fix the YAML value to match the expected type.
@@ -245,7 +248,7 @@ ForgeLM now **rejects unknown fields** in YAML configs — all sub-models enforc
 ```text
 ConfigError: Configuration validation failed: 1 validation error for ForgeConfig
 training.lerning_rate
-  Extra inputs are not permitted [type=extra_forbidden, input_value=2e-5]
+  Extra inputs are not permitted [type=extra_forbidden]
 ```
 
 This is intentional: silent typos (like `lerning_rate` instead of `learning_rate`) previously caused training to run with wrong defaults. Now they fail fast with a clear message.
diff --git a/docs/qms/sop_incident_response-tr.md b/docs/qms/sop_incident_response-tr.md
index d178022d..1c2b0f8a 100644
--- a/docs/qms/sop_incident_response-tr.md
+++ b/docs/qms/sop_incident_response-tr.md
@@ -28,7 +28,8 @@ yanında kapsar.
 ### 3.1 Tespit
 
 Olaylar şunlardan tespit edilebilir:
-- Runtime izleme alarmları (`monitoring.alert_on_drift: true` ise)
+- Harici servis/izleme altyapınızdan gelen runtime uyarıları (`monitoring.*` uyumluluk manifestine
+  yalnızca bir plan kaydeder; ForgeLM drift tespit etmez veya izleme uyarısı göndermez)
 - Kullanıcı/operatör raporları
 - Periyodik kalite denetimleri
 - ForgeLM webhook arıza bildirimleri
diff --git a/docs/qms/sop_incident_response.md b/docs/qms/sop_incident_response.md
index 4ccd2aff..bf322545 100644
--- a/docs/qms/sop_incident_response.md
+++ b/docs/qms/sop_incident_response.md
@@ -28,7 +28,8 @@ existing AI-safety incident flow.
 ### 3.1 Detection
 
 Incidents may be detected by:
-- Runtime monitoring alerts (if `monitoring.alert_on_drift: true`)
+- Runtime alerts from your external serving/monitoring stack (`monitoring.*` only records a plan
+  in the compliance manifest; ForgeLM does not detect drift or send monitoring alerts)
 - User/deployer reports
 - Periodic quality audits
 - ForgeLM webhook failure notifications
diff --git a/docs/qms/statement_of_applicability-tr.md b/docs/qms/statement_of_applicability-tr.md
index 5813aca5..6bc99bf0 100644
--- a/docs/qms/statement_of_applicability-tr.md
+++ b/docs/qms/statement_of_applicability-tr.md
@@ -35,7 +35,7 @@ operatör-tarafı eylem.
 | A.5.5 Yetkililerle iletişim | YES | EU AI Act Md. 73 ciddi-olay raporlama | Düzenleyici contact listesi tut |
 | A.5.6 Özel ilgi gruplarıyla iletişim | YES | ML safety / red-team topluluğu | İlgili threat intel'e abone ol |
 | A.5.7 Tehdit istihbaratı | YES | ML supply-chain CVE feed'leri, model-poisoning advisory'leri | İlgili feed'lere abone ol |
-| A.5.8 Proje yönetiminde bilgi güvenliği | YES | `risk_assessment` config + F-compliance-110 strict gate; Annex IV §9 metadata | Proje sign-off'unda göm |
+| A.5.8 Proje yönetiminde bilgi güvenliği | YES | `risk_assessment` config + F-compliance-110 strict gate; Annex IV point 5 metadata | Proje sign-off'unda göm |
 | A.5.9 Bilgi ve diğer ilişkili varlıkların envanteri | YES | `data_provenance.json`, `model_integrity.json`, SBOM | Kurumsal asset register tut |
 | A.5.10 Bilginin kabul edilebilir kullanımı | YES | Standart kurumsal AUP | Benimse |
 | A.5.11 Varlıkların iadesi | YES | Dağıtılan modellerin / eğitim host'larının decommissioning'i | Benimse |
diff --git a/docs/qms/statement_of_applicability.md b/docs/qms/statement_of_applicability.md
index 3d2219b2..e19021dd 100644
--- a/docs/qms/statement_of_applicability.md
+++ b/docs/qms/statement_of_applicability.md
@@ -34,7 +34,7 @@ contribution → deployer-side action.
 | A.5.5 Contact with authorities | YES | EU AI Act Art. 73 serious-incident reporting | Maintain regulator contact list |
 | A.5.6 Contact with special interest groups | YES | ML safety / red-team community | Subscribe to relevant threat intel |
 | A.5.7 Threat intelligence | YES | ML supply-chain CVE feeds, model-poisoning advisories | Subscribe to relevant feeds |
-| A.5.8 Information security in project management | YES | `risk_assessment` config + F-compliance-110 strict gate; Annex IV §9 metadata | Embed in project sign-off |
+| A.5.8 Information security in project management | YES | `risk_assessment` config + F-compliance-110 strict gate; Annex IV point 5 metadata | Embed in project sign-off |
 | A.5.9 Inventory of information and other associated assets | YES | `data_provenance.json`, `model_integrity.json`, SBOM | Maintain corporate asset register |
 | A.5.10 Acceptable use of information | YES | Standard corporate AUP | Adopt |
 | A.5.11 Return of assets | YES | Decommissioning of deployed models / training hosts | Adopt |
diff --git a/docs/reference/architecture-tr.md b/docs/reference/architecture-tr.md
index d3c4be7f..76978c09 100644
--- a/docs/reference/architecture-tr.md
+++ b/docs/reference/architecture-tr.md
@@ -167,7 +167,7 @@ HuggingFace-uyumlu README.md üretir: YAML front-matter, eğitim parametreleri t
 
 ### `wizard/`
 
-Etkileşimli CLI sihirbazı. Faz 22 modernizasyonu (2026-05-08) ile `site/js/wizard.js`'e parite getirildi: 9-adımlı state machine (welcome / use-case / model / strategy / trainer / dataset / training-params / compliance / operations), trainer-spesifik hyperparam'lar (`dpo_beta` / `simpo_beta` + `simpo_gamma` / `kto_beta` / `orpo_beta` / `grpo_*`), tam PEFT method kapsamı (`lora` / `dora` / `pissa` / `rslora`) + GaLore ayrı axis, EU AI Act Madde 9 + 10 + 11 + 12+17 uyumluluk akordeon'ları, F-compliance-110 strict-tier auto-coercion, `back` / `reset` navigasyon, XDG-aware persistence (`$XDG_CACHE_HOME/forgelm/wizard_state.yaml`), step-diff preview, beginner / expert toggle ve Phase 11.5 / 12.5 BYOD inline ingest + audit helper'ları (`_offer_ingest_for_directory`, `_offer_audit_for_jsonl`).
+Etkileşimli CLI sihirbazı. Faz 22 modernizasyonu (2026-05-08) ile `site/js/wizard.js`'e parite getirildi: 9-adımlı state machine (welcome / use-case / model / strategy / trainer / dataset / training-params / compliance / operations), trainer-spesifik hyperparam'lar (`dpo_beta` / `simpo_beta` + `simpo_gamma` / `kto_beta` / `orpo_beta` / `grpo_*`), tam PEFT method kapsamı (`lora` / `dora` / `pissa` / `rslora`) + GaLore ayrı axis, EU AI Act Madde 9 + 10 + 11 + 12 / 17(1)(h) / 72 uyumluluk akordeon'ları, F-compliance-110 strict-tier auto-coercion, `back` / `reset` navigasyon, XDG-aware persistence (`$XDG_CACHE_HOME/forgelm/wizard_state.yaml`), step-diff preview, beginner / expert toggle ve Phase 11.5 / 12.5 BYOD inline ingest + audit helper'ları (`_offer_ingest_for_directory`, `_offer_audit_for_jsonl`).
 
 ### `webhook.py`
 Eğitim başlangıcı/başarı/başarısızlık durumlarında yapılandırılmış JSON payload gönderir. Yapılandırılabilir timeout ile güvenilir hata yönetimi.
diff --git a/docs/reference/architecture.md b/docs/reference/architecture.md
index 72d01e32..cda21d17 100644
--- a/docs/reference/architecture.md
+++ b/docs/reference/architecture.md
@@ -169,7 +169,7 @@ Synthetic data generation via teacher-to-student distillation. The `SyntheticDat
 
 ### `wizard/`
 
-Interactive CLI wizard for generating valid YAML configs. Phase 22 modernisation (2026-05-08) brought the CLI to parity with `site/js/wizard.js`: 9-step state machine (welcome / use-case / model / strategy / trainer / dataset / training-params / compliance / operations), per-trainer hyperparameters (`dpo_beta` / `simpo_beta` + `simpo_gamma` / `kto_beta` / `orpo_beta` / `grpo_*`), full PEFT method coverage (`lora` / `dora` / `pissa` / `rslora`) plus GaLore axis, EU AI Act Article 9 + 10 + 11 + 12+17 compliance accordions, F-compliance-110 strict-tier auto-coercion, `back` / `reset` navigation, XDG-aware persistence at `$XDG_CACHE_HOME/forgelm/wizard_state.yaml`, step-diff preview, beginner / expert toggle, and the Phase 11.5 / 12.5 BYOD inline ingest + audit helpers (`_offer_ingest_for_directory`, `_offer_audit_for_jsonl`).
+Interactive CLI wizard for generating valid YAML configs. Phase 22 modernisation (2026-05-08) brought the CLI to parity with `site/js/wizard.js`: 9-step state machine (welcome / use-case / model / strategy / trainer / dataset / training-params / compliance / operations), per-trainer hyperparameters (`dpo_beta` / `simpo_beta` + `simpo_gamma` / `kto_beta` / `orpo_beta` / `grpo_*`), full PEFT method coverage (`lora` / `dora` / `pissa` / `rslora`) plus GaLore axis, EU AI Act Article 9 + 10 + 11 + 12 / 17(1)(h) / 72 compliance accordions, F-compliance-110 strict-tier auto-coercion, `back` / `reset` navigation, XDG-aware persistence at `$XDG_CACHE_HOME/forgelm/wizard_state.yaml`, step-diff preview, beginner / expert toggle, and the Phase 11.5 / 12.5 BYOD inline ingest + audit helpers (`_offer_ingest_for_directory`, `_offer_audit_for_jsonl`).
 
 ### `webhook.py`
 Sends structured JSON payloads to Slack/Teams/generic webhooks on training start, success, and failure. Supports URL from config or environment variable. Graceful error handling with configurable timeout.
diff --git a/docs/reference/compliance_summary-tr.md b/docs/reference/compliance_summary-tr.md
index fcac92cc..1437b221 100644
--- a/docs/reference/compliance_summary-tr.md
+++ b/docs/reference/compliance_summary-tr.md
@@ -52,7 +52,7 @@ Düzenleyicinin sorduğu vs ForgeLM'in cevapladığı:
 | Veri kaynağı | `data_provenance.json`; `compute_dataset_fingerprint` (SHA-256 + size + mtime); `hf_revision_source` ile derecelendirilmiş dataset Hub commit SHA'sı |
 | Teknik dokümantasyon | `annex_iv_metadata.json`; Annex IV §§1-9 kanonik düzen |
 | Uygunluk kanıtı | `compliance_report.json`; `model_card.md`; `model_integrity.json` |
-| İzleme + post-market gözetim | Webhook lifecycle (`notify_*`); `safety_trend.jsonl` cross-run trend |
+| Harici izleme planı beyanı | `monitoring:` yalnızca metadata kaydeder; Madde 72 izleme sistemini ayrıca işletin. Webhook’lar yaşam döngüsü bildirimleridir; `safety_trend.jsonl` koşular arası değerlendirme eğilimlerini kaydeder. |
 | İnsan gözetimi | Madde 14 staging gate; `human_approval.required/granted/rejected` |
 
 ## ForgeLM her gereksinimi nerede karşılar
diff --git a/docs/reference/compliance_summary.md b/docs/reference/compliance_summary.md
index bc7a9019..1bd714c3 100644
--- a/docs/reference/compliance_summary.md
+++ b/docs/reference/compliance_summary.md
@@ -53,7 +53,7 @@ What the regulator asks vs. how ForgeLM answers:
 | Data provenance | `data_provenance.json`; `compute_dataset_fingerprint` (SHA-256 + size + mtime); dataset Hub commit SHA graded by `hf_revision_source` |
 | Technical documentation | `annex_iv_metadata.json`; Annex IV §§1-9 canonical layout |
 | Conformity evidence | `compliance_report.json`; `model_card.md`; `model_integrity.json` |
-| Monitoring + post-market surveillance | Webhook lifecycle (`notify_*`); `safety_trend.jsonl` cross-run trend |
+| External monitoring-plan declaration | `monitoring:` metadata only; operate the Art. 72 monitoring system separately. Webhooks notify lifecycle events; `safety_trend.jsonl` records cross-run evaluation trends. |
 | Human oversight | Article 14 staging gate; `human_approval.required/granted/rejected` |
 
 ## Where ForgeLM meets each requirement
diff --git a/docs/reference/configuration-tr.md b/docs/reference/configuration-tr.md
index bc7c49ef..20506bce 100644
--- a/docs/reference/configuration-tr.md
+++ b/docs/reference/configuration-tr.md
@@ -316,16 +316,81 @@ uzatmasını engeller.
 
 > **Sıkı kapı:** yukarıdaki [`compliance.risk_classification`](#compliance-isteğe-bağlı--eu-ai-act-madde-11--annex-iv) ile aynı — `risk_category`'i `"high-risk"` veya `"unacceptable"` olarak ayarlamak `evaluation.safety.enabled: true` gerektirir, aksi halde config-load `ConfigError` (çıkış kodu 1) fırlatır. Kapı her iki alan üzerinden OR'lanır: ikisinden biri sıkı bir tier'daysa kapı tetiklenir.
 
-## `monitoring` (İsteğe bağlı — EU AI Act Madde 12+17)
+## `monitoring` (İsteğe bağlı — Harici izleme planı, Madde 72 / 17(1)(h))
+
+Bu blok **yalnızca beyan niteliğinde metadata** içerir. `enabled: true` olduğunda ForgeLM, harici
+pazar sonrası izleme planınızı uyumluluk manifestine ve Annex IV `monitoring_and_logging` bölümüne
+`declarative_only: true` etiketiyle kaydeder. Endpoint'e bağlanmaz, metrikleri dışa aktarmaz, drift
+tespit etmez, drift uyarısı göndermez veya kontrol zamanlamaz. İzleme altyapınızı ayrıca işletin;
+bu beyan izlemenin çalıştığını kanıtlamaz ve izleme yükümlülüğünü tek başına karşılamaz.
+Pazar sonrası izleme planı [Madde 72](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng#art_72),
+QMS ise [Madde 17(1)(h)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng#art_17) kapsamındadır.
 
 | Alan | Tip | Varsayılan | Açıklama |
 |------|-----|-----------|----------|
-| `enabled` | bool | `false` | İzleme hook'larını etkinleştir |
-| `endpoint` | string | `""` | İzleme webhook URL'si |
-| `endpoint_env` | string | `null` | Endpoint için ortam değişkeni adı |
-| `metrics_export` | string | `"none"` | `"none"`, `"prometheus"`, `"datadog"`, `"custom_webhook"` |
-| `alert_on_drift` | bool | `true` | Model sapmasında uyar |
-| `check_interval_hours` | int | `24` | İzleme kontrol aralığı (saat) |
+| `enabled` | bool | `false` | Harici planı manifeste kaydet; metadata kaydı isteğe bağlıdır; yasal yükümlülükler ayrıca geçerlidir |
+| `endpoint` | string | `""` | İsteğe bağlı URL; kanıta yalnızca scheme/host/port yazılır, bağlantı kurulmaz |
+| `endpoint_env` | string | `null` | ASCII tanımlayıcısı `endpoint`'ten önceliklidir; düz metin kanıta `$NAME` yazılır, değeri okunmaz; kimlik bilgisi yazmayın |
+| `metrics_export` | string | `"none"` | Beyan edilen harici aktarım sistemi: `"none"`, `"prometheus"`, `"datadog"`, `"custom_webhook"`; etkin planda `none` dışındaki seçim yapılandırma yüklenirken uyarı üretir |
+| `alert_on_drift` | bool | `true` | Harici planda drift uyarısı öngörülüp öngörülmediği; uyarıları etkinleştirmez |
+| `check_interval_hours` | int | `24` | Beyan edilen harici kontrol aralığı, saat (en az 1); kontrol zamanlamaz |
+
+Her iki endpoint alanı da boşsa manifeste `endpoint: null` yazılır. Adlandırılan ortam değişkeni
+tanımlı olsa bile referansı çözümlenmez. Kapalı veya belirtilmemiş planlar kaydedilmez.
+
+ForgeLM'de metadata kaydı isteğe bağlıdır; Madde 72, yüksek riskli yapay zekâ sistemlerinin sağlayıcılarının pazar
+sonrası izleme sistemi kurmasını ve belgelemesini gerektirir. Bu alanlar bir özettir; Annex IV madde 9'un (Madde
+72(3)) gerektirdiği ayrıntılı plan değildir.
+
+`load_config`, etkin ve `none` dışında exporter beyan eden plan için uyarı verir; doğrudan `ForgeConfig(...)` kurulumu
+bu dosya yükleme uyarısını vermez.
+
+İzleme kanıtı dosyaları (`compliance_report.json` ve üretildiğinde `annex_iv_metadata.json`) düz metindir.
+Düzleştirilmiş `training_manifest.yaml` özeti izleme metadatası içermez.
+
+Literal `endpoint` kırpılır; `https://monitor.example.com:8443/plan` gibi host ve geçerli port içeren mutlak bir
+HTTP(S) URL olmalıdır. Şemasız değerler reddedilir. Kanıtta scheme, host ve varsa port korunur; IPv6 köşeli
+parantezleri kalır, kullanıcı bilgisi, path, query ve fragment çıkarılır. `endpoint_redacted: true`, URL zaten
+yalnızca scheme/host/port içerse bile literal endpoint'e bu işlemin uygulandığını belirtir. Kopyalanan girdi YAML'ı,
+tarayıcı önizlemesi ve yerel depolama tam URL'yi korur; bu nedenle URL'ye kimlik bilgisi yazmayın. Sır içeren URL için
+`endpoint_env` tercih edin.
+
+`endpoint_env` adının çevresindeki boşluklar kırpılır; ASCII tanımlayıcısı olmalıdır (önce harf veya alt çizgi,
+ardından harf, rakam veya alt çizgi). Değişkenin değeri okunmadan adı `$NAME` olarak aynen kaydedilir. Boş ad literal
+endpoint'e geri döner. Doğrulama mesajları, Pydantic'in ham girdi alıntılarını göstermeden alan yolunu ve hata
+nedenini korur.
+
+`defaulted_fields`, operatörün açıkça yazdığı değerler yerine şema varsayılanlarından gelen kayıt alanlarını listeler.
+Kütüphane kullanıcıları, yapılandırmayı yeniden kurarken bu ayrımı korumak için `model_dump(exclude_unset=True)`
+kullanmalıdır; tam döküm varsayılanları açık girdiye dönüştürür. Yalnızca planı etkinleştirmek, 24 saatlik kontrol
+aralığı veya drift uyarısı politikası beyan etmez. Örneğin `monitoring: {enabled: true}`, `compliance_report.json`
+içinde şu kaydı üretir:
+
+```json
+{
+  "monitoring": {
+    "declarative_only": true,
+    "endpoint": null,
+    "endpoint_redacted": false,
+    "metrics_export": "none",
+    "alert_on_drift": true,
+    "check_interval_hours": 24,
+    "defaulted_fields": [
+      "endpoint",
+      "metrics_export",
+      "alert_on_drift",
+      "check_interval_hours"
+    ]
+  }
+}
+```
+
+Endpoint yoksa kanıta `null` yazılır; girdi YAML'ında `endpoint` alanını atlayın veya `endpoint: ""` kullanın (`null`
+değil). `compliance:` içindeki sağlayıcı adı, sistem adı veya kullanım amacından en az biri doluysa Annex IV dosyası
+üretilir; doğrulama üçünün de dolu olmasını gerektirir. `declarative_only` anahtarı bulunmayan eski izleme kayıtlarını
+da beyan olarak değerlendirin; çalışan izleme kanıtı saymayın. Eski kanıtta `defaulted_fields` yoksa varsayılanların
+kaynağı kaydedilmemiştir.
+
 
 ## `distributed` (İsteğe bağlı)
 
diff --git a/docs/reference/configuration.md b/docs/reference/configuration.md
index 69f04b76..77690e50 100644
--- a/docs/reference/configuration.md
+++ b/docs/reference/configuration.md
@@ -374,16 +374,78 @@ silently extend the retention horizon by re-using a stale workspace.
 
 ---
 
-## `monitoring` (Optional — EU AI Act Art. 12+17)
+## `monitoring` (Optional — External monitoring plan, Art. 72 / 17(1)(h))
+
+This block is **declarative metadata only**. With `enabled: true`, ForgeLM records your external
+post-market monitoring plan in the compliance manifest and Annex IV `monitoring_and_logging` section,
+labelled `declarative_only: true`. It does not contact the endpoint, export metrics, detect drift,
+send drift alerts, or schedule checks. Operate your monitoring stack separately; this declaration
+does not establish that monitoring is running or satisfy the monitoring obligation by itself.
+The post-market monitoring plan is covered by [Article 72](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng#art_72)
+and the QMS by [Article 17(1)(h)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng#art_17).
 
 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
-| `enabled` | bool | `false` | Enable monitoring hooks |
-| `endpoint` | string | `""` | Monitoring webhook URL |
-| `endpoint_env` | string | `null` | Env var name for endpoint |
-| `metrics_export` | string | `"none"` | `"none"`, `"prometheus"`, `"datadog"`, `"custom_webhook"` |
-| `alert_on_drift` | bool | `true` | Alert on model drift |
-| `check_interval_hours` | int | `24` | Monitoring check interval |
+| `enabled` | bool | `false` | Record the external plan in the manifest; optional metadata recording; legal obligations apply separately |
+| `endpoint` | string | `""` | Optional URL; evidence keeps scheme/host/port only, never contacted |
+| `endpoint_env` | string | `null` | ASCII identifier takes precedence over `endpoint`; recorded as `$NAME` in plain-text evidence, never read; no credentials |
+| `metrics_export` | string | `"none"` | Declared external exporter: `"none"`, `"prometheus"`, `"datadog"`, `"custom_webhook"`; an enabled non-`none` declaration emits a config-load warning |
+| `alert_on_drift` | bool | `true` | Whether the external plan calls for drift alerts; does not activate alerts |
+| `check_interval_hours` | int | `24` | Declared external cadence in hours (minimum 1); does not schedule checks |
+
+If both endpoint fields are empty, the manifest records `endpoint: null`. Environment-variable
+references remain unresolved even if the named variable is set. Disabled or omitted plans are not recorded.
+
+An enabled non-`none` exporter declaration emits a warning at `load_config`; direct `ForgeConfig(...)` construction
+does not emit this file-load warning.
+
+Recording metadata in ForgeLM is optional; Article 72 requires providers of high-risk AI systems to establish and
+document a post-market monitoring system. These fields are a summary, not the detailed plan required by Annex IV point
+9 (Article 72(3)).
+
+Monitoring evidence files (`compliance_report.json` and, when emitted, `annex_iv_metadata.json`) are plain text. The
+flattened `training_manifest.yaml` summary does not contain monitoring metadata.
+
+A literal `endpoint` is trimmed and must be an absolute HTTP(S) URL with a host and valid port, such as
+`https://monitor.example.com:8443/plan`; schemeless values are rejected. Evidence retains scheme, host and optional
+port, preserving IPv6 brackets; userinfo, path, query and fragment are removed. `endpoint_redacted: true` means the
+literal endpoint passed through this projection, even if it already contained only scheme/host/port. The copied input
+YAML, browser preview and local storage retain the full URL, so never put credentials in it. Prefer `endpoint_env` for
+a secret-bearing URL.
+
+The `endpoint_env` name is trimmed, must be an ASCII identifier (letters or underscore first, then letters, digits, or
+underscores), and is recorded verbatim as `$NAME` without reading the variable. A blank name falls back to the literal
+endpoint. Validation diagnostics preserve field paths and reasons without Pydantic raw-input excerpts.
+
+`defaulted_fields` lists recorded fields whose values came from schema defaults, rather than explicit operator input.
+Library callers must use `model_dump(exclude_unset=True)` when rebuilding a config to preserve that distinction; a
+full dump makes defaults explicit. An enabled plan alone therefore does not declare a 24-hour check cadence or a
+drift-alert policy. For example, `monitoring: {enabled: true}` produces this entry in `compliance_report.json`:
+
+```json
+{
+  "monitoring": {
+    "declarative_only": true,
+    "endpoint": null,
+    "endpoint_redacted": false,
+    "metrics_export": "none",
+    "alert_on_drift": true,
+    "check_interval_hours": 24,
+    "defaulted_fields": [
+      "endpoint",
+      "metrics_export",
+      "alert_on_drift",
+      "check_interval_hours"
+    ]
+  }
+}
+```
+
+An absent endpoint is `null` in evidence; in input YAML omit `endpoint` or use `endpoint: ""` (not `null`). Annex IV
+output is emitted when `compliance:` contains at least one non-blank provider name, system name, or intended purpose;
+verification requires all three. Legacy monitoring blocks without `declarative_only` must also be treated as
+declarations, never proof of runtime monitoring. Missing `defaulted_fields` in old evidence means the provenance of
+defaults was not recorded.
 
 ---
 
diff --git a/docs/reference/iso27001_control_mapping-tr.md b/docs/reference/iso27001_control_mapping-tr.md
index 6ff24161..f3f85762 100644
--- a/docs/reference/iso27001_control_mapping-tr.md
+++ b/docs/reference/iso27001_control_mapping-tr.md
@@ -28,7 +28,7 @@
 | A.5.5 Yetkililerle iletişim | OOS | — |
 | A.5.6 Özel ilgi gruplarıyla iletişim | OOS | — |
 | A.5.7 Tehdit istihbaratı | OOS | — |
-| A.5.8 Proje yönetiminde bilgi güvenliği | FL-helps | `risk_assessment` config; F-compliance-110 strict gate; Annex IV §9 |
+| A.5.8 Proje yönetiminde bilgi güvenliği | FL-helps | `risk_assessment` config; F-compliance-110 strict gate; Annex IV point 5 |
 | A.5.9 Bilgi ve ilişkili varlık envanteri | FL-helps | `data_provenance.json`; `model_integrity.json`; SBOM |
 | A.5.10 Bilginin kabul edilebilir kullanımı | OOS | — |
 | A.5.11 Varlıkların iadesi | OOS | — |
diff --git a/docs/reference/iso27001_control_mapping.md b/docs/reference/iso27001_control_mapping.md
index 3915aa7c..57508fe9 100644
--- a/docs/reference/iso27001_control_mapping.md
+++ b/docs/reference/iso27001_control_mapping.md
@@ -29,7 +29,7 @@
 | A.5.5 Contact with authorities | OOS | — |
 | A.5.6 Contact with special interest groups | OOS | — |
 | A.5.7 Threat intelligence | OOS | — |
-| A.5.8 Information security in project management | FL-helps | `risk_assessment` config; F-compliance-110 strict gate; Annex IV §9 |
+| A.5.8 Information security in project management | FL-helps | `risk_assessment` config; F-compliance-110 strict gate; Annex IV point 5 |
 | A.5.9 Inventory of information and other associated assets | FL-helps | `data_provenance.json`; `model_integrity.json`; SBOM |
 | A.5.10 Acceptable use of information | OOS | — |
 | A.5.11 Return of assets | OOS | — |
diff --git a/docs/reference/verify_annex_iv_subcommand-tr.md b/docs/reference/verify_annex_iv_subcommand-tr.md
index 699d0a41..5627854e 100644
--- a/docs/reference/verify_annex_iv_subcommand-tr.md
+++ b/docs/reference/verify_annex_iv_subcommand-tr.md
@@ -163,12 +163,16 @@ Doğrulayıcı statik bir katalogu (`_ANNEX_IV_REQUIRED_FIELDS`) yürür; böyle
 | `system_identification` | §1 — sistem tanıtımı (ad, sürüm, sağlayıcı, intended_purpose). |
 | `intended_purpose` | §1 — amaçlanan kullanım beyanı. |
 | `system_components` | §2 — yazılım / donanım bileşenleri + tedarikçi listesi. |
-| `computational_resources` | §2(g) — eğitim sırasında kullanılan hesaplama kaynakları. |
+| `computational_resources` | Annex IV point 2(c) — eğitim sırasında kullanılan hesaplama kaynakları. |
 | `data_governance` | §2(d) — veri kaynakları, yönetişim, doğrulama metodolojisi. |
 | `technical_documentation` | §3-5 — tasarım + geliştirme metodolojisi. |
-| `monitoring_and_logging` | §6 — pazara-sonrası izleme + audit-log varlığı. |
-| `performance_metrics` | §7 — doğruluk / dayanıklılık / siber güvenlik metrikleri. |
-| `risk_management` | §9 — risk yönetim sistemi referansı (Madde 9 hizalaması). |
+| `monitoring_and_logging` | Annex IV madde 9 (Madde 72(3)) — Beyan edilen harici izleme planı (`declarative_only: true`) veya audit log referansı; çalışan izleme kanıtı değildir. |
+| `performance_metrics` | Annex IV points 2(g) / 4 — doğruluk / dayanıklılık / siber güvenlik metrikleri. |
+| `risk_management` | Annex IV point 5 — risk yönetim sistemi referansı (Madde 9 hizalaması). |
+
+Bu tablo kısmi metadata eşlemelerini gösterir; tam bir Annex IV belgesi değildir. Diğer eski bölüm etiketlerinin
+düzeltilmesi [#539](https://github.com/HodeTech/ForgeLM/issues/539), doğrulayıcının tamlık iddiaları
+[#573](https://github.com/HodeTech/ForgeLM/issues/573) kapsamında takip edilir.
 
 ## Emit edilen audit event'leri
 
diff --git a/docs/reference/verify_annex_iv_subcommand.md b/docs/reference/verify_annex_iv_subcommand.md
index c8fdc943..01ab3c82 100644
--- a/docs/reference/verify_annex_iv_subcommand.md
+++ b/docs/reference/verify_annex_iv_subcommand.md
@@ -163,12 +163,16 @@ The verifier walks a static catalog (`_ANNEX_IV_REQUIRED_FIELDS`) so a future sc
 | `system_identification` | §1 — system identification (name, version, provider, intended_purpose). |
 | `intended_purpose` | §1 — intended purpose statement. |
 | `system_components` | §2 — software / hardware components + supplier list. |
-| `computational_resources` | §2(g) — compute resources used during training. |
+| `computational_resources` | Annex IV point 2(c) — compute resources used during training. |
 | `data_governance` | §2(d) — data sources, governance, validation methodology. |
 | `technical_documentation` | §3-5 — design + development methodology. |
-| `monitoring_and_logging` | §6 — post-market monitoring + audit-log presence. |
-| `performance_metrics` | §7 — accuracy / robustness / cybersecurity metrics. |
-| `risk_management` | §9 — risk management system reference (Article 9 alignment). |
+| `monitoring_and_logging` | Annex IV point 9 (Art. 72(3)) — Declared external monitoring-plan metadata (`declarative_only: true`), or an audit-log reference; no runtime monitoring evidence. |
+| `performance_metrics` | Annex IV points 2(g) / 4 — accuracy / robustness / cybersecurity metrics. |
+| `risk_management` | Annex IV point 5 — risk management system reference (Article 9 alignment). |
+
+This table maps partial metadata, not complete Annex IV sections. Remaining legacy section labels are tracked in
+[#539](https://github.com/HodeTech/ForgeLM/issues/539); verifier completeness claims are tracked in
+[#573](https://github.com/HodeTech/ForgeLM/issues/573).
 
 ## Audit events emitted
 
diff --git a/docs/standards/testing.md b/docs/standards/testing.md
index 12e14948..68398aed 100644
--- a/docs/standards/testing.md
+++ b/docs/standards/testing.md
@@ -59,6 +59,18 @@ tests/
 - `conftest.py` holds shared fixtures only. Domain-specific helpers live in the test files that need them.
 - A new feature PR adds the matching `test_*.py` in the **same** PR. No "tests in next PR."
 
+## Browser wizard regressions
+
+Browser wizard regressions live in `tests/site/*.test.cjs`. Run
+`node --test tests/site/*.test.cjs` with Node 22 or newer and the development
+Python environment available. The harness uses the active virtual environment, then the checkout's
+`.venv`, then `python3`; set `FORGELM_TEST_PYTHON` to an explicit interpreter path if needed.
+CI runs them on Node 24 in the validation job after installing the Python development dependencies.
+They use Node's built-in test runner and a VM to execute the actual wizard;
+generated YAML is also parsed by the real `ForgeConfig`. No npm packages,
+browser download, GPU, or network calls are needed. DOM shims exercise the
+Operations renderer only; they do not replace browser or native-language review.
+
 ## Fixtures
 
 The canonical factory lives in
diff --git a/docs/usermanuals/en/compliance/annex-iv.md b/docs/usermanuals/en/compliance/annex-iv.md
index a2aa31ce..d5562496 100644
--- a/docs/usermanuals/en/compliance/annex-iv.md
+++ b/docs/usermanuals/en/compliance/annex-iv.md
@@ -72,19 +72,23 @@ A field is considered missing when the key is absent OR the value is `None`, an
 
 `system_identification` is additionally validated **at sub-field depth**: `provider_name`, `system_name` and `intended_purpose` must each be non-empty, otherwise a populated-looking container of blank placeholders would pass. When one of those is empty, `missing_fields` carries a **dotted path** rather than a bare key — e.g. `["system_identification.provider_name", "system_identification.system_name"]`. Tooling that assumes `missing_fields` entries are always top-level keys will mis-handle these.
 
-The nine required keys map onto Annex IV §1-9:
+The verifier requires nine ForgeLM metadata keys; these do not represent nine complete Annex IV sections:
 
 | Top-level key | Annex IV section |
 |---|---|
 | `system_identification` | §1 — `provider_name`, `system_name`, `intended_purpose` (each validated non-empty), plus `provider_contact`, `system_version`, `risk_classification`. |
 | `intended_purpose` | §1 — intended purpose statement. |
 | `system_components` | §2 — software / hardware components + supplier list. |
-| `computational_resources` | §2(g) — compute resources used during training. |
+| `computational_resources` | Annex IV point 2(c) — compute resources used during training. |
 | `data_governance` | §2(d) — data sources, governance, validation methodology. |
 | `technical_documentation` | §3-5 — design + development methodology. |
-| `monitoring_and_logging` | §6 — post-market monitoring + audit-log presence. |
-| `performance_metrics` | §7 — accuracy / robustness / cybersecurity metrics. |
-| `risk_management` | §9 — risk management system reference (Article 9 alignment). |
+| `monitoring_and_logging` | Annex IV point 9 (Art. 72(3)) — Declared external monitoring-plan metadata (`declarative_only: true`), or an audit-log reference; no runtime monitoring evidence. |
+| `performance_metrics` | Annex IV points 2(g) / 4 — accuracy / robustness / cybersecurity metrics. |
+| `risk_management` | Annex IV point 5 — risk management system reference (Article 9 alignment). |
+
+This table maps partial metadata, not complete Annex IV sections. Remaining legacy section labels are tracked in
+[#539](https://github.com/HodeTech/ForgeLM/issues/539); verifier completeness claims are tracked in
+[#573](https://github.com/HodeTech/ForgeLM/issues/573).
 
 ### What "manifest hash mismatch" means
 
diff --git a/docs/usermanuals/en/compliance/overview.md b/docs/usermanuals/en/compliance/overview.md
index 1a800083..7afdd5b4 100644
--- a/docs/usermanuals/en/compliance/overview.md
+++ b/docs/usermanuals/en/compliance/overview.md
@@ -122,21 +122,26 @@ Every field from `compliance:` flows into `annex_iv_metadata.json`. Required fie
 
 ## What goes into Annex IV
 
-The artifact carries **nine** top-level keys mapped onto Annex IV §1-9, populated from the training manifest and your `compliance:` block:
+The artifact carries **nine** ForgeLM metadata keys, populated from the training manifest and your `compliance:`
+block; these do not represent nine complete Annex IV sections:
 
 | Top-level key | Annex IV section |
 |---|---|
 | `system_identification` | §1 — provider_name, system_name, system_version, provider_contact, intended_purpose, risk_classification. |
 | `intended_purpose` | §1 — intended purpose statement. |
 | `system_components` | §2 — model lineage + training parameters. |
-| `computational_resources` | §2(g) — compute used during training. |
+| `computational_resources` | Annex IV point 2(c) — compute used during training. |
 | `data_governance` | §2(d) — data provenance and validation methodology. |
 | `technical_documentation` | §3-5 — ForgeLM version, generation timestamp, known limitations. |
-| `monitoring_and_logging` | §6 — post-market monitoring + audit-log reference. |
-| `performance_metrics` | §7 — accuracy / robustness metrics from the evaluation results. |
-| `risk_management` | §9 — the top-level `risk_assessment:` block, or an explicit "no risk_assessment block configured" marker. |
+| `monitoring_and_logging` | Annex IV point 9 (Art. 72(3)) — Declared external monitoring-plan metadata (`declarative_only: true`), or an audit-log reference; no runtime monitoring evidence. |
+| `performance_metrics` | Annex IV points 2(g) / 4 — accuracy / robustness metrics from the evaluation results. |
+| `risk_management` | Annex IV point 5 — the top-level `risk_assessment:` block, or an explicit "no risk_assessment block configured" marker. |
 
-There is no deployment-geography, lifecycle, standards-list, declaration-of-conformity or post-market-monitoring-plan section — earlier drafts of this page listed those, and none is emitted. [Annex IV](#/compliance/annex-iv) is the single source of truth for this table.
+This table maps partial metadata, not complete Annex IV sections. Remaining legacy section labels are tracked in
+[#539](https://github.com/HodeTech/ForgeLM/issues/539); verifier completeness claims are tracked in
+[#573](https://github.com/HodeTech/ForgeLM/issues/573).
+
+There is no deployment-geography, lifecycle, standards-list, declaration-of-conformity section — earlier drafts of this page listed those, and none is emitted. The monitoring entry above is a metadata summary, not the detailed post-market monitoring plan required by Annex IV point 9. [Annex IV](#/compliance/annex-iv) is the single source of truth for this table.
 
 ## Operational responsibilities (you, not ForgeLM)
 
diff --git a/docs/usermanuals/en/operations/air-gap.md b/docs/usermanuals/en/operations/air-gap.md
index f396e88b..aacf64e5 100644
--- a/docs/usermanuals/en/operations/air-gap.md
+++ b/docs/usermanuals/en/operations/air-gap.md
@@ -47,23 +47,23 @@ $ export HF_DATASETS_OFFLINE=1
 $ forgelm --config configs/run.yaml --offline
 ```
 
-The `--offline` flag tells ForgeLM to refuse any network call. If your config references a model that isn't cached, training fails with a clear error message:
+The `--offline` flag sets HuggingFace offline environment variables. Configure other network clients separately; the flag does not disable them. Required models, tokenisers, and datasets must be cached before disconnecting:
 
 ```text
 ERROR: model "meta-llama/Llama-Guard-3-8B" not in local cache.
        Pre-cache from a connected machine: forgelm cache-models --safety meta-llama/Llama-Guard-3-8B
 ```
 
-## What `--offline` enforces
+## `--offline` scope and disconnected-run preparation
 
-| Subsystem | Behaviour with `--offline` |
+| Subsystem | Flag effect or required operator preparation |
 |---|---|
-| HuggingFace Hub downloads | Disabled. Fail-fast if a model/tokeniser isn't cached. |
-| W&B / MLflow / Comet | Disabled (or local-only for MLflow with file URI). |
-| OpenAI / Anthropic judges | Disabled. Use a local judge model. |
-| Webhooks | Disabled (or restricted to internal hosts via `webhook.allow_private`). |
-| pip install (during runtime) | Disabled. |
-| Telemetry | None — ForgeLM never phones home regardless. |
+| HuggingFace Hub downloads | `--offline` sets the HF offline flags; pre-cache the required artifacts. |
+| W&B / MLflow / Comet | Disable reporting or configure local-only MLflow with a file URI; `--offline` does not disable these integrations. |
+| OpenAI / Anthropic judges | Disable API-backed evaluation or use a local judge; `--offline` does not disable API clients. |
+| Webhooks | `--offline` does not disable configured delivery. Omit the `webhook:` block, or leave `url` / `url_env` unset, to disable it; `allow_private_destinations` permits private destinations in addition to public ones. |
+| pip install (during runtime) | Install required dependencies before disconnecting; `--offline` does not control pip. |
+| Telemetry | ForgeLM has no own telemetry service; configure external integrations separately. |
 
 ## Local synthetic data
 
@@ -76,7 +76,10 @@ synthetic:
   teacher_backend: "local"                    # not "api" — no network call
 ```
 
-There is no nested `synthetic.teacher:` sub-block (`provider` / `model` / `load_in_4bit`) — `teacher_model` and `teacher_backend` are flat fields on `synthetic:`. An API-backed teacher (`teacher_backend: "api"`, e.g. OpenAI / Anthropic) fails on `--offline`.
+There is no nested `synthetic.teacher:` sub-block (`provider` / `model` / `load_in_4bit`) — `teacher_model` and
+`teacher_backend` are flat fields on `synthetic:`. An API-backed teacher (`teacher_backend: "api"`, e.g. OpenAI /
+Anthropic) requires network access and is not disabled by `--offline`; use `teacher_backend: "local"` or `"file"` in
+an air-gap.
 
 ## Local LLM-as-judge
 
diff --git a/docs/usermanuals/en/operations/troubleshooting.md b/docs/usermanuals/en/operations/troubleshooting.md
index b043eb6f..97894962 100644
--- a/docs/usermanuals/en/operations/troubleshooting.md
+++ b/docs/usermanuals/en/operations/troubleshooting.md
@@ -152,11 +152,11 @@ audit:
 ## Webhooks not firing
 
 **Causes:**
-- Webhook URL points at a private IP and `webhook.allow_private` is false.
+- Webhook URL points at a private IP and `webhook.allow_private_destinations` is false.
 - TLS certificate validation fails.
-- Endpoint returns 4xx (silent in logs by default).
+- Endpoint returns 4xx (reported at WARNING level; training continues).
 
-**Fix:** Check `audit_log.jsonl` for `webhook_failed` events; they include the response status and body.
+**Fix:** Search stderr or captured CI logs for `WARNING` lines from `forgelm.webhook`. Failed deliveries are non-fatal and are not written to `audit_log.jsonl`; there is no `webhook_failed` event. Response bodies are suppressed to avoid exposing secrets.
 
 ## Where to file a bug
 
diff --git a/docs/usermanuals/en/reference/configuration.md b/docs/usermanuals/en/reference/configuration.md
index 280bf944..89ccbb1d 100644
--- a/docs/usermanuals/en/reference/configuration.md
+++ b/docs/usermanuals/en/reference/configuration.md
@@ -200,7 +200,7 @@ synthetic:
   sanity_failure_rate: 0.2                    # failure rate above which a WARNING is logged (warn-only)
 ```
 
-There is no nested `synthetic.teacher:` sub-block and no `rate_limit:` block — `teacher_model`, `teacher_backend`, `api_base`, `api_delay`, and `api_timeout` are flat fields on `synthetic:` directly. Prefer `api_key_env` (an environment-variable name) over the inline `api_key` field to avoid committing secrets. ForgeLM's YAML loader has no `${VAR}` interpolation mechanism anywhere — `*_env` fields name an environment variable that is read directly via `os.environ`, never substituted into a string.
+There is no nested `synthetic.teacher:` sub-block and no `rate_limit:` block — `teacher_model`, `teacher_backend`, `api_base`, `api_delay`, and `api_timeout` are flat fields on `synthetic:` directly. Prefer `api_key_env` (an environment-variable name) over the inline `api_key` field to avoid committing secrets. ForgeLM's YAML loader has no `${VAR}` interpolation mechanism anywhere — `*_env` fields name an environment variable that is read directly via `os.environ`, never substituted into a string, except `monitoring.endpoint_env`: that name is recorded as an unresolved `$NAME` reference only.
 
 Under `teacher_backend: "local"` you may also set `teacher_revision` to pin the teacher checkpoint to a Hub commit SHA or ref; the teacher's generations become training data, so an unpinned teacher is a data-provenance gap. It is rejected outright with `teacher_backend: "api"` or `"file"`, which never load from the Hub.
 
@@ -286,16 +286,75 @@ risk_assessment:
 
 ## `monitoring:`
 
+This block declares an external post-market monitoring plan (Art. 72 / 17(1)(h)). With `enabled: true`,
+its values are recorded in the compliance manifest and Annex IV `monitoring_and_logging` section with
+`declarative_only: true`. ForgeLM does not contact the endpoint, export metrics, detect drift, send
+drift alerts, or schedule checks. Operate your monitoring stack separately; the declaration is not
+proof that monitoring is running. The block remains optional for every risk tier.
+
 ```yaml
 monitoring:
-  enabled: false                              # Enable Article 12 post-market monitoring
-  endpoint: ""                               # Monitoring webhook URL (Prometheus / Datadog / custom)
-  endpoint_env: null                          # env var overriding endpoint
-  metrics_export: "none"                     # none | prometheus | datadog | custom_webhook
-  alert_on_drift: true                       # webhook alert on drift-detector regression
-  check_interval_hours: 24                   # monitoring cadence in hours
+  enabled: false                            # Record external plan metadata only
+  endpoint: ""                              # Optional URL; evidence keeps scheme/host/port only
+  endpoint_env: null                        # ASCII name overrides endpoint as "$NAME"; never read
+  metrics_export: "none"                    # Declared only: none | prometheus | datadog | custom_webhook
+  alert_on_drift: true                      # Declared external drift-alert policy; no alerts sent
+  check_interval_hours: 24                  # Declared external cadence; no checks scheduled
+```
+
+An enabled non-`none` exporter declaration emits a warning at `load_config`; direct `ForgeConfig(...)` construction
+does not emit this file-load warning. If neither endpoint field is set, the manifest endpoint is `null`. Environment
+references are recorded by name even when the
+variable is set; disabled or omitted plans are not recorded.
+
+Recording metadata in ForgeLM is optional; Article 72 requires providers of high-risk AI systems to establish and
+document a post-market monitoring system. These fields are a summary, not the detailed plan required by Annex IV point
+9 (Article 72(3)).
+
+Monitoring evidence files (`compliance_report.json` and, when emitted, `annex_iv_metadata.json`) are plain text. The
+flattened `training_manifest.yaml` summary does not contain monitoring metadata.
+
+A literal `endpoint` is trimmed and must be an absolute HTTP(S) URL with a host and valid port, such as
+`https://monitor.example.com:8443/plan`; schemeless values are rejected. Evidence retains scheme, host and optional
+port, preserving IPv6 brackets; userinfo, path, query and fragment are removed. `endpoint_redacted: true` means the
+literal endpoint passed through this projection, even if it already contained only scheme/host/port. The copied input
+YAML, browser preview and local storage retain the full URL, so never put credentials in it. Prefer `endpoint_env` for
+a secret-bearing URL.
+
+The `endpoint_env` name is trimmed, must be an ASCII identifier (letters or underscore first, then letters, digits, or
+underscores), and is recorded verbatim as `$NAME` without reading the variable. A blank name falls back to the literal
+endpoint. Validation diagnostics preserve field paths and reasons without Pydantic raw-input excerpts.
+
+`defaulted_fields` lists recorded fields whose values came from schema defaults, rather than explicit operator input.
+Library callers must use `model_dump(exclude_unset=True)` when rebuilding a config to preserve that distinction; a
+full dump makes defaults explicit. An enabled plan alone therefore does not declare a 24-hour check cadence or a
+drift-alert policy. For example, `monitoring: {enabled: true}` produces this entry in `compliance_report.json`:
+
+```json
+{
+  "monitoring": {
+    "declarative_only": true,
+    "endpoint": null,
+    "endpoint_redacted": false,
+    "metrics_export": "none",
+    "alert_on_drift": true,
+    "check_interval_hours": 24,
+    "defaulted_fields": [
+      "endpoint",
+      "metrics_export",
+      "alert_on_drift",
+      "check_interval_hours"
+    ]
+  }
+}
 ```
 
+An absent endpoint is `null` in evidence; in input YAML omit `endpoint` or use `endpoint: ""` (not `null`). Annex IV
+output is emitted when `compliance:` contains at least one non-blank provider name, system name, or intended purpose;
+verification requires all three. Legacy monitoring blocks without `declarative_only` must also be treated as
+declarations, never proof of runtime monitoring. Missing `defaulted_fields` in old evidence means the provenance of
+defaults was not recorded.
+
 ## `retention:`
 
 ```yaml
diff --git a/docs/usermanuals/tr/compliance/annex-iv.md b/docs/usermanuals/tr/compliance/annex-iv.md
index 76560ba4..40aa38e5 100644
--- a/docs/usermanuals/tr/compliance/annex-iv.md
+++ b/docs/usermanuals/tr/compliance/annex-iv.md
@@ -72,19 +72,23 @@ Bir alan; anahtar yoksa VEYA değer `None`, boş string, boş liste ya da boş d
 
 `system_identification` ayrıca **alt-alan derinliğinde** doğrulanır: `provider_name`, `system_name` ve `intended_purpose` her biri boş olmamalıdır; aksi halde boş placeholder'lardan oluşan, dolu görünen bir kapsayıcı testi geçerdi. Bunlardan biri boş olduğunda `missing_fields` düz bir anahtar değil, **noktalı yol** taşır — örn. `["system_identification.provider_name", "system_identification.system_name"]`. `missing_fields` girdilerinin daima üst düzey anahtar olduğunu varsayan araçlar bunları yanlış işler.
 
-Dokuz gerekli anahtar Annex IV §1-9 ile eşleşir:
+Doğrulayıcı dokuz ForgeLM metadata anahtarı gerektirir; bunlar dokuz eksiksiz Annex IV bölümünü temsil etmez:
 
 | Üst-seviye anahtar | Annex IV bölümü |
 |---|---|
 | `system_identification` | §1 — `provider_name`, `system_name`, `intended_purpose` (her biri boş-olmama koşuluyla doğrulanır), ayrıca `provider_contact`, `system_version`, `risk_classification`. |
 | `intended_purpose` | §1 — amaçlanan kullanım beyanı. |
 | `system_components` | §2 — yazılım / donanım bileşenleri + tedarikçi listesi. |
-| `computational_resources` | §2(g) — eğitim sırasında kullanılan hesaplama kaynakları. |
+| `computational_resources` | Annex IV point 2(c) — eğitim sırasında kullanılan hesaplama kaynakları. |
 | `data_governance` | §2(d) — veri kaynakları, yönetişim, doğrulama metodolojisi. |
 | `technical_documentation` | §3-5 — tasarım + geliştirme metodolojisi. |
-| `monitoring_and_logging` | §6 — pazara-sonrası izleme + audit-log varlığı. |
-| `performance_metrics` | §7 — doğruluk / dayanıklılık / siber güvenlik metrikleri. |
-| `risk_management` | §9 — risk yönetim sistemi referansı (Madde 9 hizalaması). |
+| `monitoring_and_logging` | Annex IV madde 9 (Madde 72(3)) — Beyan edilen harici izleme planı (`declarative_only: true`) veya audit log referansı; çalışan izleme kanıtı değildir. |
+| `performance_metrics` | Annex IV points 2(g) / 4 — doğruluk / dayanıklılık / siber güvenlik metrikleri. |
+| `risk_management` | Annex IV point 5 — risk yönetim sistemi referansı (Madde 9 hizalaması). |
+
+Bu tablo kısmi metadata eşlemelerini gösterir; tam bir Annex IV belgesi değildir. Diğer eski bölüm etiketlerinin
+düzeltilmesi [#539](https://github.com/HodeTech/ForgeLM/issues/539), doğrulayıcının tamlık iddiaları
+[#573](https://github.com/HodeTech/ForgeLM/issues/573) kapsamında takip edilir.
 
 ### "Manifest hash uyuşmazlığı" ne demek
 
diff --git a/docs/usermanuals/tr/compliance/overview.md b/docs/usermanuals/tr/compliance/overview.md
index 36a6c4a4..1e4fcced 100644
--- a/docs/usermanuals/tr/compliance/overview.md
+++ b/docs/usermanuals/tr/compliance/overview.md
@@ -122,21 +122,26 @@ evaluation:
 
 ## Annex IV neyi içerir
 
-Artefact, eğitim manifestinden ve `compliance:` bloğunuzdan doldurulan, Annex IV §1-9'a eşlenmiş **dokuz** üst düzey anahtar taşır:
+Artefact, eğitim manifestinden ve `compliance:` bloğunuzdan doldurulan **dokuz** ForgeLM metadata anahtarı taşır;
+bunlar dokuz eksiksiz Annex IV bölümünü temsil etmez:
 
 | Üst düzey anahtar | Annex IV bölümü |
 |---|---|
 | `system_identification` | §1 — provider_name, system_name, system_version, provider_contact, intended_purpose, risk_classification. |
 | `intended_purpose` | §1 — kullanım amacı beyanı. |
 | `system_components` | §2 — model soyağacı + eğitim parametreleri. |
-| `computational_resources` | §2(g) — eğitim sırasında kullanılan hesaplama kaynağı. |
+| `computational_resources` | Annex IV point 2(c) — eğitim sırasında kullanılan hesaplama kaynağı. |
 | `data_governance` | §2(d) — veri kökeni ve doğrulama metodolojisi. |
 | `technical_documentation` | §3-5 — ForgeLM sürümü, üretim zaman damgası, bilinen sınırlamalar. |
-| `monitoring_and_logging` | §6 — pazar-sonrası izleme + audit log referansı. |
-| `performance_metrics` | §7 — değerlendirme sonuçlarından doğruluk / sağlamlık metrikleri. |
-| `risk_management` | §9 — top-level `risk_assessment:` bloğu ya da açık bir "no risk_assessment block configured" işareti. |
+| `monitoring_and_logging` | Annex IV madde 9 (Madde 72(3)) — Beyan edilen harici izleme planı (`declarative_only: true`) veya audit log referansı; çalışan izleme kanıtı değildir. |
+| `performance_metrics` | Annex IV points 2(g) / 4 — değerlendirme sonuçlarından doğruluk / sağlamlık metrikleri. |
+| `risk_management` | Annex IV point 5 — top-level `risk_assessment:` bloğu ya da açık bir "no risk_assessment block configured" işareti. |
 
-Dağıtım coğrafyası, yaşam döngüsü, standartlar listesi, uygunluk beyanı veya pazar-sonrası izleme planı bölümü yoktur — bu sayfanın eski taslakları bunları listeliyordu, hiçbiri emit edilmiyor. Bu tablonun tek doğruluk kaynağı [Annex IV](#/compliance/annex-iv) sayfasıdır.
+Bu tablo kısmi metadata eşlemelerini gösterir; tam bir Annex IV belgesi değildir. Diğer eski bölüm etiketlerinin
+düzeltilmesi [#539](https://github.com/HodeTech/ForgeLM/issues/539), doğrulayıcının tamlık iddiaları
+[#573](https://github.com/HodeTech/ForgeLM/issues/573) kapsamında takip edilir.
+
+Dağıtım coğrafyası, yaşam döngüsü, standartlar listesi, uygunluk beyanı bölümü yoktur — bu sayfanın eski taslakları bunları listeliyordu, hiçbiri üretilmez. Yukarıdaki izleme kaydı metadata özetidir; Annex IV madde 9’un gerektirdiği ayrıntılı pazar sonrası izleme planı değildir. Bu tablonun tek doğruluk kaynağı [Annex IV](#/compliance/annex-iv) sayfasıdır.
 
 ## Operasyonel sorumluluklar (siz, ForgeLM değil)
 
diff --git a/docs/usermanuals/tr/operations/air-gap.md b/docs/usermanuals/tr/operations/air-gap.md
index 40a7307d..034c2ff4 100644
--- a/docs/usermanuals/tr/operations/air-gap.md
+++ b/docs/usermanuals/tr/operations/air-gap.md
@@ -47,23 +47,23 @@ $ export HF_DATASETS_OFFLINE=1
 $ forgelm --config configs/run.yaml --offline
 ```
 
-`--offline` bayrağı ForgeLM'e ağ çağrısı yapmamasını söyler. Config'iniz cache'lenmemiş bir modele referans verirse eğitim açık bir hata mesajıyla başarısız olur:
+`--offline` bayrağı HuggingFace çevrimdışı ortam değişkenlerini ayarlar. Diğer ağ istemcilerini ayrıca yapılandırın; bayrak bunları kapatmaz. Gerekli model, tokenizer ve veri kümelerini bağlantıyı kesmeden önce önbelleğe alın:
 
 ```text
 ERROR: model "meta-llama/Llama-Guard-3-8B" not in local cache.
        Bağlı bir makineden önceden cache'leyin: forgelm cache-models --safety meta-llama/Llama-Guard-3-8B
 ```
 
-## `--offline`'ın zorladığı
+## `--offline` kapsamı ve bağlantısız koşu hazırlığı
 
-| Alt sistem | `--offline` ile davranış |
+| Alt sistem | Bayrağın etkisi veya operatörün yapması gereken hazırlık |
 |---|---|
-| HuggingFace Hub indirmeleri | Devre dışı. Model/tokenizer cache'de yoksa hızlı başarısız. |
-| W&B / MLflow / Comet | Devre dışı (veya MLflow için file URI ile sadece-yerel). |
-| OpenAI / Anthropic judge'lar | Devre dışı. Yerel judge model kullanın. |
-| Webhook'lar | Devre dışı (veya `webhook.allow_private` ile iç host'larla sınırlı). |
-| Runtime'da pip install | Devre dışı. |
-| Telemetry | Yok — ForgeLM zaten asla eve telefon etmez. |
+| HuggingFace Hub indirmeleri | `--offline`, HF çevrimdışı ayarlarını yapar; gerekli artefaktları önceden önbelleğe alın. |
+| W&B / MLflow / Comet | Raporlamayı kapatın veya MLflow’u file URI ile yalnızca yerel çalışacak şekilde yapılandırın; `--offline` bu entegrasyonları kapatmaz. |
+| OpenAI / Anthropic judge’lar | API tabanlı değerlendirmeyi kapatın veya yerel judge kullanın; `--offline` API istemcilerini kapatmaz. |
+| Webhook'lar | `--offline`, yapılandırılmış teslimatları kapatmaz. Kapatmak için `webhook:` bloğunu çıkarın veya `url` / `url_env` değerlerini boş bırakın; `allow_private_destinations`, genel hedeflerin yanında özel hedeflere de izin verir. |
+| Runtime’da pip install | Gerekli bağımlılıkları bağlantıyı kesmeden kurun; `--offline`, pip’i yönetmez. |
+| Telemetry | ForgeLM’in kendi telemetri servisi yoktur; harici entegrasyonları ayrıca yapılandırın. |
 
 ## Yerel sentetik veri
 
@@ -76,7 +76,10 @@ synthetic:
   teacher_backend: "local"                    # "api" değil — network çağrısı yok
 ```
 
-Nested `synthetic.teacher:` alt-bloğu (`provider` / `model` / `load_in_4bit`) yoktur — `teacher_model` ve `teacher_backend` `synthetic:` üzerinde düz alanlardır. API-destekli bir teacher (`teacher_backend: "api"`, örn. OpenAI / Anthropic) `--offline`'da başarısız olur.
+Nested `synthetic.teacher:` alt-bloğu (`provider` / `model` / `load_in_4bit`) yoktur — `teacher_model` ve
+`teacher_backend` `synthetic:` üzerinde düz alanlardır. API tabanlı teacher (`teacher_backend: "api"`, örn. OpenAI /
+Anthropic) ağ erişimi gerektirir ve `--offline` ile kapatılmaz; air-gap'te `teacher_backend: "local"` veya `"file"`
+kullanın.
 
 ## Yerel LLM-as-judge
 
diff --git a/docs/usermanuals/tr/operations/troubleshooting.md b/docs/usermanuals/tr/operations/troubleshooting.md
index dc93356e..8f9b9841 100644
--- a/docs/usermanuals/tr/operations/troubleshooting.md
+++ b/docs/usermanuals/tr/operations/troubleshooting.md
@@ -152,11 +152,11 @@ audit:
 ## Webhook'lar fırlamıyor
 
 **Sebepler:**
-- Webhook URL özel IP'ye işaret ediyor ve `webhook.allow_private` false.
+- Webhook URL özel IP'ye işaret ediyor ve `webhook.allow_private_destinations` false.
 - TLS sertifika doğrulaması başarısız.
-- Endpoint 4xx döndürüyor (varsayılan log'da sessiz).
+- Endpoint 4xx döndürüyor (WARNING düzeyinde bildirilir; eğitim devam eder).
 
-**Çözüm:** `audit_log.jsonl` içinde `webhook_failed` olaylarına bakın; yanıt durumu ve gövdeyi içerirler.
+**Çözüm:** stderr veya kaydedilen CI log’larında `forgelm.webhook` kaynaklı `WARNING` satırlarını arayın. Başarısız teslimatlar eğitimi durdurmaz ve `audit_log.jsonl` dosyasına yazılmaz; `webhook_failed` olayı yoktur. Sırların açığa çıkmasını önlemek için yanıt gövdeleri gösterilmez.
 
 ## Bug bildirimi nereye
 
diff --git a/docs/usermanuals/tr/reference/configuration.md b/docs/usermanuals/tr/reference/configuration.md
index f45afc40..869121b7 100644
--- a/docs/usermanuals/tr/reference/configuration.md
+++ b/docs/usermanuals/tr/reference/configuration.md
@@ -202,7 +202,7 @@ synthetic:
   sanity_failure_rate: 0.2                    # üzerinde bir WARNING loglanan başarısızlık oranı (yalnızca uyarı)
 ```
 
-Nested `synthetic.teacher:` alt-bloğu ve `rate_limit:` bloğu yoktur — `teacher_model`, `teacher_backend`, `api_base`, `api_delay`, `api_timeout` doğrudan `synthetic:` üzerinde düz alanlardır. Secret'ları commit etmemek için inline `api_key` alanı yerine `api_key_env` (bir ortam-değişkeni adı) tercih edin. ForgeLM'in YAML loader'ında hiçbir yerde `${VAR}` interpolasyon mekanizması yoktur — `*_env` alanları doğrudan `os.environ` ile okunan bir ortam değişkeni adlandırır, bir string içine substitüe edilmez.
+Nested `synthetic.teacher:` alt-bloğu ve `rate_limit:` bloğu yoktur — `teacher_model`, `teacher_backend`, `api_base`, `api_delay`, `api_timeout` doğrudan `synthetic:` üzerinde düz alanlardır. Secret'ları commit etmemek için inline `api_key` alanı yerine `api_key_env` (bir ortam-değişkeni adı) tercih edin. ForgeLM'in YAML loader'ında hiçbir yerde `${VAR}` interpolasyon mekanizması yoktur — `*_env` alanları doğrudan `os.environ` ile okunan bir ortam değişkeni adlandırır, bir string içine substitüe edilmez; istisna `monitoring.endpoint_env` alanıdır: bu ad yalnızca çözümlenmemiş bir `$NAME` referansı olarak kaydedilir.
 
 `teacher_backend: "local"` altında ayrıca `teacher_revision` ayarlayarak teacher checkpoint'ini bir Hub commit SHA'sına veya ref'ine sabitleyebilirsiniz; teacher'ın ürettikleri eğitim verisi hâline gelir, dolayısıyla sabitlenmemiş bir teacher bir veri-kökeni boşluğudur. `teacher_backend: "api"` veya `"file"` ile — ki bunlar Hub'dan hiç yüklemez — doğrudan reddedilir.
 
@@ -288,16 +288,81 @@ risk_assessment:
 
 ## `monitoring:`
 
+Bu blok harici bir pazar sonrası izleme planını beyan eder (Madde 72 / 17(1)(h)). `enabled: true`
+olduğunda değerleri uyumluluk manifestine ve Annex IV `monitoring_and_logging` bölümüne
+`declarative_only: true` etiketiyle kaydedilir. ForgeLM endpoint'e bağlanmaz, metrikleri dışa aktarmaz,
+drift tespit etmez, drift uyarısı göndermez veya kontrol zamanlamaz. İzleme altyapınızı ayrıca
+işletin; bu beyan izlemenin çalıştığını kanıtlamaz. ForgeLM'de metadata kaydı isteğe bağlıdır; yasal yükümlülükler
+ayrıca geçerlidir.
+
 ```yaml
 monitoring:
-  enabled: false                              # Madde 12 pazar-sonrası izlemeyi etkinleştir
-  endpoint: ""                               # İzleme webhook URL'si (Prometheus / Datadog / özel)
-  endpoint_env: null                          # endpoint'i geçersiz kılan env değişkeni
-  metrics_export: "none"                     # none | prometheus | datadog | custom_webhook
-  alert_on_drift: true                       # drift tespitinde webhook uyarısı
-  check_interval_hours: 24                   # izleme periyodu (saat)
+  enabled: false                            # Yalnızca harici plan metadata'sını kaydet
+  endpoint: ""                              # İsteğe bağlı URL; kanıtta yalnızca scheme/host/port
+  endpoint_env: null                        # ASCII ad endpoint'ten öncelikli; "$NAME" yazılır, okunmaz
+  metrics_export: "none"                    # Yalnızca beyan: none | prometheus | datadog | custom_webhook
+  alert_on_drift: true                      # Harici drift uyarısı politikası; uyarı gönderilmez
+  check_interval_hours: 24                  # Harici kontrol aralığı; kontrol zamanlanmaz
+```
+
+Etkin planda `none` dışındaki aktarım sistemi seçimi yapılandırma yüklenirken uyarı üretir. Her iki
+endpoint alanı da boşsa manifeste `null` yazılır. Ortam değişkeni tanımlı olsa bile yalnızca adı
+kaydedilir; kapalı veya belirtilmemiş planlar kaydedilmez.
+
+ForgeLM'de metadata kaydı isteğe bağlıdır; Madde 72, yüksek riskli yapay zekâ sistemlerinin sağlayıcılarının pazar
+sonrası izleme sistemi kurmasını ve belgelemesini gerektirir. Bu alanlar bir özettir; Annex IV madde 9'un (Madde
+72(3)) gerektirdiği ayrıntılı plan değildir.
+
+`load_config`, etkin ve `none` dışında exporter beyan eden plan için uyarı verir; doğrudan `ForgeConfig(...)` kurulumu
+bu dosya yükleme uyarısını vermez.
+
+İzleme kanıtı dosyaları (`compliance_report.json` ve üretildiğinde `annex_iv_metadata.json`) düz metindir.
+Düzleştirilmiş `training_manifest.yaml` özeti izleme metadatası içermez.
+
+Literal `endpoint` kırpılır; `https://monitor.example.com:8443/plan` gibi host ve geçerli port içeren mutlak bir
+HTTP(S) URL olmalıdır. Şemasız değerler reddedilir. Kanıtta scheme, host ve varsa port korunur; IPv6 köşeli
+parantezleri kalır, kullanıcı bilgisi, path, query ve fragment çıkarılır. `endpoint_redacted: true`, URL zaten
+yalnızca scheme/host/port içerse bile literal endpoint'e bu işlemin uygulandığını belirtir. Kopyalanan girdi YAML'ı,
+tarayıcı önizlemesi ve yerel depolama tam URL'yi korur; bu nedenle URL'ye kimlik bilgisi yazmayın. Sır içeren URL için
+`endpoint_env` tercih edin.
+
+`endpoint_env` adının çevresindeki boşluklar kırpılır; ASCII tanımlayıcısı olmalıdır (önce harf veya alt çizgi,
+ardından harf, rakam veya alt çizgi). Değişkenin değeri okunmadan adı `$NAME` olarak aynen kaydedilir. Boş ad literal
+endpoint'e geri döner. Doğrulama mesajları, Pydantic'in ham girdi alıntılarını göstermeden alan yolunu ve hata
+nedenini korur.
+
+`defaulted_fields`, operatörün açıkça yazdığı değerler yerine şema varsayılanlarından gelen kayıt alanlarını listeler.
+Kütüphane kullanıcıları, yapılandırmayı yeniden kurarken bu ayrımı korumak için `model_dump(exclude_unset=True)`
+kullanmalıdır; tam döküm varsayılanları açık girdiye dönüştürür. Yalnızca planı etkinleştirmek, 24 saatlik kontrol
+aralığı veya drift uyarısı politikası beyan etmez. Örneğin `monitoring: {enabled: true}`, `compliance_report.json`
+içinde şu kaydı üretir:
+
+```json
+{
+  "monitoring": {
+    "declarative_only": true,
+    "endpoint": null,
+    "endpoint_redacted": false,
+    "metrics_export": "none",
+    "alert_on_drift": true,
+    "check_interval_hours": 24,
+    "defaulted_fields": [
+      "endpoint",
+      "metrics_export",
+      "alert_on_drift",
+      "check_interval_hours"
+    ]
+  }
+}
 ```
 
+Endpoint yoksa kanıta `null` yazılır; girdi YAML'ında `endpoint` alanını atlayın veya `endpoint: ""` kullanın (`null`
+değil). `compliance:` içindeki sağlayıcı adı, sistem adı veya kullanım amacından en az biri doluysa Annex IV dosyası
+üretilir; doğrulama üçünün de dolu olmasını gerektirir. `declarative_only` anahtarı bulunmayan eski izleme kayıtlarını
+da beyan olarak değerlendirin; çalışan izleme kanıtı saymayın. Eski kanıtta `defaulted_fields` yoksa varsayılanların
+kaynağı kaydedilmemiştir.
+
+
 ## `retention:`
 
 ```yaml
diff --git a/forgelm/compliance.py b/forgelm/compliance.py
index 9bd94b2a..b21976b9 100644
--- a/forgelm/compliance.py
+++ b/forgelm/compliance.py
@@ -19,6 +19,7 @@
 from dataclasses import dataclass, field
 from datetime import datetime, timezone
 from typing import Any, Dict, List, Optional, Set, Tuple
+from urllib.parse import urlsplit
 
 from ._strict_json import dumps_strict, sanitize_non_finite
 from ._version import __version__ as _forgelm_version
@@ -1488,6 +1489,17 @@ def _component_revisions_block(_config: Any = None) -> List[Dict[str, Any]]:
     ]
 
 
+def _redact_monitoring_endpoint(endpoint: str) -> str:
+    """Project a validated URL into evidence without losing its port or IPv6 brackets."""
+    parts = urlsplit(endpoint)
+    host = parts.hostname
+    assert host is not None  # MonitoringConfig has already validated this URL.
+    if ":" in host:
+        host = f"[{host}]"
+    port = f":{parts.port}" if parts.port is not None else ""
+    return f"{parts.scheme}://{host}{port}"
+
+
 def generate_training_manifest(
     config: Any,
     metrics: Dict[str, float],
@@ -1578,14 +1590,28 @@ def generate_training_manifest(
             {"path": p, "fingerprint": compute_dataset_fingerprint(p, offline=_offline)} for p in extra_datasets
         ]
 
-    # Monitoring config
+    # Record the operator's plan, not evidence of an active monitoring system.
+    # Reconstruct library configs with exclude_unset=True to preserve this
+    # provenance; a full model_dump() makes schema defaults look explicit.
     mon_cfg = getattr(config, "monitoring", None)
     if mon_cfg and mon_cfg.enabled:
         manifest["monitoring"] = {
-            "endpoint": mon_cfg.endpoint or f"${mon_cfg.endpoint_env}",
+            "declarative_only": True,
+            "endpoint": (
+                f"${mon_cfg.endpoint_env}"
+                if mon_cfg.endpoint_env
+                else (_redact_monitoring_endpoint(mon_cfg.endpoint) if mon_cfg.endpoint else None)
+            ),
+            "endpoint_redacted": bool(mon_cfg.endpoint and not mon_cfg.endpoint_env),
             "metrics_export": mon_cfg.metrics_export,
             "alert_on_drift": mon_cfg.alert_on_drift,
             "check_interval_hours": mon_cfg.check_interval_hours,
+            "defaulted_fields": [
+                name
+                for name in ("endpoint", "metrics_export", "alert_on_drift", "check_interval_hours")
+                if ("endpoint_env" if name == "endpoint" and mon_cfg.endpoint_env else name)
+                not in mon_cfg.model_fields_set
+            ],
         }
 
     # Webhook config — preserved into the compliance report so the
@@ -1853,11 +1879,11 @@ def build_annex_iv_artifact(manifest: Dict[str, Any]) -> Optional[Dict[str, Any]
             "generated_at": manifest.get("generated_at", ""),
             "known_limitations": operator_block.get("known_limitations", ""),
         },
-        # Annex IV §6: post-market monitoring + audit-log presence.
+        # Annex IV point 9 (Art. 72(3)): external plan metadata, not the full plan.
         "monitoring_and_logging": (manifest.get("monitoring") or {"audit_log": "audit_log.jsonl"}),
         # Annex IV §7: accuracy / robustness metrics.
         "performance_metrics": manifest.get("evaluation_results", {}).get("metrics", {}),
-        # Annex IV §9: risk management system reference.
+        # Annex IV point 5: risk management system reference.
         "risk_management": manifest.get("risk_assessment")
         or {
             "art9_reference": "no risk_assessment block configured",
diff --git a/forgelm/config.py b/forgelm/config.py
index a098f3c8..4114f2df 100644
--- a/forgelm/config.py
+++ b/forgelm/config.py
@@ -4,6 +4,7 @@
 import re
 import warnings
 from typing import Annotated, Any, Dict, List, Literal, Optional, Tuple
+from urllib.parse import urlsplit
 
 import yaml
 from pydantic import BaseModel, ConfigDict, Field, FiniteFloat, ValidationError, field_validator, model_validator
@@ -1288,25 +1289,73 @@ class RiskAssessmentConfig(BaseModel):
 
 
 class MonitoringConfig(BaseModel):
-    """Art. 12+17: Post-market monitoring hooks."""
+    """Art. 72 / 17(1)(h): declarative external monitoring plan, without runtime hooks."""
 
-    model_config = ConfigDict(extra="forbid")
+    model_config = ConfigDict(extra="forbid", hide_input_in_errors=True)
 
-    enabled: bool = Field(default=False, description="Enable Article 12 post-market monitoring hooks.")
+    enabled: bool = Field(
+        default=False, description="Record external monitoring-plan metadata in the compliance manifest."
+    )
     endpoint: str = Field(
-        default="", description="Monitoring-system webhook URL (Prometheus push gateway / Datadog / custom)."
+        default="",
+        description="Optional http(s)://host plan URL; evidence keeps scheme/host/port only. Never contacted. Prefer endpoint_env.",
     )
     endpoint_env: Optional[str] = Field(
-        default=None, description="Env var name carrying the endpoint URL (overrides `endpoint` when set)."
+        default=None,
+        description="ASCII env name recorded as `$NAME` in plain-JSON evidence, overriding endpoint; never resolved. No credentials.",
     )
     metrics_export: Literal["none", "prometheus", "datadog", "custom_webhook"] = Field(
         default="none",
-        description="Metrics exporter: `none`, `prometheus`, `datadog`, or `custom_webhook`.",
+        description="Declared external exporter: `none`, `prometheus`, `datadog`, or `custom_webhook`; no metrics are exported.",
     )
     alert_on_drift: bool = Field(
-        default=True, description="Emit a webhook alert when drift detector flags a regression."
+        default=True,
+        description="Declared external drift-alert policy; ForgeLM does not detect drift or send these alerts.",
+    )
+    check_interval_hours: int = Field(
+        default=24, ge=1, description="Declared external check cadence in hours; ForgeLM does not schedule checks."
     )
-    check_interval_hours: int = Field(default=24, ge=1, description="Monitoring check cadence in hours.")
+
+    @field_validator("endpoint")
+    @classmethod
+    def _strip_endpoint(cls, value: str) -> str:
+        """Validate an optional literal URL without echoing secret-bearing input."""
+        value = value.strip()
+        if not value:
+            return value
+        try:
+            parts = urlsplit(value)
+            if (
+                parts.scheme not in ("http", "https")
+                or not parts.hostname
+                or "\\" in value
+                or any(ord(char) <= 32 or ord(char) == 127 for char in value)
+            ):
+                raise ValueError
+            _ = parts.port  # Access validates malformed and out-of-range ports.
+        except ValueError:
+            raise ValueError("monitoring.endpoint must be an http(s)://host URL with a valid port, or empty.") from None
+        return value
+
+    @field_validator("endpoint_env")
+    @classmethod
+    def _validate_endpoint_env(cls, value: Optional[str]) -> Optional[str]:
+        """Accept an unresolved ASCII identifier, with surrounding whitespace removed."""
+        if value is None or not value.strip():
+            return None
+        value = value.strip()
+        if not value.isascii() or not value.isidentifier():
+            raise ValueError("monitoring.endpoint_env must be an ASCII environment-variable name, without env: or $.")
+        return value
+
+    def _warn_declarative_exporter(self) -> None:
+        """Warn once at the file-load boundary, never during pipeline stage validation."""
+        if self.enabled and self.metrics_export != "none":
+            logger.warning(
+                "monitoring.metrics_export is declarative metadata only: ForgeLM does not export metrics, "
+                "contact the endpoint, detect drift, send drift alerts, or schedule checks. "
+                "Operate your external monitoring stack separately."
+            )
 
 
 class ComplianceMetadataConfig(BaseModel):
@@ -1613,7 +1662,8 @@ class RetentionConfig(BaseModel):
 
 
 class ForgeConfig(BaseModel):
-    model_config = ConfigDict(extra="forbid")
+    # Nested validation errors otherwise repeat secret-bearing raw inputs in CLI logs.
+    model_config = ConfigDict(extra="forbid", hide_input_in_errors=True)
 
     model: ModelConfig = Field(description="Base-model + quantisation + backend block (required).")
     lora: LoraConfigModel = Field(description="PEFT / LoRA configuration block (required).")
@@ -1642,7 +1692,7 @@ class ForgeConfig(BaseModel):
         description="EU AI Act Article 9 risk-management block (intended use, foreseeable misuse, risk tier).",
     )
     monitoring: Optional[MonitoringConfig] = Field(
-        default=None, description="EU AI Act Article 12 + 17 post-market monitoring block."
+        default=None, description="EU AI Act Art. 72 / 17(1)(h) external monitoring-plan metadata; no runtime hooks."
     )
     synthetic: Optional[SyntheticConfig] = Field(
         default=None, description="Teacher→student synthetic-data generation block."
@@ -2279,4 +2329,6 @@ def load_config(config_path: str) -> ForgeConfig:
         # / TypeError outside Pydantic's wrapper. Same message shape.
         raise ConfigError(f"Configuration validation failed: {e}") from e
 
+    if config.monitoring is not None:
+        config.monitoring._warn_declarative_exporter()
     return config
diff --git a/forgelm/verify/_annex_iv.py b/forgelm/verify/_annex_iv.py
index 52e97f14..a3b66b83 100644
--- a/forgelm/verify/_annex_iv.py
+++ b/forgelm/verify/_annex_iv.py
@@ -35,12 +35,15 @@
     ("system_identification", "Annex IV §1 — system identification (name, version, provider, intended_purpose)"),
     ("intended_purpose", "Annex IV §1 — intended purpose statement"),
     ("system_components", "Annex IV §2 — software / hardware components + supplier list"),
-    ("computational_resources", "Annex IV §2(g) — compute resources used during training"),
+    ("computational_resources", "Annex IV point 2(c) — compute resources used during training"),
     ("data_governance", "Annex IV §2(d) — data sources, governance, validation methodology"),
     ("technical_documentation", "Annex IV §3-5 — design + development methodology"),
-    ("monitoring_and_logging", "Annex IV §6 — post-market monitoring + audit-log presence"),
-    ("performance_metrics", "Annex IV §7 — accuracy / robustness / cybersecurity metrics"),
-    ("risk_management", "Annex IV §9 — risk management system reference (Art. 9 alignment)"),
+    (
+        "monitoring_and_logging",
+        "Annex IV point 9 (Art. 72(3)) — declared external plan metadata or audit-log reference",
+    ),
+    ("performance_metrics", "Annex IV points 2(g) / 4 — accuracy / robustness / cybersecurity metrics"),
+    ("risk_management", "Annex IV point 5 — risk management system reference (Art. 9 alignment)"),
 )
 
 
diff --git a/forgelm/wizard/__init__.py b/forgelm/wizard/__init__.py
index 3e8fc4ca..8d9e8b36 100644
--- a/forgelm/wizard/__init__.py
+++ b/forgelm/wizard/__init__.py
@@ -7,7 +7,7 @@
 hyperparameters (``dpo_beta``, ``simpo_beta`` / ``simpo_gamma``,
 ``kto_beta``, ``orpo_beta``, ``grpo_*``), full PEFT method coverage
 (``lora`` / ``dora`` / ``pissa`` / ``rslora``) plus GaLore as a
-separate axis, EU AI Act Article 9 / 10 / 11 / 12+17 compliance
+separate axis, EU AI Act Article 9 / 10 / 11 / 12 / 17(1)(h) / 72 compliance
 accordions, F-compliance-110 strict-tier auto-coercion, ``back`` /
 ``reset`` navigation, XDG-aware persistence at
 ``$XDG_CACHE_HOME/forgelm/wizard_state.yaml``, step-diff preview,
diff --git a/forgelm/wizard/_collectors.py b/forgelm/wizard/_collectors.py
index 50f0a6c9..83c0039a 100644
--- a/forgelm/wizard/_collectors.py
+++ b/forgelm/wizard/_collectors.py
@@ -10,7 +10,6 @@
 
 from __future__ import annotations
 
-import re
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Dict, Optional, Tuple
@@ -90,15 +89,11 @@ def _default_safety_probes_path() -> str:
 # ---------------------------------------------------------------------------
 # Shared ``env:VAR_NAME`` reference parsing — the webhook URL prompt and
 # the monitoring-endpoint prompt both accept this indirection syntax.
-# Centralised here so the accepted env-var-name shape (POSIX: uppercase
-# letters / digits / underscores, must start with a letter) only needs
+# Centralised here so the accepted ASCII-identifier shape only needs
 # updating in one place.
 # ---------------------------------------------------------------------------
 
 
-_ENV_VAR_NAME_RE = re.compile(r"^[A-Z][A-Z0-9_]*$", flags=re.ASCII)
-
-
 def _parse_env_ref(raw: str) -> Optional[str]:
     """Parse an ``env:VAR_NAME`` reference; return the variable name.
 
@@ -116,10 +111,9 @@ def _parse_env_ref(raw: str) -> Optional[str]:
     var_name = raw[4:].strip()
     if not var_name:
         raise ValueError("`env:` prefix needs a non-empty variable name (e.g. `env:SLACK_WEBHOOK_URL`).")
-    if not _ENV_VAR_NAME_RE.match(var_name):
+    if not var_name.isascii() or not var_name.isidentifier():
         raise ValueError(
-            f"`env:{var_name}` is not a POSIX environment-variable name "
-            "(uppercase letters / digits / underscores, must start with a letter)."
+            "Not a POSIX environment-variable name: use ASCII letters, digits, or underscores; no initial digit."
         )
     return var_name
 
@@ -879,28 +873,41 @@ def _collect_retention() -> Optional[Dict[str, Any]]:
 
 
 def _collect_monitoring() -> Optional[Dict[str, Any]]:
-    """Article 12+17: post-market monitoring hooks.
+    """Collect declarative external monitoring metadata for the compliance manifest.
 
-    Endpoint accepts either a literal URL or ``env:VAR_NAME`` to source
-    the URL from the environment — mirrors the webhook collector's
-    convention so the same syntax works in both places.
+    Endpoint accepts a literal URL or ``env:VAR_NAME`` recorded by name only.
     """
-    if not _prompt_yes_no("Configure post-market monitoring?", default=False):
+    if not _prompt_yes_no("Record an external post-market monitoring plan (metadata only)?", default=False):
         return None
-    endpoint_raw = _prompt(
-        "Monitoring endpoint URL (or `env:VAR_NAME` to source from environment)",
-        "",
-    )
+    _print("  Metadata only: ForgeLM does not export metrics, detect drift, send drift alerts, or schedule checks.")
+    from pydantic import ValidationError
+
+    from ..config import MonitoringConfig
+
+    while True:
+        endpoint_raw = _prompt(
+            "External plan http(s) URL (optional; evidence keeps scheme/host/port; YAML keeps the full URL; prefer env:NAME)",
+            "",
+        )
+        if endpoint_raw.strip().lower().startswith("env:"):
+            break
+        try:
+            MonitoringConfig(endpoint=endpoint_raw)
+            break
+        except ValidationError:
+            _print("  Use an http(s)://host URL with a valid port, env:NAME, or leave the endpoint empty.")
     metrics_export = _prompt_choice(
-        "metrics_export:",
+        "Declared external metrics_export:",
         ["none", "prometheus", "datadog", "custom_webhook"],
         default=1,
     )
     monitoring: Dict[str, Any] = {
         "enabled": True,
         "metrics_export": metrics_export,
-        "alert_on_drift": _prompt_yes_no("alert_on_drift?", default=True),
-        "check_interval_hours": _prompt_int("check_interval_hours", 24, min_val=1, max_val=720),
+        "alert_on_drift": _prompt_yes_no("Does your external monitoring stack alert on drift?", default=False),
+        "check_interval_hours": _prompt_int(
+            "How often does your external stack check the model (hours)?", 24, min_val=1, max_val=720
+        ),
     }
     # P9: support env:VAR_NAME indirection for the monitoring endpoint
     # the same way the webhook collector does.  ``MonitoringConfig`` has
@@ -911,7 +918,7 @@ def _collect_monitoring() -> Optional[Dict[str, Any]]:
         try:
             monitoring["endpoint_env"] = _parse_env_ref(endpoint_stripped)
         except ValueError:
-            _print(f"  '{endpoint_stripped}' is not a valid env-var reference — monitoring endpoint left unset.")
+            _print("  Input is not a valid env-var reference — monitoring endpoint left unset.")
     elif endpoint_stripped:
         monitoring["endpoint"] = endpoint_stripped
     return monitoring
diff --git a/site/js/translations.js b/site/js/translations.js
index cc00c511..fe8431f6 100644
--- a/site/js/translations.js
+++ b/site/js/translations.js
@@ -1677,8 +1677,8 @@
     "wizard.step.compliance.tutorial": "ForgeLM checks the fine-tuned model BEFORE you ship it. Auto-revert undoes the fine-tune if it's worse than the base. Llama Guard scans outputs for harmful content. The EU AI Act risk tier you pick determines which compliance artefacts get auto-generated and whether human approval is REQUIRED before deployment.",
     "wizard.step.operations.title": "Operations & advanced settings",
     "wizard.step.operations.name": "Operations",
-    "wizard.step.operations.desc": "Notifications, post-market monitoring, synthetic data generation, and retention horizons. All optional — skip the ones you don't need.",
-    "wizard.step.operations.tutorial": "These four blocks cover everything ForgeLM ships beyond the core training + compliance flow. Webhook posts a Slack/Teams/Discord alert when training starts, succeeds, or fails. Post-market monitoring (Article 12+17) wires drift detection into Prometheus / Datadog after deployment. Synthetic data generation lets a stronger teacher model produce extra training examples. Retention horizons set GDPR-compliant TTLs for the audit log and intermediate artefacts. None of these are required — leave them collapsed if your run doesn't need them.",
+    "wizard.step.operations.desc": "Notifications, external monitoring-plan metadata, synthetic data, and retention settings. Configure the ForgeLM options you need.",
+    "wizard.step.operations.tutorial": "These four blocks cover everything ForgeLM ships beyond the core training + compliance flow. Webhook posts a Slack/Teams/Discord alert when training starts, succeeds, or fails. Monitoring (Art. 72 / 17(1)(h)) records an external plan as metadata; ForgeLM does not export metrics, detect drift, send drift alerts, or schedule checks. Synthetic data generation lets a stronger teacher model produce extra training examples. Retention horizons set GDPR-compliant TTLs for the audit log and intermediate artefacts. These ForgeLM settings are optional; external legal obligations still apply.",
     "wizard.risk.label":          "EU AI Act risk classification",
     "wizard.risk.hint":           "Pick the tier your deployment falls under. High-risk and unacceptable trigger mandatory audit + human-approval obligations.",
     "wizard.risk.unknown":        "Unknown — pre-classification placeholder",
@@ -1775,25 +1775,29 @@
     "wizard.webhook.url.detail": "You can paste the URL directly (fine for non-secret test channels) or use the env: prefix to point at an environment variable. The env-var route keeps the URL out of YAML and out of the audit log — required for production secrets.",
     "wizard.webhook.enabled.label": "Send webhook notifications",
     "wizard.webhook.enabled.hint": "When on, ForgeLM posts a webhook payload at the events you select. Off keeps the YAML free of any webhook block.",
-    "wizard.webhook.enabled.detail": "Notifications wire training events into your team chat (Slack / Teams / Discord) or any HTTPS endpoint that accepts JSON. Audit log entries record every dispatch — including failures — so you have an independent trail of what was sent. Leave off if you don't want the YAML to carry webhook config at all.",
+    "wizard.webhook.enabled.detail": "Notifications wire training events into your team chat (Slack / Teams / Discord) or any HTTPS endpoint that accepts JSON. Delivery failures are warnings in the process log only. The audit log records the underlying training events, not delivery attempts. Leave off if you don't want the YAML to carry webhook config at all.",
     "wizard.webhook.notify_start.label": "Notify on training start",
     "wizard.webhook.notify_success.label": "Notify on success",
     "wizard.webhook.notify_failure.label": "Notify on failure",
-    "wizard.monitoring.title": "Post-market monitoring (Art. 12+17)",
-    "wizard.monitoring.hint": "Wire training metrics into Prometheus / Datadog / a custom webhook for ongoing observability.",
-    "wizard.monitoring.enabled.label": "Enable monitoring hooks",
-    "wizard.monitoring.enabled.hint": "Required for high-risk systems under EU AI Act Articles 12 + 17.",
-    "wizard.monitoring.enabled.detail": "Once a high-risk system is deployed, the EU AI Act requires ongoing observation — drift detection, error reporting, periodic accuracy re-checks. ForgeLM doesn't run the monitoring itself, but it wires the right hooks into the training run so your existing observability stack picks them up.",
-    "wizard.monitoring.endpoint.label": "Monitoring endpoint",
-    "wizard.monitoring.endpoint.hint": "URL of your Prometheus push gateway, Datadog, or custom webhook receiver.",
-    "wizard.monitoring.endpoint.placeholder": "https://prometheus.example.com/metrics OR env:DATADOG_URL",
-    "wizard.monitoring.metrics_export.label": "Metrics format",
+    "wizard.monitoring.title": "External monitoring plan (Art. 72 / 17(1)(h))",
+    "wizard.monitoring.hint": "Declare your external monitoring setup in the compliance manifest / Annex IV. Metadata only; ForgeLM runs no monitoring.",
+    "wizard.monitoring.enabled.label": "Record monitoring-plan metadata",
+    "wizard.monitoring.enabled.hint": "Recording metadata in ForgeLM is optional. Art. 72 requires providers of high-risk AI systems to establish and document a post-market monitoring system; Annex IV point 9 includes the detailed plan.",
+    "wizard.monitoring.enabled.detail": "ForgeLM records the declared plan with declarative_only: true. It does not contact the endpoint, export metrics, detect drift, send drift alerts, or schedule checks. Operate your external stack separately; this declaration alone does not satisfy a post-market monitoring obligation.",
+    "wizard.monitoring.endpoint.label": "Declared endpoint (optional)",
+    "wizard.monitoring.endpoint.hint": "Use http(s)://host or env:NAME; never contacted. Evidence keeps only scheme/host/port, but copied YAML, preview and local storage keep the full URL, so never put credentials in it. Prefer env:NAME: only $NAME is recorded, never its value.",
+    "wizard.monitoring.endpoint.placeholder": "https://monitor.example.com OR env:MONITORING_URL",
+    "wizard.monitoring.metrics_export.label": "Declared external exporter",
     "wizard.monitoring.metrics_export.none": "None",
     "wizard.monitoring.metrics_export.prometheus": "Prometheus",
     "wizard.monitoring.metrics_export.datadog": "Datadog",
     "wizard.monitoring.metrics_export.custom": "Custom webhook",
-    "wizard.monitoring.alert_on_drift.label": "Alert on drift",
-    "wizard.monitoring.alert_on_drift.hint": "Emit a webhook alert when the drift detector flags a regression.",
+    "wizard.monitoring.alert_on_drift.label": "My external stack alerts on drift",
+    "wizard.monitoring.alert_on_drift.hint": "Records the policy of your external stack. ForgeLM does not detect drift or send these alerts.",
+    "wizard.error.env_ref.invalid": "Use env:NAME with an ASCII identifier: letters or underscore first, then letters, digits, or underscores.",
+    "wizard.monitoring.interval.label": "External check interval (hours, 1–720)",
+    "wizard.error.monitoring.endpoint.invalid": "Use an http(s)://host URL with a valid port, env:NAME, or leave the endpoint empty.",
+    "wizard.error.monitoring.interval.invalid": "Enter a whole number from 1 to 720 hours.",
     "wizard.synthetic.title": "Synthetic data generation (advanced)",
     "wizard.synthetic.hint": "Generate teacher → student training data using a stronger model.",
     "wizard.synthetic.enabled.label": "Generate synthetic data",
@@ -1834,7 +1838,6 @@
     "wizard.error.model.format": "HuggingFace Hub IDs use the org/repo shape (e.g. meta-llama/Llama-3.1-8B-Instruct). A bare name without a slash won’t resolve at training time.",
     "wizard.error.dataset.required": "Pick a dataset before continuing — without a path or HuggingFace name the trainer has nothing to learn from.",
     "wizard.error.webhook.required": "You enabled notifications, so we need a webhook URL or env var to send them to. Either turn notifications off above or paste your Slack / Teams / Discord URL (or env:SLACK_WEBHOOK_URL for production secrets).",
-    "wizard.error.monitoring.endpoint.required": "You enabled post-market monitoring, so we need an endpoint to push metrics to. Either turn monitoring off above or paste your Prometheus / Datadog / custom-webhook URL (use env:VAR_NAME for secrets).",
     "wizard.error.synthetic.teacher_model.required": "You enabled synthetic data generation, so we need a teacher model name (e.g. gpt-4o, claude-opus-4-7, or a local model path).",
     "wizard.error.synthetic.seed_file.required": "You enabled synthetic data generation, so we need a seed prompts file — the teacher will produce a response for each prompt in it (e.g. ./data/seed_prompts.jsonl).",
     "wizard.error.synthetic.api_key_env.required": "You enabled synthetic data generation, so we need the env var that carries the teacher API key — typically OPENAI_API_KEY or ANTHROPIC_API_KEY. Leave the synthetic toggle off if you don’t have one.",
@@ -2072,8 +2075,8 @@
     "wizard.step.compliance.tutorial": "ForgeLM, fine-tune edilmiş modeli teslim etmeden ÖNCE kontrol eder. Auto-revert, fine-tune base'den kötüyse geri alır. Llama Guard çıktılarda zararlı içerik tarar. Seçtiğin EU AI Act risk seviyesi hangi uyumluluk artefaktlarının otomatik üretileceğini ve dağıtım öncesi insan onayının ZORUNLU olup olmayacağını belirler.",
     "wizard.step.operations.title": "Operasyon ve gelişmiş ayarlar",
     "wizard.step.operations.name": "Operasyon",
-    "wizard.step.operations.desc": "Bildirimler, post-market monitoring, sentetik veri üretimi ve saklama ufukları. Hepsi opsiyonel — ihtiyacın olmayanları atla.",
-    "wizard.step.operations.tutorial": "Bu dört blok, ForgeLM'in temel eğitim + compliance akışının ötesinde sağladığı her şeyi kapsar. Webhook, eğitim başladığında / başarılı olduğunda / hata verdiğinde Slack/Teams/Discord uyarısı gönderir. Post-market monitoring (Madde 12+17) deploy sonrasında drift detection'ı Prometheus / Datadog'a bağlar. Sentetik veri üretimi, daha güçlü bir teacher modelin ek eğitim örnekleri üretmesini sağlar. Saklama ufukları audit log'u ve ara artefact'ları için GDPR uyumlu TTL'ler ayarlar. Hiçbiri zorunlu değil — eğitimin ihtiyaç duymuyorsa kapalı bırak.",
+    "wizard.step.operations.desc": "Bildirimler, harici izleme planı metadata’sı, sentetik veri ve saklama ayarları. Gereken ForgeLM seçeneklerini yapılandırın.",
+    "wizard.step.operations.tutorial": "Bu dört blok, ForgeLM'in temel eğitim + compliance akışının ötesinde sağladığı her şeyi kapsar. Webhook, eğitim başladığında / başarılı olduğunda / hata verdiğinde Slack/Teams/Discord uyarısı gönderir. İzleme (Madde 72 / 17(1)(h)) harici planı metadata olarak kaydeder; ForgeLM metrikleri dışa aktarmaz, drift tespit etmez, drift uyarısı göndermez veya kontrol zamanlamaz. Sentetik veri üretimi, daha güçlü bir teacher modelin ek eğitim örnekleri üretmesini sağlar. Saklama ufukları audit log'u ve ara artefact'ları için GDPR uyumlu TTL'ler ayarlar. Bu ForgeLM ayarları isteğe bağlıdır; harici yasal yükümlülükler ayrıca geçerlidir.",
     "wizard.risk.label":          "EU AI Act risk sınıflandırması",
     "wizard.risk.hint":           "Dağıtımın hangi seviyeye girdiğini seç. High-risk ve unacceptable, zorunlu denetim + insan onay yükümlülüklerini tetikler.",
     "wizard.risk.unknown":        "Unknown — ön sınıflandırma yer tutucusu",
@@ -2170,25 +2173,29 @@
     "wizard.webhook.url.detail": "URL'yi doğrudan yapıştırabilirsin (gizli olmayan test kanalları için iyi) ya da bir ortam değişkenine işaret etmek için env: önekini kullan. Env-var yolu URL'yi YAML'den ve audit log'tan dışarıda tutar — production gizliler için zorunlu.",
     "wizard.webhook.enabled.label": "Webhook bildirimleri gönder",
     "wizard.webhook.enabled.hint": "Açıkken, ForgeLM seçtiğin olaylarda bir webhook payload gönderir. Kapalıyken YAML herhangi bir webhook bloğu içermez.",
-    "wizard.webhook.enabled.detail": "Bildirimler eğitim olaylarını ekip sohbetine (Slack / Teams / Discord) veya JSON kabul eden herhangi bir HTTPS endpoint'ine bağlar. Audit log her gönderimi — hatalar dahil — kaydeder, böylece neyin gönderildiğine dair bağımsız bir iz kalır. YAML'in webhook config taşımasını hiç istemiyorsan kapalı bırak.",
+    "wizard.webhook.enabled.detail": "Bildirimler eğitim olaylarını ekip sohbetine (Slack / Teams / Discord) veya JSON kabul eden herhangi bir HTTPS endpoint'ine bağlar. Teslimat hataları yalnızca süreç günlüğüne uyarı olarak yazılır. Audit log teslimat denemelerini değil, ilgili eğitim olaylarını kaydeder. YAML'in webhook config taşımasını hiç istemiyorsan kapalı bırak.",
     "wizard.webhook.notify_start.label": "Eğitim başında bildir",
     "wizard.webhook.notify_success.label": "Başarıda bildir",
     "wizard.webhook.notify_failure.label": "Hatada bildir",
-    "wizard.monitoring.title": "Post-market monitoring (Madde 12+17)",
-    "wizard.monitoring.hint": "Eğitim metriklerini sürekli observability için Prometheus / Datadog / özel webhook'a bağla.",
-    "wizard.monitoring.enabled.label": "Monitoring hook'ları aç",
-    "wizard.monitoring.enabled.hint": "High-risk sistemler için EU AI Act Madde 12 + 17 zorunlu kılar.",
-    "wizard.monitoring.enabled.detail": "High-risk bir sistem deploy edildikten sonra EU AI Act sürekli gözlem zorunlu kılar — drift detection, hata raporlama, periyodik doğruluk yeniden-kontrolleri. ForgeLM monitoring'i kendisi çalıştırmaz ama doğru hook'ları training run'a yerleştirir; mevcut observability stack'in alır.",
-    "wizard.monitoring.endpoint.label": "Monitoring endpoint",
-    "wizard.monitoring.endpoint.hint": "Prometheus push gateway, Datadog veya özel webhook receiver URL'si.",
-    "wizard.monitoring.endpoint.placeholder": "https://prometheus.example.com/metrics VEYA env:DATADOG_URL",
-    "wizard.monitoring.metrics_export.label": "Metrik formatı",
+    "wizard.monitoring.title": "Harici izleme planı (Madde 72 / 17(1)(h))",
+    "wizard.monitoring.hint": "Harici izleme planını uyumluluk manifestine / Annex IV'e kaydet. Yalnızca metadata; ForgeLM izleme çalıştırmaz.",
+    "wizard.monitoring.enabled.label": "İzleme planı metadatasını kaydet",
+    "wizard.monitoring.enabled.hint": "ForgeLM'de metadata kaydı isteğe bağlıdır. Madde 72, yüksek riskli yapay zekâ sistemlerinin sağlayıcılarının pazar sonrası izleme sistemi kurmasını ve belgelemesini gerektirir; Annex IV madde 9 ayrıntılı planı içerir.",
+    "wizard.monitoring.enabled.detail": "ForgeLM beyan edilen planı declarative_only: true etiketiyle kaydeder. Endpoint'e bağlanmaz, metrikleri dışa aktarmaz, drift tespit etmez, drift uyarısı göndermez veya kontrol zamanlamaz. Harici altyapını ayrıca işlet; bu beyan pazar sonrası izleme yükümlülüğünü tek başına karşılamaz.",
+    "wizard.monitoring.endpoint.label": "Beyan edilen endpoint (isteğe bağlı)",
+    "wizard.monitoring.endpoint.hint": "http(s)://host veya env:NAME kullan; bağlantı kurulmaz. Kanıtta yalnızca scheme/host/port kalır; kopyalanan YAML, önizleme ve yerel depolama tam URL'yi korur, bu yüzden kimlik bilgisi yazma. env:NAME tercih et: değeri okunmaz, yalnızca $NAME kaydedilir.",
+    "wizard.monitoring.endpoint.placeholder": "https://monitor.example.com VEYA env:MONITORING_URL",
+    "wizard.monitoring.metrics_export.label": "Beyan edilen harici aktarım sistemi",
     "wizard.monitoring.metrics_export.none": "Yok",
     "wizard.monitoring.metrics_export.prometheus": "Prometheus",
     "wizard.monitoring.metrics_export.datadog": "Datadog",
     "wizard.monitoring.metrics_export.custom": "Özel webhook",
-    "wizard.monitoring.alert_on_drift.label": "Drift'te uyar",
-    "wizard.monitoring.alert_on_drift.hint": "Drift detector regression bayrağı koyduğunda webhook uyarısı gönder.",
+    "wizard.monitoring.alert_on_drift.label": "Harici altyapım drift durumunda uyarır",
+    "wizard.monitoring.alert_on_drift.hint": "Harici altyapının politikasını kaydeder. ForgeLM drift tespit etmez veya bu uyarıları göndermez.",
+    "wizard.error.env_ref.invalid": "env:NAME için ASCII tanımlayıcısı kullan: önce harf veya alt çizgi, sonra harf, rakam veya alt çizgi.",
+    "wizard.monitoring.interval.label": "Harici kontrol aralığı (saat, 1–720)",
+    "wizard.error.monitoring.endpoint.invalid": "Geçerli port içeren http(s)://host URL'si veya env:NAME kullan; endpoint'i boş da bırakabilirsin.",
+    "wizard.error.monitoring.interval.invalid": "1–720 arasında tam sayı saat gir.",
     "wizard.synthetic.title": "Sentetik veri üretimi (gelişmiş)",
     "wizard.synthetic.hint": "Daha güçlü bir model kullanarak teacher → student eğitim verisi üret.",
     "wizard.synthetic.enabled.label": "Sentetik veri üret",
@@ -2229,7 +2236,6 @@
     "wizard.error.model.format": "HuggingFace Hub ID’leri org/repo formatında olmalı (örn. meta-llama/Llama-3.1-8B-Instruct). Slash içermeyen düz bir isim eğitim sırasında bulunamaz.",
     "wizard.error.dataset.required": "Devam etmeden önce bir dataset seç — yol veya HuggingFace ismi olmadan trainer’ın öğreneceği bir şey yok.",
     "wizard.error.webhook.required": "Bildirimleri açtın, yani göndereceğimiz bir webhook URL’si veya env değişkeni gerekli. Ya yukarıdan bildirimleri kapat ya da Slack / Teams / Discord URL’ini yapıştır (production gizliler için env:SLACK_WEBHOOK_URL kullan).",
-    "wizard.error.monitoring.endpoint.required": "Post-market monitoring’i açtın, yani metrikleri yollayacağımız bir endpoint gerekli. Ya yukarıdan monitoring’i kapat ya da Prometheus / Datadog / özel-webhook URL’ini yapıştır (gizliler için env:DEĞİŞKEN_ADI).",
     "wizard.error.synthetic.teacher_model.required": "Sentetik veri üretimini açtın, yani bir teacher model adı gerekli (örn. gpt-4o, claude-opus-4-7 veya yerel model yolu).",
     "wizard.error.synthetic.seed_file.required": "Sentetik veri üretimini açtın, yani bir seed prompts dosyası gerekli — teacher her prompt için yanıt üretir (örn. ./data/seed_prompts.jsonl).",
     "wizard.error.synthetic.api_key_env.required": "Sentetik veri üretimini açtın, yani teacher API anahtarını taşıyan env değişkeni gerekli — genellikle OPENAI_API_KEY veya ANTHROPIC_API_KEY. Bir tane yoksa sentetik anahtarını kapalı bırak.",
@@ -2467,8 +2473,8 @@
     "wizard.step.compliance.tutorial": "ForgeLM prüft das fine-getunte Modell, BEVOR Sie es ausliefern. Auto-Revert macht das Fine-Tune rückgängig, wenn es schlechter als das Basismodell ist. Llama Guard scannt Ausgaben auf schädliche Inhalte. Die EU-AI-Act-Risikoklasse, die Sie wählen, bestimmt, welche Compliance-Artefakte automatisch generiert werden und ob menschliche Freigabe vor dem Deployment ERFORDERLICH ist.",
     "wizard.step.operations.title": "Betrieb & erweiterte Einstellungen",
     "wizard.step.operations.name": "Betrieb",
-    "wizard.step.operations.desc": "Benachrichtigungen, Post-Market-Monitoring, synthetische Datengenerierung und Retentions-Horizonte. Alle optional — überspringen, was Sie nicht brauchen.",
-    "wizard.step.operations.tutorial": "Diese vier Blöcke decken alles ab, was ForgeLM jenseits des Kern-Training + Compliance-Flows liefert. Webhook sendet Slack/Teams/Discord-Alerts bei Trainingsstart/Erfolg/Fehler. Post-Market-Monitoring (Artikel 12+17) verdrahtet Drift-Detection nach Deployment in Prometheus/Datadog. Synthetische Datengenerierung lässt ein stärkeres Teacher-Modell zusätzliche Trainingsbeispiele produzieren. Retentions-Horizonte setzen DSGVO-konforme TTLs für Audit-Log und ephemere Artefakte. Nichts davon ist Pflicht — eingeklappt lassen wenn Ihr Run sie nicht braucht.",
+    "wizard.step.operations.desc": "Benachrichtigungen, Metadaten externer Überwachungspläne, synthetische Daten und Aufbewahrung. Konfigurieren Sie die benötigten ForgeLM-Optionen.",
+    "wizard.step.operations.tutorial": "Diese vier Blöcke decken alles ab, was ForgeLM jenseits des Kern-Training + Compliance-Flows liefert. Webhook sendet Slack/Teams/Discord-Alerts bei Trainingsstart/Erfolg/Fehler. Monitoring (Art. 72 / 17(1)(h)) erfasst einen externen Plan als Metadaten; ForgeLM exportiert keine Metriken, erkennt keine Drift, sendet keine Drift-Warnungen und plant keine Kontrollen. Synthetische Datengenerierung lässt ein stärkeres Teacher-Modell zusätzliche Trainingsbeispiele produzieren. Retentions-Horizonte setzen DSGVO-konforme TTLs für Audit-Log und ephemere Artefakte. Diese ForgeLM-Einstellungen sind optional; externe gesetzliche Pflichten gelten weiterhin.",
     "wizard.risk.label":          "EU-AI-Act-Risikoklassifizierung",
     "wizard.risk.hint":           "Wählen Sie die Klasse, unter die Ihr Deployment fällt. High-risk und unacceptable lösen verpflichtende Audit + Human-Approval-Pflichten aus.",
     "wizard.risk.unknown":        "Unknown — Vorklassifizierungs-Platzhalter",
@@ -2565,25 +2571,25 @@
     "wizard.webhook.url.detail": "URL kann direkt eingefügt werden, oder via env:-Präfix auf eine Umgebungsvariable verweisen. Env-Variante hält die URL aus YAML + Audit-Log.",
     "wizard.webhook.enabled.label": "Webhook-Benachrichtigungen senden",
     "wizard.webhook.enabled.hint": "Wenn aktiv, sendet ForgeLM eine Webhook-Payload bei den ausgewählten Events. Aus = die YAML enthält keinen Webhook-Block.",
-    "wizard.webhook.enabled.detail": "Benachrichtigungen leiten Trainings-Events in Ihren Team-Chat (Slack / Teams / Discord) oder einen HTTPS-Endpoint, der JSON akzeptiert. Audit-Log-Einträge dokumentieren jeden Versand — auch Fehler — als unabhängige Spur. Aus lassen, wenn die YAML keinerlei Webhook-Config tragen soll.",
+    "wizard.webhook.enabled.detail": "Benachrichtigungen leiten Trainings-Events in Ihren Team-Chat (Slack / Teams / Discord) oder einen HTTPS-Endpoint, der JSON akzeptiert. Zustellfehler erscheinen nur als Warnungen im Prozesslog. Das Audit-Log erfasst die zugrunde liegenden Trainingsereignisse, keine Zustellversuche. Aus lassen, wenn die YAML keinerlei Webhook-Config tragen soll.",
     "wizard.webhook.notify_start.label": "Bei Trainingsstart benachrichtigen",
     "wizard.webhook.notify_success.label": "Bei Erfolg benachrichtigen",
     "wizard.webhook.notify_failure.label": "Bei Fehler benachrichtigen",
-    "wizard.monitoring.title": "Post-Market-Monitoring (Art. 12+17)",
-    "wizard.monitoring.hint": "Trainings-Metriken in Prometheus/Datadog/Custom-Webhook für laufende Observability.",
-    "wizard.monitoring.enabled.label": "Monitoring-Hooks aktivieren",
-    "wizard.monitoring.enabled.hint": "Pflicht für High-risk-Systeme nach EU AI Act Artikel 12 + 17.",
-    "wizard.monitoring.enabled.detail": "Nach Deployment eines High-risk-Systems verlangt der EU AI Act laufende Beobachtung — Drift-Detection, Fehler-Reporting, periodische Genauigkeitschecks.",
-    "wizard.monitoring.endpoint.label": "Monitoring-Endpoint",
-    "wizard.monitoring.endpoint.hint": "URL Ihres Prometheus-Push-Gateways, Datadogs oder Custom-Webhook-Receivers.",
-    "wizard.monitoring.endpoint.placeholder": "https://prometheus.example.com/metrics ODER env:DATADOG_URL",
-    "wizard.monitoring.metrics_export.label": "Metrik-Format",
+    "wizard.monitoring.title": "Externer Monitoring-Plan (Art. 72 / 17(1)(h))",
+    "wizard.monitoring.hint": "Externe Überwachung im Compliance-Manifest / Annex IV dokumentieren. Nur Metadaten; ForgeLM führt keine Überwachung aus.",
+    "wizard.monitoring.enabled.label": "Monitoring-Plan als Metadaten erfassen",
+    "wizard.monitoring.enabled.hint": "Die Metadatenerfassung in ForgeLM ist optional. Art. 72 verpflichtet Anbieter von Hochrisiko-KI-Systemen, ein System zur Überwachung nach dem Inverkehrbringen einzurichten und zu dokumentieren; Annex IV Punkt 9 umfasst den detaillierten Plan.",
+    "wizard.monitoring.enabled.detail": "ForgeLM erfasst den Plan mit declarative_only: true. Es kontaktiert keinen Endpoint, exportiert keine Metriken, erkennt keine Drift, sendet keine Drift-Warnungen und plant keine Kontrollen. Betreiben Sie Ihren externen Stack separat; die Erklärung allein erfüllt keine Pflicht zur Überwachung nach dem Inverkehrbringen.",
+    "wizard.monitoring.endpoint.label": "Deklarierter Endpoint (optional)",
+    "wizard.monitoring.endpoint.hint": "http(s)://host oder env:NAME verwenden; nie kontaktiert. Belege enthalten nur Schema/Host/Port, aber kopiertes YAML, Vorschau und lokaler Speicher behalten die vollständige URL. Daher keine Zugangsdaten eingeben. env:NAME bevorzugen: nur $NAME wird erfasst, nie der Wert.",
+    "wizard.monitoring.endpoint.placeholder": "https://monitor.example.com ODER env:MONITORING_URL",
+    "wizard.monitoring.metrics_export.label": "Deklarierter externer Exporter",
     "wizard.monitoring.metrics_export.none": "Keine",
     "wizard.monitoring.metrics_export.prometheus": "Prometheus",
     "wizard.monitoring.metrics_export.datadog": "Datadog",
     "wizard.monitoring.metrics_export.custom": "Custom Webhook",
-    "wizard.monitoring.alert_on_drift.label": "Bei Drift alerten",
-    "wizard.monitoring.alert_on_drift.hint": "Webhook-Alert wenn Drift-Detector eine Regression markiert.",
+    "wizard.monitoring.alert_on_drift.label": "Externe Drift-Warnungen deklarieren",
+    "wizard.monitoring.alert_on_drift.hint": "Erfasst die Richtlinie Ihres externen Stacks. ForgeLM erkennt keine Drift und sendet diese Warnungen nicht.",
     "wizard.synthetic.title": "Synthetische Datengenerierung (erweitert)",
     "wizard.synthetic.hint": "Teacher → Student Trainingsdaten mit stärkerem Modell generieren.",
     "wizard.synthetic.enabled.label": "Synthetische Daten generieren",
@@ -2624,7 +2630,6 @@
     "wizard.error.model.format": "HuggingFace-Hub-IDs verwenden das Format org/repo (z. B. meta-llama/Llama-3.1-8B-Instruct). Ein Name ohne Schrägstrich wird zur Trainingszeit nicht aufgelöst.",
     "wizard.error.dataset.required": "Wählen Sie ein Dataset, bevor Sie fortfahren — ohne Pfad oder HuggingFace-Namen hat der Trainer nichts zum Lernen.",
     "wizard.error.webhook.required": "Sie haben Benachrichtigungen aktiviert — wir brauchen also eine Webhook-URL oder Env-Variable. Schalten Sie Benachrichtigungen oben aus oder fügen Sie Ihre Slack-/Teams-/Discord-URL ein (für Produktion env:SLACK_WEBHOOK_URL).",
-    "wizard.error.monitoring.endpoint.required": "Sie haben Post-Market-Monitoring aktiviert — wir brauchen einen Endpoint für Metriken. Schalten Sie Monitoring oben aus oder fügen Sie Ihre Prometheus-/Datadog-/Custom-Webhook-URL ein (für Secrets env:VAR_NAME).",
     "wizard.error.synthetic.teacher_model.required": "Sie haben synthetische Datengenerierung aktiviert — wir brauchen einen Teacher-Modellnamen (z. B. gpt-4o, claude-opus-4-7 oder lokaler Pfad).",
     "wizard.error.synthetic.seed_file.required": "Sie haben synthetische Datengenerierung aktiviert — wir brauchen eine Seed-Prompts-Datei (z. B. ./data/seed_prompts.jsonl); der Teacher generiert pro Prompt eine Antwort.",
     "wizard.error.synthetic.api_key_env.required": "Sie haben synthetische Datengenerierung aktiviert — wir brauchen die Env-Variable mit dem Teacher-API-Key (üblich: OPENAI_API_KEY oder ANTHROPIC_API_KEY).",
@@ -2862,8 +2867,8 @@
     "wizard.step.compliance.tutorial": "ForgeLM vérifie le modèle fine-tuné AVANT que vous le livriez. L'auto-revert annule le fine-tune s'il est pire que le modèle de base. Llama Guard scanne les sorties pour du contenu nuisible. Le niveau de risque EU AI Act que vous choisissez détermine quels artefacts de conformité sont auto-générés et si l'approbation humaine est REQUISE avant déploiement.",
     "wizard.step.operations.title": "Opérations et paramètres avancés",
     "wizard.step.operations.name": "Opérations",
-    "wizard.step.operations.desc": "Notifications, surveillance post-marché, génération de données synthétiques et horizons de rétention. Tous optionnels — passez ceux dont vous n'avez pas besoin.",
-    "wizard.step.operations.tutorial": "Ces quatre blocs couvrent tout ce que ForgeLM offre au-delà du flux core entraînement + compliance. Webhook envoie une alerte Slack/Teams/Discord au démarrage / succès / échec. La surveillance post-marché (Articles 12+17) connecte la détection de dérive à Prometheus / Datadog après déploiement. La génération synthétique laisse un modèle teacher plus fort produire des exemples supplémentaires. Les horizons de rétention fixent des TTL conformes RGPD pour audit log et artefacts éphémères. Aucun n'est obligatoire — laissez replié ce dont votre run n'a pas besoin.",
+    "wizard.step.operations.desc": "Notifications, métadonnées du plan de surveillance externe, données synthétiques et conservation. Configurez les options ForgeLM nécessaires.",
+    "wizard.step.operations.tutorial": "Ces quatre blocs couvrent tout ce que ForgeLM offre au-delà du flux core entraînement + compliance. Webhook envoie une alerte Slack/Teams/Discord au démarrage / succès / échec. La surveillance (Art. 72 / 17(1)(h)) enregistre un plan externe comme métadonnées ; ForgeLM n’exporte pas de métriques, ne détecte pas de dérive, n’envoie pas d’alertes de dérive et ne programme aucun contrôle. La génération synthétique laisse un modèle teacher plus fort produire des exemples supplémentaires. Les horizons de rétention fixent des TTL conformes RGPD pour audit log et artefacts éphémères. Ces paramètres ForgeLM sont facultatifs ; les obligations légales externes restent applicables.",
     "wizard.risk.label":          "Classification de risque EU AI Act",
     "wizard.risk.hint":           "Choisissez le niveau dans lequel votre déploiement tombe. High-risk et unacceptable déclenchent des obligations d'audit + approbation humaine obligatoires.",
     "wizard.risk.unknown":        "Unknown — placeholder de pré-classification",
@@ -2960,25 +2965,25 @@
     "wizard.webhook.url.detail": "Vous pouvez coller URL directement, ou utiliser préfixe env: pour pointer une variable env.",
     "wizard.webhook.enabled.label": "Envoyer des notifications webhook",
     "wizard.webhook.enabled.hint": "Quand activé, ForgeLM publie un payload webhook aux événements sélectionnés. Désactivé = la YAML ne contient aucun bloc webhook.",
-    "wizard.webhook.enabled.detail": "Les notifications connectent les événements d'entraînement à votre chat d'équipe (Slack / Teams / Discord) ou tout endpoint HTTPS acceptant du JSON. Les entrées de l'audit log enregistrent chaque envoi — y compris les échecs — pour une piste indépendante. Laissez désactivé si vous ne voulez aucune config webhook dans la YAML.",
+    "wizard.webhook.enabled.detail": "Les notifications connectent les événements d'entraînement à votre chat d'équipe (Slack / Teams / Discord) ou tout endpoint HTTPS acceptant du JSON. Les échecs de livraison sont uniquement des avertissements dans le journal du processus. L’audit log enregistre les événements d’entraînement sous-jacents, pas les tentatives de livraison. Laissez désactivé si vous ne voulez aucune config webhook dans la YAML.",
     "wizard.webhook.notify_start.label": "Notifier au démarrage",
     "wizard.webhook.notify_success.label": "Notifier au succès",
     "wizard.webhook.notify_failure.label": "Notifier à l'échec",
-    "wizard.monitoring.title": "Surveillance post-marché (Art. 12+17)",
-    "wizard.monitoring.hint": "Branchez métriques d'entraînement à Prometheus/Datadog/webhook custom.",
-    "wizard.monitoring.enabled.label": "Activer hooks monitoring",
-    "wizard.monitoring.enabled.hint": "Obligatoire pour systèmes high-risk selon EU AI Act Articles 12 + 17.",
-    "wizard.monitoring.enabled.detail": "Une fois système high-risk déployé, EU AI Act exige observation continue — détection de drift, reporting d'erreurs.",
-    "wizard.monitoring.endpoint.label": "Endpoint monitoring",
-    "wizard.monitoring.endpoint.hint": "URL de votre Prometheus push gateway, Datadog ou récepteur webhook custom.",
-    "wizard.monitoring.endpoint.placeholder": "https://prometheus.example.com/metrics OU env:DATADOG_URL",
-    "wizard.monitoring.metrics_export.label": "Format métriques",
+    "wizard.monitoring.title": "Plan de surveillance externe (Art. 72 / 17(1)(h))",
+    "wizard.monitoring.hint": "Déclarez votre surveillance externe dans le manifeste de conformité / Annex IV. Métadonnées uniquement ; ForgeLM n’exécute aucune surveillance.",
+    "wizard.monitoring.enabled.label": "Enregistrer les métadonnées du plan",
+    "wizard.monitoring.enabled.hint": "L'enregistrement dans ForgeLM est facultatif. L'Art. 72 impose aux fournisseurs de systèmes d'IA à haut risque d'établir et de documenter un système de surveillance après commercialisation ; Annex IV, point 9, comprend le plan détaillé.",
+    "wizard.monitoring.enabled.detail": "ForgeLM enregistre le plan avec declarative_only: true. Il ne contacte pas l'endpoint, n’exporte pas de métriques, ne détecte pas de dérive, n’envoie pas d’alertes de dérive et ne programme aucun contrôle. Exploitez votre infrastructure externe séparément ; cette déclaration seule ne remplit pas une obligation de surveillance après commercialisation.",
+    "wizard.monitoring.endpoint.label": "Endpoint déclaré (facultatif)",
+    "wizard.monitoring.endpoint.hint": "Utilisez http(s)://host ou env:NAME ; jamais contacté. Les preuves ne conservent que schéma/hôte/port, mais le YAML copié, l'aperçu et le stockage local gardent l'URL complète. N'y mettez donc jamais d'identifiants. Préférez env:NAME : seul $NAME est enregistré, jamais sa valeur.",
+    "wizard.monitoring.endpoint.placeholder": "https://monitor.example.com OU env:MONITORING_URL",
+    "wizard.monitoring.metrics_export.label": "Outil d’export externe déclaré",
     "wizard.monitoring.metrics_export.none": "Aucun",
     "wizard.monitoring.metrics_export.prometheus": "Prometheus",
     "wizard.monitoring.metrics_export.datadog": "Datadog",
     "wizard.monitoring.metrics_export.custom": "Webhook custom",
-    "wizard.monitoring.alert_on_drift.label": "Alerte sur drift",
-    "wizard.monitoring.alert_on_drift.hint": "Émettre alert webhook quand détecteur drift signale régression.",
+    "wizard.monitoring.alert_on_drift.label": "Déclarer les alertes de dérive externes",
+    "wizard.monitoring.alert_on_drift.hint": "Enregistre la politique de votre infrastructure externe. ForgeLM ne détecte pas de dérive et n’envoie pas ces alertes.",
     "wizard.synthetic.title": "Génération données synthétiques (avancé)",
     "wizard.synthetic.hint": "Générez données entraînement teacher → student avec modèle plus fort.",
     "wizard.synthetic.enabled.label": "Générer données synthétiques",
@@ -3019,7 +3024,6 @@
     "wizard.error.model.format": "Les ID HuggingFace Hub utilisent le format org/repo (ex. meta-llama/Llama-3.1-8B-Instruct). Un nom sans barre oblique ne sera pas résolu lors de l’entraînement.",
     "wizard.error.dataset.required": "Choisissez un dataset avant de continuer — sans chemin ni nom HuggingFace, le trainer n’a rien à apprendre.",
     "wizard.error.webhook.required": "Vous avez activé les notifications — il nous faut donc une URL webhook ou une variable d’env. Désactivez les notifications ci-dessus ou collez votre URL Slack / Teams / Discord (pour la production utilisez env:SLACK_WEBHOOK_URL).",
-    "wizard.error.monitoring.endpoint.required": "Vous avez activé la surveillance post-marché — il nous faut un endpoint pour les métriques. Désactivez la surveillance ci-dessus ou collez votre URL Prometheus / Datadog / webhook custom (pour les secrets utilisez env:VAR_NAME).",
     "wizard.error.synthetic.teacher_model.required": "Vous avez activé la génération de données synthétiques — il nous faut un nom de modèle teacher (ex. gpt-4o, claude-opus-4-7 ou chemin local).",
     "wizard.error.synthetic.seed_file.required": "Vous avez activé la génération de données synthétiques — il nous faut un fichier de seed prompts (ex. ./data/seed_prompts.jsonl) ; le teacher produira une réponse par prompt.",
     "wizard.error.synthetic.api_key_env.required": "Vous avez activé la génération de données synthétiques — il nous faut la variable d’env qui porte la clé API du teacher (généralement OPENAI_API_KEY ou ANTHROPIC_API_KEY).",
@@ -3257,8 +3261,8 @@
     "wizard.step.compliance.tutorial": "ForgeLM verifica el modelo fine-tuned ANTES de que lo entregue. Auto-revert deshace el fine-tune si es peor que el modelo base. Llama Guard escanea las salidas en busca de contenido dañino. El nivel de riesgo EU AI Act que elija determina qué artefactos de cumplimiento se autogeneran y si la aprobación humana es REQUERIDA antes del despliegue.",
     "wizard.step.operations.title": "Operaciones y ajustes avanzados",
     "wizard.step.operations.name": "Operaciones",
-    "wizard.step.operations.desc": "Notificaciones, monitoreo post-mercado, generación de datos sintéticos y horizontes de retención. Todos opcionales — salte los que no necesite.",
-    "wizard.step.operations.tutorial": "Estos cuatro bloques cubren todo lo que ForgeLM ofrece más allá del flujo central de entrenamiento + compliance. Webhook envía alerta Slack/Teams/Discord en inicio / éxito / fallo. El monitoreo post-mercado (Artículos 12+17) conecta detección de drift a Prometheus / Datadog tras despliegue. La generación sintética permite a un modelo teacher más fuerte producir ejemplos adicionales. Los horizontes de retención fijan TTLs conformes a RGPD para audit log y artefactos efímeros. Ninguno es obligatorio — déjelos cerrados si su run no los necesita.",
+    "wizard.step.operations.desc": "Notificaciones, metadatos del plan de monitoreo externo, datos sintéticos y retención. Configure las opciones de ForgeLM que necesite.",
+    "wizard.step.operations.tutorial": "Estos cuatro bloques cubren todo lo que ForgeLM ofrece más allá del flujo central de entrenamiento + compliance. Webhook envía alerta Slack/Teams/Discord en inicio / éxito / fallo. El monitoreo (Art. 72 / 17(1)(h)) registra un plan externo como metadatos; ForgeLM no exporta métricas, no detecta deriva, no envía alertas de deriva ni programa comprobaciones. La generación sintética permite a un modelo teacher más fuerte producir ejemplos adicionales. Los horizontes de retención fijan TTLs conformes a RGPD para audit log y artefactos efímeros. Estas opciones de ForgeLM son opcionales; las obligaciones legales externas siguen vigentes.",
     "wizard.risk.label":          "Clasificación de riesgo EU AI Act",
     "wizard.risk.hint":           "Elija el nivel en el que cae su despliegue. High-risk y unacceptable activan obligaciones obligatorias de auditoría + aprobación humana.",
     "wizard.risk.unknown":        "Unknown — placeholder de pre-clasificación",
@@ -3355,25 +3359,25 @@
     "wizard.webhook.url.detail": "Puede pegar URL directamente, o usar prefijo env: para apuntar a una variable de entorno.",
     "wizard.webhook.enabled.label": "Enviar notificaciones webhook",
     "wizard.webhook.enabled.hint": "Cuando está activo, ForgeLM publica un payload webhook en los eventos seleccionados. Apagado = la YAML no contiene ningún bloque webhook.",
-    "wizard.webhook.enabled.detail": "Las notificaciones conectan eventos de entrenamiento a su chat de equipo (Slack / Teams / Discord) o cualquier endpoint HTTPS que acepte JSON. Las entradas del audit log registran cada envío — incluyendo fallos — como un rastro independiente. Déjelo apagado si no quiere ninguna config de webhook en la YAML.",
+    "wizard.webhook.enabled.detail": "Las notificaciones conectan eventos de entrenamiento a su chat de equipo (Slack / Teams / Discord) o cualquier endpoint HTTPS que acepte JSON. Los fallos de entrega solo se registran como advertencias en el registro del proceso. El audit log registra los eventos de entrenamiento subyacentes, no los intentos de entrega. Déjelo apagado si no quiere ninguna config de webhook en la YAML.",
     "wizard.webhook.notify_start.label": "Notificar al inicio",
     "wizard.webhook.notify_success.label": "Notificar en éxito",
     "wizard.webhook.notify_failure.label": "Notificar en fallo",
-    "wizard.monitoring.title": "Monitoreo post-mercado (Art. 12+17)",
-    "wizard.monitoring.hint": "Conecte métricas entrenamiento a Prometheus/Datadog/webhook custom para observabilidad continua.",
-    "wizard.monitoring.enabled.label": "Habilitar hooks de monitoreo",
-    "wizard.monitoring.enabled.hint": "Obligatorio para sistemas high-risk bajo EU AI Act Artículos 12 + 17.",
-    "wizard.monitoring.enabled.detail": "Una vez desplegado un sistema high-risk, EU AI Act requiere observación continua — detección de drift, reporte de errores.",
-    "wizard.monitoring.endpoint.label": "Endpoint monitoreo",
-    "wizard.monitoring.endpoint.hint": "URL de su Prometheus push gateway, Datadog o receptor webhook custom.",
-    "wizard.monitoring.endpoint.placeholder": "https://prometheus.example.com/metrics O env:DATADOG_URL",
-    "wizard.monitoring.metrics_export.label": "Formato métricas",
+    "wizard.monitoring.title": "Plan de monitoreo externo (Art. 72 / 17(1)(h))",
+    "wizard.monitoring.hint": "Declare su monitoreo externo en el manifiesto de cumplimiento / Annex IV. Solo metadatos; ForgeLM no ejecuta el monitoreo.",
+    "wizard.monitoring.enabled.label": "Registrar metadatos del plan",
+    "wizard.monitoring.enabled.hint": "El registro en ForgeLM es opcional. El Art. 72 exige a los proveedores de sistemas de IA de alto riesgo establecer y documentar un sistema de monitoreo posterior a la comercialización; Annex IV, punto 9, incluye el plan detallado.",
+    "wizard.monitoring.enabled.detail": "ForgeLM registra el plan con declarative_only: true. No contacta el endpoint, no exporta métricas, no detecta deriva, no envía alertas de deriva ni programa comprobaciones. Opere su infraestructura externa por separado; esta declaración por sí sola no cumple una obligación de monitoreo posterior a la comercialización.",
+    "wizard.monitoring.endpoint.label": "Endpoint declarado (opcional)",
+    "wizard.monitoring.endpoint.hint": "Use http(s)://host o env:NAME; nunca se contacta. La evidencia conserva solo esquema/host/puerto, pero el YAML copiado, la vista previa y el almacenamiento local conservan la URL completa. Nunca incluya credenciales. Prefiera env:NAME: solo se registra $NAME, nunca su valor.",
+    "wizard.monitoring.endpoint.placeholder": "https://monitor.example.com O env:MONITORING_URL",
+    "wizard.monitoring.metrics_export.label": "Exportador externo declarado",
     "wizard.monitoring.metrics_export.none": "Ninguno",
     "wizard.monitoring.metrics_export.prometheus": "Prometheus",
     "wizard.monitoring.metrics_export.datadog": "Datadog",
     "wizard.monitoring.metrics_export.custom": "Webhook custom",
-    "wizard.monitoring.alert_on_drift.label": "Alertar en drift",
-    "wizard.monitoring.alert_on_drift.hint": "Emite alerta webhook cuando detector drift marca regresión.",
+    "wizard.monitoring.alert_on_drift.label": "Declarar alertas externas de deriva",
+    "wizard.monitoring.alert_on_drift.hint": "Registra la política de su infraestructura externa. ForgeLM no detecta deriva ni envía estas alertas.",
     "wizard.synthetic.title": "Generación datos sintéticos (avanzado)",
     "wizard.synthetic.hint": "Genere datos entrenamiento teacher → student usando modelo más fuerte.",
     "wizard.synthetic.enabled.label": "Generar datos sintéticos",
@@ -3414,7 +3418,6 @@
     "wizard.error.model.format": "Los ID de HuggingFace Hub usan el formato org/repo (ej. meta-llama/Llama-3.1-8B-Instruct). Un nombre sin barra no se resolverá al entrenar.",
     "wizard.error.dataset.required": "Elija un dataset antes de continuar — sin una ruta o nombre de HuggingFace, el trainer no tiene de qué aprender.",
     "wizard.error.webhook.required": "Activó las notificaciones, así que necesitamos una URL de webhook o variable de entorno. Desactive las notificaciones arriba o pegue su URL de Slack / Teams / Discord (para producción use env:SLACK_WEBHOOK_URL).",
-    "wizard.error.monitoring.endpoint.required": "Activó el monitoreo post-mercado, así que necesitamos un endpoint para enviar métricas. Desactive el monitoreo arriba o pegue su URL de Prometheus / Datadog / webhook custom (para secretos use env:NOMBRE_VAR).",
     "wizard.error.synthetic.teacher_model.required": "Activó la generación de datos sintéticos, así que necesitamos un nombre de modelo teacher (ej. gpt-4o, claude-opus-4-7 o ruta local).",
     "wizard.error.synthetic.seed_file.required": "Activó la generación de datos sintéticos, así que necesitamos un archivo de seed prompts (ej. ./data/seed_prompts.jsonl); el teacher producirá una respuesta por cada prompt.",
     "wizard.error.synthetic.api_key_env.required": "Activó la generación de datos sintéticos, así que necesitamos la variable de entorno con la clave API del teacher (típicamente OPENAI_API_KEY o ANTHROPIC_API_KEY).",
@@ -3652,8 +3655,8 @@
     "wizard.step.compliance.tutorial": "ForgeLM 在你交付之前检查微调后的模型。Auto-revert 在微调比基础模型差时撤销微调。Llama Guard 扫描输出中的有害内容。你选择的 EU AI Act 风险等级决定哪些合规 artefact 自动生成,以及在部署之前是否需要人工批准。",
     "wizard.step.operations.title": "运维与高级设置",
     "wizard.step.operations.name": "运维",
-    "wizard.step.operations.desc": "通知、上市后监控、合成数据生成和保留期限。全部可选 — 不需要的可以跳过。",
-    "wizard.step.operations.tutorial": "这四个区块涵盖了 ForgeLM 在核心训练 + 合规流程之外提供的一切。Webhook 在训练开始 / 成功 / 失败时发送 Slack/Teams/Discord 警报。上市后监控(第 12+17 条)在部署后将漂移检测连接到 Prometheus / Datadog。合成数据生成让更强的 teacher 模型产生额外训练样本。保留期限为审计日志和临时 artefact 设置符合 GDPR 的 TTL。没有任何项是强制的 — 你的训练不需要的话保持折叠状态即可。",
+    "wizard.step.operations.desc": "通知、外部监控计划元数据、合成数据和保留设置。配置所需的 ForgeLM 选项。",
+    "wizard.step.operations.tutorial": "这四个区块涵盖了 ForgeLM 在核心训练 + 合规流程之外提供的一切。Webhook 在训练开始 / 成功 / 失败时发送 Slack/Teams/Discord 警报。监控（第 72 条及第 17(1)(h) 条）仅将外部计划记录为元数据；ForgeLM 不导出指标、检测漂移、发送漂移警报或安排检查。合成数据生成让更强的 teacher 模型产生额外训练样本。保留期限为审计日志和临时 artefact 设置符合 GDPR 的 TTL。这些 ForgeLM 设置是可选的；外部法律义务仍然适用。",
     "wizard.risk.label":          "EU AI Act 风险分类",
     "wizard.risk.hint":           "选择你的部署所属的等级。High-risk 和 unacceptable 触发强制审计 + 人工批准义务。",
     "wizard.risk.unknown":        "Unknown — 预分类占位符",
@@ -3750,25 +3753,25 @@
     "wizard.webhook.url.detail": "你可以直接粘贴 URL,或使用 env: 前缀指向环境变量。",
     "wizard.webhook.enabled.label": "发送 webhook 通知",
     "wizard.webhook.enabled.hint": "启用后,ForgeLM 在你选择的事件发送 webhook payload。关闭 = YAML 不包含任何 webhook 块。",
-    "wizard.webhook.enabled.detail": "通知将训练事件连接到你的团队聊天(Slack / Teams / Discord)或任何接受 JSON 的 HTTPS endpoint。审计日志条目记录每次发送 — 包括失败 — 提供独立追踪。如果不想 YAML 中包含任何 webhook 配置,保持关闭。",
+    "wizard.webhook.enabled.detail": "通知将训练事件连接到你的团队聊天(Slack / Teams / Discord)或任何接受 JSON 的 HTTPS endpoint。发送失败仅在进程日志中记录为警告。审计日志记录对应的训练事件，不记录发送尝试。如果不想 YAML 中包含任何 webhook 配置,保持关闭。",
     "wizard.webhook.notify_start.label": "训练开始时通知",
     "wizard.webhook.notify_success.label": "成功时通知",
     "wizard.webhook.notify_failure.label": "失败时通知",
-    "wizard.monitoring.title": "上市后监控(第 12+17 条)",
-    "wizard.monitoring.hint": "将训练指标连接到 Prometheus/Datadog/自定义 webhook 以持续观测。",
-    "wizard.monitoring.enabled.label": "启用监控 hook",
-    "wizard.monitoring.enabled.hint": "EU AI Act 第 12 + 17 条要求 high-risk 系统强制启用。",
-    "wizard.monitoring.enabled.detail": "high-risk 系统部署后,EU AI Act 要求持续观察 — 漂移检测、错误报告、定期准确性复查。",
-    "wizard.monitoring.endpoint.label": "监控 endpoint",
-    "wizard.monitoring.endpoint.hint": "你的 Prometheus push gateway、Datadog 或自定义 webhook 接收器的 URL。",
-    "wizard.monitoring.endpoint.placeholder": "https://prometheus.example.com/metrics 或 env:DATADOG_URL",
-    "wizard.monitoring.metrics_export.label": "指标格式",
+    "wizard.monitoring.title": "外部监控计划（第 72 条及第 17(1)(h) 条）",
+    "wizard.monitoring.hint": "在合规清单 / Annex IV 中声明外部监控设置。仅记录元数据；ForgeLM 不运行监控。",
+    "wizard.monitoring.enabled.label": "记录监控计划元数据",
+    "wizard.monitoring.enabled.hint": "在 ForgeLM 中记录元数据是可选的。第 72 条要求高风险 AI 系统提供者建立上市后监控系统并形成文档；Annex IV 第 9 点包括详细计划。",
+    "wizard.monitoring.enabled.detail": "ForgeLM 使用 declarative_only: true 标记记录计划。它不会连接端点、导出指标、检测漂移、发送漂移警报或安排检查。请单独运行外部监控系统；仅此声明并不能履行上市后监控义务。",
+    "wizard.monitoring.endpoint.label": "声明的端点（可选）",
+    "wizard.monitoring.endpoint.hint": "使用 http(s)://host 或 env:NAME；不会连接。证据仅保留协议、主机和端口，但复制的 YAML、预览和本地存储保留完整 URL，因此切勿输入凭据。优先使用 env:NAME：只记录 $NAME，不读取变量值。",
+    "wizard.monitoring.endpoint.placeholder": "https://monitor.example.com 或 env:MONITORING_URL",
+    "wizard.monitoring.metrics_export.label": "声明的外部指标导出器",
     "wizard.monitoring.metrics_export.none": "无",
     "wizard.monitoring.metrics_export.prometheus": "Prometheus",
     "wizard.monitoring.metrics_export.datadog": "Datadog",
     "wizard.monitoring.metrics_export.custom": "自定义 webhook",
-    "wizard.monitoring.alert_on_drift.label": "漂移时报警",
-    "wizard.monitoring.alert_on_drift.hint": "当漂移检测器标记回归时发出 webhook 警报。",
+    "wizard.monitoring.alert_on_drift.label": "声明外部漂移警报策略",
+    "wizard.monitoring.alert_on_drift.hint": "记录外部系统的策略。ForgeLM 不检测漂移，也不发送这些警报。",
     "wizard.synthetic.title": "合成数据生成(高级)",
     "wizard.synthetic.hint": "使用更强的模型生成 teacher → student 训练数据。",
     "wizard.synthetic.enabled.label": "生成合成数据",
@@ -3809,7 +3812,6 @@
     "wizard.error.model.format": "HuggingFace Hub ID 使用 org/repo 格式(如 meta-llama/Llama-3.1-8B-Instruct)。不带斜杠的纯名称在训练时无法解析。",
     "wizard.error.dataset.required": "继续之前请选择一个数据集 — 没有路径或 HuggingFace 名称,trainer 没有可学习的对象。",
     "wizard.error.webhook.required": "你启用了通知,所以需要一个 webhook URL 或环境变量。要么关闭上方的通知,要么粘贴你的 Slack / Teams / Discord URL(生产环境密钥请使用 env:SLACK_WEBHOOK_URL)。",
-    "wizard.error.monitoring.endpoint.required": "你启用了上市后监控,所以需要一个推送指标的 endpoint。要么关闭上方的监控,要么粘贴你的 Prometheus / Datadog / 自定义 webhook URL(密钥请使用 env:变量名)。",
     "wizard.error.synthetic.teacher_model.required": "你启用了合成数据生成,所以需要一个 teacher 模型名(如 gpt-4o、claude-opus-4-7 或本地模型路径)。",
     "wizard.error.synthetic.seed_file.required": "你启用了合成数据生成,所以需要一个 seed prompts 文件(如 ./data/seed_prompts.jsonl);teacher 会为每个 prompt 生成一个回应。",
     "wizard.error.synthetic.api_key_env.required": "你启用了合成数据生成,所以需要承载 teacher API 密钥的环境变量(通常是 OPENAI_API_KEY 或 ANTHROPIC_API_KEY)。",
diff --git a/site/js/wizard.js b/site/js/wizard.js
index 87ffb806..687bd763 100644
--- a/site/js/wizard.js
+++ b/site/js/wizard.js
@@ -251,12 +251,12 @@
       webhookNotifyStart: false,
       webhookNotifySuccess: true,
       webhookNotifyFailure: true,
-      // Article 12+17 post-market monitoring. Auto-coerced to true
-      // for high-risk tiers.
+      // External monitoring-plan metadata; opt-in for every risk tier.
       monitoringEnabled: false,
       monitoringEndpoint: '',
       metricsExport: 'none',  // 'none' | 'prometheus' | 'datadog' | 'custom_webhook'
-      alertOnDrift: true,
+      alertOnDrift: false,
+      monitoringCheckIntervalHours: 24,
       // Synthetic data generation (teacher → student). Optional
       // pre-training stage; ``syntheticTeacherModel`` is the API
       // name or local path for the teacher.
@@ -362,6 +362,28 @@
     }
   }
 
+  // Shared ASCII env-reference syntax for monitoring metadata and webhook secrets.
+  function parseEnvRef(value) {
+    var raw = (value || '').trim();
+    if (raw.slice(0, 4).toLowerCase() !== 'env:') return { literal: raw };
+    var name = raw.slice(4).trim();
+    return /^[A-Za-z_][A-Za-z0-9_]*$/.test(name) ? { name: name } : { invalid: true };
+  }
+
+  function validMonitoringUrl(value) {
+    if (!value) return true;
+    var lower = value.toLowerCase();
+    if (!lower.startsWith('http://') && !lower.startsWith('https://')) return false;
+    for (var i = 0; i < value.length; i++) {
+      if (value.charCodeAt(i) <= 32 || value.charCodeAt(i) === 127 || value[i] === '\\') return false;
+    }
+    try {
+      return !!new URL(value).hostname;
+    } catch (e) {
+      return false;
+    }
+  }
+
   /* ── Required-field validation ────────────────────────── */
   // Some steps require operator input that has no sensible default —
   // e.g. a custom HuggingFace model ID, or a dataset path / Hub name.
@@ -447,9 +469,17 @@
       // Off → skip; the operator opted out of the whole feature.
       if (state.webhookEnabled && blank(state.webhookUrl)) {
         errors.push({ field: 'webhookUrl', key: 'wizard.error.webhook.required' });
+      } else if (state.webhookEnabled && parseEnvRef(state.webhookUrl).invalid) {
+        errors.push({ field: 'webhookUrl', key: 'wizard.error.env_ref.invalid' });
+      }
+      if (state.monitoringEnabled && parseEnvRef(state.monitoringEndpoint).invalid) {
+        errors.push({ field: 'monitoringEndpoint', key: 'wizard.error.env_ref.invalid' });
+      } else if (state.monitoringEnabled && !validMonitoringUrl(parseEnvRef(state.monitoringEndpoint).literal)) {
+        errors.push({ field: 'monitoringEndpoint', key: 'wizard.error.monitoring.endpoint.invalid' });
       }
-      if (state.monitoringEnabled && blank(state.monitoringEndpoint)) {
-        errors.push({ field: 'monitoringEndpoint', key: 'wizard.error.monitoring.endpoint.required' });
+      if (state.monitoringEnabled && (!Number.isInteger(state.monitoringCheckIntervalHours) ||
+          state.monitoringCheckIntervalHours < 1 || state.monitoringCheckIntervalHours > 720)) {
+        errors.push({ field: 'monitoringCheckIntervalHours', key: 'wizard.error.monitoring.interval.invalid' });
       }
       if (state.syntheticEnabled) {
         if (blank(state.syntheticTeacherModel)) {
@@ -806,11 +836,11 @@
     // is a secret.
     if (state.webhookEnabled && state.webhookUrl) {
       header(0, 'webhook');
-      var raw = state.webhookUrl.trim();
-      if (raw.indexOf('env:') === 0) {
-        k(1, 'url_env', JSON.stringify(raw.slice(4)));
-      } else {
-        k(1, 'url', JSON.stringify(raw));
+      var webhookRef = parseEnvRef(state.webhookUrl);
+      if (webhookRef.name) {
+        k(1, 'url_env', JSON.stringify(webhookRef.name));
+      } else if (!webhookRef.invalid) {
+        k(1, 'url', JSON.stringify(webhookRef.literal));
       }
       if (state.webhookNotifyStart)   k(1, 'notify_on_start', 'true');
       if (!state.webhookNotifySuccess) k(1, 'notify_on_success', 'false');
@@ -818,24 +848,25 @@
       blank();
     }
 
-    // Article 12+17 post-market monitoring. Auto-coerced enabled for
-    // strict tiers (operator can't disable it under high-risk).
-    var monitoringActive = state.monitoringEnabled || isStrictRisk(state);
-    if (monitoringActive || state.monitoringEndpoint) {
+    // Record an external plan without activating monitoring hooks.
+    if (state.monitoringEnabled) {
+      comment('External monitoring plan metadata only; ForgeLM runs no exporters, drift alerts, or scheduled checks.');
       header(0, 'monitoring');
-      k(1, 'enabled', monitoringActive ? 'true' : 'false');
+      k(1, 'enabled', 'true');
       if (state.monitoringEndpoint) {
-        var rawM = state.monitoringEndpoint.trim();
-        if (rawM.indexOf('env:') === 0) {
-          k(1, 'endpoint_env', JSON.stringify(rawM.slice(4)));
-        } else {
-          k(1, 'endpoint', JSON.stringify(rawM));
+        var monitoringRef = parseEnvRef(state.monitoringEndpoint);
+        if (monitoringRef.name) {
+          k(1, 'endpoint_env', JSON.stringify(monitoringRef.name));
+        } else if (!monitoringRef.invalid && monitoringRef.literal && validMonitoringUrl(monitoringRef.literal)) {
+          k(1, 'endpoint', JSON.stringify(monitoringRef.literal));
         }
       }
-      if (state.metricsExport && state.metricsExport !== 'none') {
-        k(1, 'metrics_export', JSON.stringify(state.metricsExport));
+      k(1, 'metrics_export', JSON.stringify(state.metricsExport || 'none'));
+      k(1, 'alert_on_drift', state.alertOnDrift ? 'true' : 'false');
+      if (Number.isInteger(state.monitoringCheckIntervalHours) && state.monitoringCheckIntervalHours >= 1 &&
+          state.monitoringCheckIntervalHours <= 720) {
+        k(1, 'check_interval_hours', String(state.monitoringCheckIntervalHours));
       }
-      if (!state.alertOnDrift) k(1, 'alert_on_drift', 'false');
       blank();
     }
 
@@ -1120,6 +1151,7 @@
       judgeApiKeyEnv:         'eval_gates',
       webhookUrl:             'webhook',
       monitoringEndpoint:     'monitoring',
+      monitoringCheckIntervalHours: 'monitoring',
       syntheticTeacherModel:  'synthetic',
       syntheticSeedFile:      'synthetic',
       syntheticApiKeyEnv:     'synthetic'
@@ -2701,26 +2733,17 @@
       body.appendChild(makeWebhookToggle('webhookNotifyFailure', 'wizard.webhook.notify_failure.label'));
     }));
 
-    // Post-market monitoring (Article 12+17) accordion.
+    // Declarative external monitoring plan (Article 72 / 17(1)(h)).
     pane.appendChild(makeAccordion(state, persist, 'monitoring',
       'wizard.monitoring.title', 'wizard.monitoring.hint', function (body) {
-      var monLocked = isStrictRisk(state);
       var monInput = el('input', { type: 'checkbox' });
-      monInput.checked = !!state.monitoringEnabled || monLocked;
-      if (monLocked) monInput.disabled = true;
+      monInput.checked = !!state.monitoringEnabled;
       monInput.addEventListener('change', function () {
         state.monitoringEnabled = monInput.checked;
         persist();
         rerender();
       });
       var monLabelEl = el('span', { class: 'wizard-toggle-label', text: tr('wizard.monitoring.enabled.label') });
-      if (monLocked) {
-        monLabelEl.appendChild(el('span', {
-          class: 'badge',
-          text: tr('wizard.locked.badge'),
-          style: 'margin-left: 0.4rem;'
-        }));
-      }
       var monRow = el('div', { class: 'wizard-row' }, [
         el('label', { class: 'wizard-toggle' }, [monInput, monLabelEl]),
         el('span', { class: 'wizard-row-hint', text: tr('wizard.monitoring.enabled.hint') })
@@ -2728,15 +2751,12 @@
       appendDetail(monRow, 'wizard.monitoring.enabled.detail');
       body.appendChild(monRow);
 
-      var endpointShown = !!state.monitoringEnabled || monLocked;
+      var endpointShown = !!state.monitoringEnabled;
       if (endpointShown) {
-        // Endpoint is required when monitoring is enabled — same
-        // pattern as webhook.url. fieldHasError gates the visual so
-        // the error appears only after a failed Next attempt.
-        var monEndpointErrKey = fieldHasError(state, 'operations', 'monitoringEndpoint');
+        var monEndpointErr = fieldHasError(state, 'operations', 'monitoringEndpoint');
         var epInput = el('input', {
           type: 'text',
-          class: 'wizard-input' + (monEndpointErrKey ? ' is-invalid' : ''),
+          class: 'wizard-input' + (monEndpointErr ? ' is-invalid' : ''),
           placeholder: tr('wizard.monitoring.endpoint.placeholder'),
           value: state.monitoringEndpoint || ''
         });
@@ -2745,23 +2765,43 @@
           epInput,
           el('span', { class: 'wizard-row-hint', text: tr('wizard.monitoring.endpoint.hint') })
         ];
-        if (monEndpointErrKey) {
-          epRowChildren.push(el('div', { class: 'wizard-error', role: 'alert', text: tr(monEndpointErrKey) }));
-        }
         var epRow = el('div', { class: 'wizard-row' }, epRowChildren);
+        if (monEndpointErr) epRow.appendChild(el('div', { class: 'wizard-error', role: 'alert', text: tr(monEndpointErr) }));
         epInput.addEventListener('input', function () {
           state.monitoringEndpoint = epInput.value.trim();
           persist();
           liveYaml();
-          var stillErr = validateStep(state, 'operations').some(function (e) { return e.field === 'monitoringEndpoint'; });
-          if (!stillErr) {
+          if (!fieldHasError(state, 'operations', 'monitoringEndpoint')) {
             epInput.classList.remove('is-invalid');
-            var errEl = epRow.querySelector('.wizard-error');
-            if (errEl) errEl.remove();
+            var error = epRow.querySelector('.wizard-error');
+            if (error) error.remove();
           }
         });
         body.appendChild(epRow);
 
+        var intervalError = fieldHasError(state, 'operations', 'monitoringCheckIntervalHours');
+        var intervalInput = el('input', {
+          id: 'wizard-monitoring-interval', type: 'number', min: '1', max: '720', step: '1',
+          class: 'wizard-input' + (intervalError ? ' is-invalid' : ''),
+          value: String(state.monitoringCheckIntervalHours)
+        });
+        intervalInput.addEventListener('input', function () {
+          state.monitoringCheckIntervalHours = Number(intervalInput.value);
+          persist();
+          liveYaml();
+          if (!fieldHasError(state, 'operations', 'monitoringCheckIntervalHours')) {
+            intervalInput.classList.remove('is-invalid');
+            var error = intervalRow.querySelector('.wizard-error');
+            if (error) error.remove();
+          }
+        });
+        var intervalRow = el('div', { class: 'wizard-row' }, [
+          el('label', { for: 'wizard-monitoring-interval', class: 'wizard-row-label', text: tr('wizard.monitoring.interval.label') }),
+          intervalInput
+        ]);
+        if (intervalError) intervalRow.appendChild(el('div', { class: 'wizard-error', role: 'alert', text: tr(intervalError) }));
+        body.appendChild(intervalRow);
+
         var meSelect = el('select', { class: 'wizard-select' });
         [
           { id: 'none',           labelKey: 'wizard.monitoring.metrics_export.none' },
diff --git a/tests/site/wizard_monitoring.test.cjs b/tests/site/wizard_monitoring.test.cjs
new file mode 100644
index 00000000..b7e866bb
--- /dev/null
+++ b/tests/site/wizard_monitoring.test.cjs
@@ -0,0 +1,210 @@
+'use strict';
+
+const assert = require('node:assert/strict');
+const fs = require('node:fs');
+const path = require('node:path');
+const { spawnSync } = require('node:child_process');
+const { test } = require('node:test');
+const vm = require('node:vm');
+
+const root = path.resolve(__dirname, '../..');
+const pythonBin = process.platform === 'win32' ? ['Scripts', 'python.exe'] : ['bin', 'python'];
+const venvPython = path.join(process.env.VIRTUAL_ENV || path.join(root, '.venv'), ...pythonBin);
+const python = process.env.FORGELM_TEST_PYTHON || (fs.existsSync(venvPython) ? venvPython : 'python3');
+function element(tag) {
+  const node = {
+    tag, children: [], attrs: {}, events: {},
+    appendChild(child) { this.children.push(child); child.parent = this; },
+    setAttribute(key, value) { this.attrs[key] = value; },
+    addEventListener(key, handler) { this.events[key] = handler; },
+    remove() { this.parent.children = this.parent.children.filter(child => child !== this); },
+    querySelector(selector) {
+      return descendants(this).find(child => (child.className || '').split(' ').includes(selector.slice(1))) || null;
+    }
+  };
+  node.classList = {
+    add() {}, toggle() {},
+    remove(value) { node.className = (node.className || '').split(' ').filter(item => item !== value).join(' '); }
+  };
+  return node;
+}
+function loadWizard() {
+  const document = {
+    readyState: 'loading', addEventListener() {}, querySelector() { return null; }, createElement: element,
+    createElementNS: (_, tag) => element(tag), createTextNode: text => ({ text })
+  };
+  const context = vm.createContext({ window: {}, document, URL });
+  for (const file of ['wizard_defaults.js', 'translations.js']) {
+    vm.runInContext(fs.readFileSync(path.join(root, 'site/js', file), 'utf8'), context);
+  }
+  const marker = '  /* ── Boot ─';
+  const source = fs.readFileSync(path.join(root, 'site/js/wizard.js'), 'utf8');
+  assert.equal(source.split(marker).length, 2, 'test seam must identify the boot boundary exactly once');
+  // Expose closure functions only in the VM; production has no test-only API.
+  vm.runInContext(source.replace(marker,
+    '  window.testWizard = {defaultState, buildYaml, validateStep, coerceForRisk, parseEnvRef, STEP_RENDERERS};\n' + marker), context);
+  return { ...context.window.testWizard, translations: context.window.ForgeLMTranslations };
+}
+function state(wizard, overrides = {}) {
+  return Object.assign(wizard.defaultState(), { model: 'org/model', datasetName: 'org/dataset' }, overrides);
+}
+function descendants(node) {
+  return [node, ...(node.children || []).flatMap(descendants)];
+}
+
+for (const tier of ['minimal-risk', 'limited-risk', 'high-risk', 'unacceptable']) {
+  test(`monitoring remains opt-in at ${tier}; empty endpoint is valid`, () => {
+    const wizard = loadWizard();
+    const value = state(wizard, { riskClassification: tier });
+    wizard.coerceForRisk(value);
+    assert.equal(value.monitoringEnabled, false);
+    assert.doesNotMatch(wizard.buildYaml(value), /\nmonitoring:/);
+    value.monitoringEnabled = true;
+    wizard.coerceForRisk(value);
+    assert.equal(value.monitoringEnabled, true, 'an enabled declaration must survive tier coercion');
+    assert.equal(value.alertOnDrift, false, 'drift alerts require an affirmative operator answer');
+    assert.equal(wizard.validateStep(value, 'operations').length, 0);
+    assert.match(wizard.buildYaml(value), /monitoring:\n  enabled: true/);
+    assert.doesNotMatch(wizard.buildYaml(value), /\n  endpoint(?:_env)?:/);
+  });
+}
+
+for (const [raw, expected] of [['env:MON_URL', 'MON_URL'], ['ENV: MON_URL ', 'MON_URL'], [' env:_mon_url ', '_mon_url']]) {
+  test(`both endpoint inputs normalize ${JSON.stringify(raw)}`, () => {
+    const wizard = loadWizard();
+    const value = state(wizard, { monitoringEnabled: true, monitoringEndpoint: raw, webhookEnabled: true, webhookUrl: raw });
+    assert.equal(wizard.validateStep(value, 'operations').length, 0);
+    const yaml = wizard.buildYaml(value);
+    assert.match(yaml, new RegExp(`endpoint_env: "${expected}"`));
+    assert.match(yaml, new RegExp(`url_env: "${expected}"`));
+  });
+}
+for (const raw of ['env:', 'env:   ', 'env:$MON_URL', 'env:https://host/token', 'env:MON URL', 'env:1MON', 'env:MÖN']) {
+  test(`both endpoint inputs reject ${JSON.stringify(raw)}`, () => {
+    const wizard = loadWizard();
+    const value = state(wizard, { monitoringEnabled: true, monitoringEndpoint: raw, webhookEnabled: true, webhookUrl: raw });
+    const fields = Array.from(wizard.validateStep(value, 'operations'), item => item.field);
+    assert.deepEqual(fields.sort(), ['monitoringEndpoint', 'webhookUrl']);
+    assert.doesNotMatch(wizard.buildYaml(value), /(?:endpoint_env|url_env):/);
+  });
+}
+
+test('disabled plans omit stale endpoint, exporter, and drift policy', () => {
+  const wizard = loadWizard();
+  const value = state(wizard, { monitoringEnabled: false, monitoringEndpoint: 'env:MON_URL', metricsExport: 'prometheus', alertOnDrift: true });
+  assert.doesNotMatch(wizard.buildYaml(value), /\nmonitoring:/);
+});
+
+for (const endpoint of ['prometheus.internal:9090', 'localhost:8080', 'monitor.example.com/x', 'https:host', 'https://host:65536', 'http://[::1', 'https://host/a b']) {
+  test(`malformed literal endpoint ${JSON.stringify(endpoint)} blocks advancement`, () => {
+    const wizard = loadWizard();
+    const value = state(wizard, { monitoringEnabled: true, monitoringEndpoint: endpoint });
+    assert.deepEqual(Array.from(wizard.validateStep(value, 'operations'), item => item.key), ['wizard.error.monitoring.endpoint.invalid']);
+    assert.doesNotMatch(wizard.buildYaml(value), /\n  endpoint:/);
+  });
+}
+
+for (const interval of [0, 721, 1.5, NaN]) {
+  test(`invalid external check interval ${interval} blocks advancement`, () => {
+    const wizard = loadWizard();
+    const value = state(wizard, { monitoringEnabled: true, monitoringCheckIntervalHours: interval });
+    assert.equal(wizard.validateStep(value, 'operations')[0].field, 'monitoringCheckIntervalHours');
+    assert.doesNotMatch(wizard.buildYaml(value), /check_interval_hours:/);
+  });
+}
+
+test('operations renderer clears corrected endpoint and interval errors', () => {
+  const wizard = loadWizard();
+  const value = state(wizard, { monitoringEnabled: true, monitoringEndpoint: 'env:', monitoringCheckIntervalHours: 0, attemptedSteps: ['operations'] });
+  const pane = element('main');
+  wizard.STEP_RENDERERS.operations(pane, value, () => {}, () => {});
+  const nodes = descendants(pane);
+  assert.equal(nodes.filter(node => node.attrs.role === 'alert').length, 2);
+  assert.deepEqual(nodes.filter(node => node.attrs.role === 'alert').map(node => node.textContent),
+    ['wizard.error.env_ref.invalid', 'wizard.error.monitoring.interval.invalid']);
+  const endpoint = nodes.find(node => node.tag === 'input' && node.attrs.type === 'text');
+  endpoint.value = 'ENV: MON_URL';
+  endpoint.events.input();
+  assert.doesNotMatch(endpoint.className, /is-invalid/);
+  assert.equal(descendants(pane).filter(node => node.attrs.role === 'alert').length, 1);
+  const interval = nodes.find(node => node.attrs.id === 'wizard-monitoring-interval');
+  interval.value = '48';
+  interval.events.input();
+  assert.doesNotMatch(interval.className, /is-invalid/);
+  assert.equal(descendants(pane).filter(node => node.attrs.role === 'alert').length, 0);
+  assert.equal(value.monitoringCheckIntervalHours, 48);
+  assert.match(wizard.buildYaml(value), /check_interval_hours: 48/);
+  const toggle = nodes.find(node => node.tag === 'label' && node.children.some(child =>
+    child.textContent === 'wizard.monitoring.enabled.label')).children[0];
+  toggle.checked = false;
+  toggle.events.change();
+  assert.doesNotMatch(wizard.buildYaml(value), /\nmonitoring:/);
+});
+
+test('active locales cover all keys; deferred locales use the real English fallback for new keys', () => {
+  const { translations } = loadWizard();
+  const keys = Object.keys(translations.en).filter(key => key.startsWith('wizard.monitoring.') ||
+    key.startsWith('wizard.error.monitoring.') || key === 'wizard.error.env_ref.invalid').sort();
+  const deferred = ['wizard.error.env_ref.invalid', 'wizard.monitoring.interval.label', 'wizard.error.monitoring.interval.invalid', 'wizard.error.monitoring.endpoint.invalid'];
+  const document = { documentElement: { lang: 'en' } };
+  const window = { ForgeLMTranslations: translations };
+  const context = vm.createContext({ window, document });
+  vm.runInContext(fs.readFileSync(path.join(root, 'site/js/_shared.js'), 'utf8'), context);
+  for (const locale of ['en', 'tr', 'de', 'fr', 'es', 'zh']) {
+    const expectedKeys = ['en', 'tr'].includes(locale) ? keys : keys.filter(key => !deferred.includes(key));
+    assert.deepEqual(Object.keys(translations[locale]).filter(key => keys.includes(key)).sort(), expectedKeys);
+    assert.ok(translations[locale]['wizard.monitoring.enabled.detail'].includes('declarative_only: true'));
+    for (const key of expectedKeys) assert.ok(translations[locale][key].trim(), `${locale}: ${key}`);
+    if (!['en', 'tr'].includes(locale)) {
+      document.documentElement.lang = locale;
+      for (const key of deferred) assert.equal(window.ForgeLMShared.tr(key), translations.en[key], `${locale}: ${key}`);
+    }
+  }
+});
+
+test('generated YAML preserves monitoring and webhook values through the real ForgeConfig schema', () => {
+  const wizard = loadWizard();
+  const configs = [];
+  for (const tier of ['minimal-risk', 'limited-risk', 'high-risk', 'unacceptable']) {
+    for (const [endpoint, literal, envName] of [
+      ['', '', null], ['ENV: _mon_url ', '', '_mon_url'],
+      ['https://monitor.example.com:8443/plan', 'https://monitor.example.com:8443/plan', null],
+      ['http://[::1]:9090/m', 'http://[::1]:9090/m', null]
+    ]) {
+      for (const exporter of ['none', 'prometheus', 'datadog', 'custom_webhook']) {
+        for (const interval of [1, 720]) {
+          for (const drift of [false, true]) {
+            const webhookLiteral = 'https://notify.example.com/service';
+            const value = state(wizard, { riskClassification: tier, monitoringEnabled: true, monitoringEndpoint: endpoint,
+              metricsExport: exporter, monitoringCheckIntervalHours: interval, alertOnDrift: drift,
+              webhookEnabled: true, webhookUrl: envName ? endpoint : webhookLiteral });
+            wizard.coerceForRisk(value);
+            assert.equal(wizard.validateStep(value, 'operations').length, 0, JSON.stringify(value));
+            configs.push({ yaml: wizard.buildYaml(value), expected: {
+              endpoint: literal, endpoint_env: envName, metrics_export: exporter, check_interval_hours: interval,
+              alert_on_drift: drift, webhook_url: envName ? null : webhookLiteral, webhook_env: envName,
+              endpoint_field: envName ? 'endpoint_env' : literal ? 'endpoint' : null
+            }});
+          }
+        }
+      }
+    }
+  }
+  const result = spawnSync(python, ['-c', [
+    'import json, sys, yaml',
+    'from forgelm.config import ForgeConfig',
+    'for case in json.load(sys.stdin):',
+    '    config = ForgeConfig(**yaml.safe_load(case["yaml"]))',
+    '    expected = case["expected"]',
+    '    assert config.monitoring.enabled is True',
+    '    for field in ("endpoint", "endpoint_env", "metrics_export", "alert_on_drift", "check_interval_hours"):',
+    '        assert getattr(config.monitoring, field) == expected[field], (field, expected, case["yaml"])',
+    '    assert {"metrics_export", "alert_on_drift", "check_interval_hours"} <= config.monitoring.model_fields_set',
+    '    actual_endpoint_fields = config.monitoring.model_fields_set & {"endpoint", "endpoint_env"}',
+    '    assert actual_endpoint_fields == ({expected["endpoint_field"]} if expected["endpoint_field"] else set())',
+    '    assert config.webhook.url == expected["webhook_url"], expected',
+    '    assert config.webhook.url_env == expected["webhook_env"], expected'
+  ].join('\n')], { cwd: root, input: JSON.stringify(configs), encoding: 'utf8', timeout: 30000 });
+  assert.ifError(result.error);
+  assert.equal(result.status, 0, `Python interpreter: ${python}\n${result.stderr}`);
+});
diff --git a/tests/test_config.py b/tests/test_config.py
index ae2a0366..10b81fd1 100644
--- a/tests/test_config.py
+++ b/tests/test_config.py
@@ -3,6 +3,8 @@
 import logging
 import os
 import re
+import subprocess
+import sys
 import warnings
 from pathlib import Path
 
@@ -27,11 +29,13 @@
     MergeConfig,
     ModelConfig,
     MonitoringConfig,
+    PipelineStage,
     SafetyConfig,
     SyntheticConfig,
     TrainingConfig,
     WebhookConfig,
     load_config,
+    merge_pipeline_stage_config,
 )
 from tests._helpers.factories import minimal_config
 
@@ -79,6 +83,151 @@ def test_unsloth_backend(self):
         assert m.backend == "unsloth"
 
 
+class TestMonitoringConfig:
+    @pytest.mark.parametrize("exporter", ["prometheus", "datadog", "custom_webhook"])
+    def test_exporter_declaration_warns_at_config_load(self, tmp_path, caplog, exporter):
+        config_data = minimal_config(
+            monitoring={
+                "enabled": True,
+                "metrics_export": exporter,
+                "endpoint": "https://secret.example.com/private",
+                "endpoint_env": "PRIVATE_MONITORING_NAME",
+            }
+        )
+        path = _write_yaml(config_data, str(tmp_path / "monitoring.yaml"))
+        with caplog.at_level(logging.WARNING, logger="forgelm.config"):
+            config = load_config(path)
+        assert config.monitoring.metrics_export == exporter
+        assert "monitoring.metrics_export" in caplog.text
+        assert "declarative metadata only" in caplog.text
+        assert "does not export metrics" in caplog.text
+        assert "secret.example.com" not in caplog.text
+        assert "PRIVATE_MONITORING_NAME" not in caplog.text
+        assert sum("monitoring.metrics_export" in record.message for record in caplog.records) == 1
+        for name in ("first", "second", "third"):
+            merge_pipeline_stage_config(config, PipelineStage(name=name))
+        assert sum("monitoring.metrics_export" in record.message for record in caplog.records) == 1
+
+    @pytest.mark.parametrize("value", ["MON_URL", " _mon_url ", "A1"])
+    def test_env_reference_is_normalized_without_resolution(self, value, monkeypatch):
+        monkeypatch.setenv(value.strip(), "https://secret.example.com/token")
+        assert MonitoringConfig(endpoint_env=value).endpoint_env == value.strip()
+
+    @pytest.mark.parametrize("value", [None, "", "   ", "\t\n"])
+    def test_blank_env_reference_is_absent(self, value):
+        assert MonitoringConfig(endpoint_env=value).endpoint_env is None
+
+    @pytest.mark.parametrize("value", ["env:MON_URL", "$MON_URL", "https://host/token", "MON URL", "1MON", "MÖN"])
+    def test_invalid_env_reference_is_rejected(self, value):
+        with pytest.raises(ValidationError, match="ASCII environment-variable name"):
+            MonitoringConfig(endpoint_env=value)
+
+    @pytest.mark.parametrize("root", [False, True], ids=["monitoring-model", "root-model"])
+    def test_invalid_env_reference_diagnostics_omit_raw_input(self, root):
+        monitoring = {
+            "enabled": True,
+            "endpoint_env": "https://user:fake-password@private.example/path?token=fake-token",
+        }
+        with pytest.raises(ValidationError) as exc:
+            if root:
+                ForgeConfig(**minimal_config(monitoring=monitoring))
+            else:
+                MonitoringConfig(**monitoring)
+        message = str(exc.value)
+        assert "endpoint_env" in message
+        assert "ASCII environment-variable name" in message
+        for secret in ("private.example", "fake-password", "fake-token"):
+            assert secret not in message
+
+    def test_config_load_error_omits_invalid_env_reference(self, tmp_path):
+        config_data = minimal_config(monitoring={"enabled": True, "endpoint_env": "https://private.example/token"})
+        path = _write_yaml(config_data, str(tmp_path / "monitoring.yaml"))
+        with pytest.raises(ConfigError) as exc:
+            load_config(path)
+        assert "monitoring.endpoint_env" in str(exc.value)
+        assert "ASCII environment-variable name" in str(exc.value)
+        assert "private.example" not in str(exc.value)
+
+    @pytest.mark.parametrize("value,expected", [("   ", ""), (" https://host/path\n", "https://host/path")])
+    def test_literal_endpoint_is_trimmed(self, value, expected):
+        assert MonitoringConfig(endpoint=value).endpoint == expected
+
+    @pytest.mark.parametrize(
+        "value",
+        [
+            "prometheus.internal:9090",
+            "localhost:8080",
+            "monitor.example.com/x",
+            "https:host",
+            "ftp://host",
+            "https://",
+            "https://host:bad",
+            "https://host:65536",
+            "http://[::1",
+            "https://host/a b",
+            "https://host/a\tb",
+            "https://host\\path",
+        ],
+    )
+    def test_literal_endpoint_requires_absolute_http_url(self, value):
+        with pytest.raises(ValidationError, match=r"http\(s\)://host") as exc:
+            MonitoringConfig(endpoint=value)
+        assert "input_value" not in str(exc.value)
+
+    @pytest.mark.parametrize(
+        "value", ["http://localhost:8080/m", "http://[::1]:9090/m", "https://user:fake@host:8443/p?q=fake#fragment"]
+    )
+    def test_literal_endpoint_keeps_valid_input_for_evidence_projection(self, value):
+        assert MonitoringConfig(endpoint=value).endpoint == value
+
+    @pytest.mark.parametrize(
+        "monitoring",
+        [None, {"enabled": False, "metrics_export": "prometheus"}, {"enabled": True, "metrics_export": "none"}],
+    )
+    def test_no_exporter_warning_for_inactive_export(self, tmp_path, caplog, monitoring):
+        path = _write_yaml(minimal_config(monitoring=monitoring), str(tmp_path / "monitoring.yaml"))
+        with caplog.at_level(logging.WARNING, logger="forgelm.config"):
+            load_config(path)
+        assert "monitoring.metrics_export" not in caplog.text
+
+    def test_direct_construction_does_not_emit_file_load_warning(self, caplog):
+        with caplog.at_level(logging.WARNING, logger="forgelm.config"):
+            config = ForgeConfig(**minimal_config(monitoring={"enabled": True, "metrics_export": "prometheus"}))
+        assert config.monitoring.metrics_export == "prometheus"
+        assert "monitoring.metrics_export" not in caplog.text
+
+    def test_null_literal_endpoint_is_rejected_in_yaml(self, tmp_path):
+        path = _write_yaml(minimal_config(monitoring={"enabled": True, "endpoint": None}), str(tmp_path / "null.yaml"))
+        with pytest.raises(ConfigError, match="monitoring.endpoint") as exc:
+            load_config(path)
+        assert "type=string_type" in str(exc.value)
+        assert "input_value" not in str(exc.value)
+
+    def test_pipeline_cli_dry_run_warns_once_without_endpoint_name(self, tmp_path):
+        data = minimal_config(
+            model={"name_or_path": "org/model", "offline": True},
+            training={"output_dir": str(tmp_path / "output")},
+            monitoring={"enabled": True, "metrics_export": "prometheus", "endpoint_env": "PRIVATE_MONITORING_NAME"},
+            pipeline={
+                "stages": [
+                    {"name": name, "training": {"trainer_type": "sft", "output_dir": str(tmp_path / "output" / name)}}
+                    for name in ("first", "second", "third")
+                ]
+            },
+        )
+        path = _write_yaml(data, str(tmp_path / "pipeline.yaml"))
+        result = subprocess.run(
+            [sys.executable, "-m", "forgelm", "--config", path, "--dry-run"],
+            cwd=Path(__file__).resolve().parents[1],
+            capture_output=True,
+            text=True,
+            timeout=30,
+        )
+        assert result.returncode == 0, result.stderr
+        assert result.stderr.count("monitoring.metrics_export is declarative metadata only") == 1
+        assert "PRIVATE_MONITORING_NAME" not in result.stderr
+
+
 # --- LoraConfigModel ---
 
 
diff --git a/tests/test_eu_ai_act.py b/tests/test_eu_ai_act.py
index 58ef5e30..ac3e9c13 100644
--- a/tests/test_eu_ai_act.py
+++ b/tests/test_eu_ai_act.py
@@ -11,6 +11,7 @@
 
 from forgelm.compliance import (
     AuditLogger,
+    compute_annex_iv_manifest_hash,
     export_compliance_artifacts,
     export_evidence_bundle,
     generate_deployer_instructions,
@@ -24,6 +25,7 @@
     RiskAssessmentConfig,
     load_config,
 )
+from forgelm.verify import verify_annex_iv_artifact
 
 # --- Config Models ---
 
@@ -815,6 +817,124 @@ def test_successful_export_promotes_all_artifacts(self, tmp_path):
 
 
 class TestManifestAnnexIV:
+    @pytest.mark.parametrize("env_value", [None, "https://secret.example.com/private"], ids=["env-unset", "env-set"])
+    @pytest.mark.parametrize(
+        "endpoint_fields,expected,expected_defaults",
+        [
+            ({}, None, ["endpoint"]),
+            ({"endpoint": "https://monitor.example.com"}, "https://monitor.example.com", []),
+            (
+                {"endpoint": "https://user:password@monitor.example.com:8443/secret?token=credential#fragment"},
+                "https://monitor.example.com:8443",
+                [],
+            ),
+            ({"endpoint": "http://[::1]:9090/secret?token=credential#fragment"}, "http://[::1]:9090", []),
+            (
+                {"endpoint": " https://user:password@monitor.example.com/secret?token=credential#fragment "},
+                "https://monitor.example.com",
+                [],
+            ),
+            ({"endpoint_env": "MON_URL"}, "$MON_URL", []),
+            ({"endpoint": "https://monitor.example.com", "endpoint_env": "MON_URL"}, "$MON_URL", []),
+            ({"endpoint_env": ""}, None, ["endpoint"]),
+            ({"endpoint_env": " \t\n "}, None, ["endpoint"]),
+            ({"endpoint_env": "   ", "endpoint": "   "}, None, []),
+            ({"endpoint_env": " MON_URL "}, "$MON_URL", []),
+        ],
+        ids=[
+            "absent",
+            "literal",
+            "literal-port",
+            "literal-ipv6",
+            "masked-literal",
+            "env",
+            "env-precedence",
+            "blank-env",
+            "whitespace-env",
+            "explicit-blank-endpoint",
+            "trimmed-env",
+        ],
+    )
+    def test_monitoring_declaration_survives_export(
+        self, tmp_path, minimal_config, monkeypatch, endpoint_fields, expected, expected_defaults, env_value
+    ):
+        """A plan is labelled as metadata and never resolves endpoint secrets."""
+        if env_value is None:
+            monkeypatch.delenv("MON_URL", raising=False)
+        else:
+            monkeypatch.setenv("MON_URL", env_value)
+        config = ForgeConfig(
+            **minimal_config(
+                compliance={"provider_name": "Corp", "system_name": "Bot", "intended_purpose": "Testing"},
+                monitoring={
+                    "enabled": True,
+                    "metrics_export": "prometheus",
+                    "alert_on_drift": True,
+                    "check_interval_hours": 12,
+                    **endpoint_fields,
+                },
+            )
+        )
+        manifest = generate_training_manifest(config, {"eval_loss": 0.5})
+        declaration = {
+            "declarative_only": True,
+            "endpoint": expected,
+            "endpoint_redacted": bool(config.monitoring.endpoint and not config.monitoring.endpoint_env),
+            "metrics_export": "prometheus",
+            "alert_on_drift": True,
+            "check_interval_hours": 12,
+            "defaulted_fields": expected_defaults,
+        }
+        assert manifest["monitoring"] == declaration
+        export_compliance_artifacts(manifest, str(tmp_path))
+        report = json.loads((tmp_path / "compliance_report.json").read_text(encoding="utf-8"))
+        annex = json.loads((tmp_path / "annex_iv_metadata.json").read_text(encoding="utf-8"))
+        assert report["monitoring"] == declaration
+        assert annex["monitoring_and_logging"] == declaration
+        training_summary = yaml.safe_load((tmp_path / "training_manifest.yaml").read_text(encoding="utf-8"))
+        assert "monitoring" not in training_summary
+        assert "secret.example.com" not in json.dumps(report)
+        assert "secret.example.com" not in json.dumps(annex)
+        for secret in ("password", "credential", "/secret", "#fragment"):
+            assert secret not in json.dumps(report)
+            assert secret not in json.dumps(annex)
+            assert secret not in (tmp_path / "training_manifest.yaml").read_text(encoding="utf-8")
+        assert annex["metadata"]["manifest_hash"] == compute_annex_iv_manifest_hash(annex)
+        assert verify_annex_iv_artifact(str(tmp_path / "annex_iv_metadata.json")).valid
+
+    def test_monitoring_defaults_are_not_operator_declarations(self, minimal_config):
+        from forgelm.config import PipelineStage, merge_pipeline_stage_config
+
+        config = ForgeConfig(**minimal_config(monitoring={"enabled": True}))
+        for effective in (config, merge_pipeline_stage_config(config, PipelineStage(name="first"))):
+            declaration = generate_training_manifest(effective, {})["monitoring"]
+            assert declaration["defaulted_fields"] == [
+                "endpoint",
+                "metrics_export",
+                "alert_on_drift",
+                "check_interval_hours",
+            ]
+            assert declaration["alert_on_drift"] is True
+            assert declaration["check_interval_hours"] == 24
+
+        explicit = ForgeConfig(
+            **minimal_config(
+                monitoring={
+                    "enabled": True,
+                    "endpoint": "",
+                    "metrics_export": "none",
+                    "alert_on_drift": True,
+                    "check_interval_hours": 24,
+                }
+            )
+        )
+        assert generate_training_manifest(explicit, {})["monitoring"]["defaulted_fields"] == []
+
+    @pytest.mark.parametrize("monitoring", [None, {"enabled": False, "metrics_export": "datadog"}])
+    def test_disabled_monitoring_is_not_declared(self, minimal_config, monitoring):
+        config = ForgeConfig(**minimal_config(monitoring=monitoring))
+        assert "monitoring" not in generate_training_manifest(config, {})
+
     def test_includes_annex_iv(self, minimal_config):
         config = ForgeConfig(
             **minimal_config(
diff --git a/tests/test_guard_wiring.py b/tests/test_guard_wiring.py
index 90392f3a..b96fa088 100644
--- a/tests/test_guard_wiring.py
+++ b/tests/test_guard_wiring.py
@@ -354,6 +354,7 @@ def test_gauntlet_non_guard_steps_are_wired_into_ci():
         "ruff check .",
         "ruff format",
         "pytest",
+        "node --test tests/site/*.test.cjs",
         "--config config_template.yaml --dry-run",
         # The FULL mypy argv, including the probe. A fragment stopping one
         # argument short was the bug: deleting ` tests/typing/public_surface_probe.py`
diff --git a/tests/test_wizard_phase22.py b/tests/test_wizard_phase22.py
index a1188788..0d76737b 100644
--- a/tests/test_wizard_phase22.py
+++ b/tests/test_wizard_phase22.py
@@ -78,11 +78,8 @@ def test_env_prefix_empty_var_name_rejected(self):
         with pytest.raises(ValueError, match="non-empty variable name"):
             wizard._parse_webhook_value("env:")
 
-    def test_env_prefix_lowercase_var_name_rejected(self):
-        # POSIX env-var convention enforces uppercase letters; lower-
-        # case names break ``${VAR}`` shell expansion in CI.
-        with pytest.raises(ValueError, match="POSIX environment-variable name"):
-            wizard._parse_webhook_value("env:lowercase_var")
+    def test_env_prefix_lowercase_var_name_accepted(self):
+        assert wizard._parse_webhook_value("env:lowercase_var") == {"url_env": "lowercase_var"}
 
     def test_bare_string_no_scheme_rejected(self):
         with pytest.raises(ValueError, match="not a valid URL"):
@@ -1011,20 +1008,23 @@ def test_empty_var_name_raises(self):
         with pytest.raises(ValueError, match="non-empty variable name"):
             _parse_env_ref("env:")
 
-    def test_lowercase_var_name_raises(self):
+    @pytest.mark.parametrize(
+        "reference,expected", [("env:lowercase_var", "lowercase_var"), ("ENV: _MON_URL ", "_MON_URL")]
+    )
+    def test_ascii_identifiers_are_accepted(self, reference, expected):
         from forgelm.wizard._collectors import _parse_env_ref
 
-        with pytest.raises(ValueError, match="POSIX environment-variable name"):
-            _parse_env_ref("env:lowercase_var")
+        assert _parse_env_ref(reference) == expected
 
-    def test_monitoring_collector_reuses_shared_helper(self, capsys):
+    @pytest.mark.parametrize("raw", ["env:not valid", "env:1MON", "env:MÖN", "env:https://private.example/fake-token"])
+    def test_monitoring_collector_reuses_shared_helper(self, capsys, raw):
         # _collect_monitoring's soft-fail policy (warn + drop the field,
         # not raise) must still hold with the shared helper.
         with patch(
             "builtins.input",
             side_effect=_input_returning(
                 "y",  # configure post-market monitoring
-                "env:not valid",  # malformed env ref
+                raw,  # malformed env ref
                 "1",  # metrics_export = none
                 "n",  # alert_on_drift
                 "24",  # check_interval_hours
@@ -1035,6 +1035,9 @@ def test_monitoring_collector_reuses_shared_helper(self, capsys):
         assert "endpoint" not in monitoring
         out = capsys.readouterr().out
         assert "not a valid env-var reference" in out
+        assert raw not in out
+        assert "private.example" not in out
+        assert "fake-token" not in out
 
 
 class TestUniqueFilenamePrompt:
@@ -1190,7 +1193,37 @@ def test_minimal_config_prints_checklist(self, capsys, monkeypatch):
 class TestMonitoringEnvPrefix:
     """P9 — monitoring collector accepts env:VAR_NAME like the webhook collector."""
 
-    def test_env_prefix_routes_to_endpoint_env(self):
+    def test_metadata_plan_accepts_no_endpoint(self, capsys):
+        with patch("builtins.input", side_effect=_input_returning("y", "", "1", "y", "24")):
+            result = wizard._collect_monitoring()
+        assert result == {"enabled": True, "metrics_export": "none", "alert_on_drift": True, "check_interval_hours": 24}
+        output = capsys.readouterr().out
+        assert "Metadata only" in output
+        assert "does not export metrics" in output
+        assert "schedule checks" in output
+
+    def test_accepting_wizard_defaults_declares_no_drift_alerts(self):
+        with patch("builtins.input", side_effect=_input_returning("y", "", "", "", "")):
+            result = wizard._collect_monitoring()
+        assert result == {
+            "enabled": True,
+            "metrics_export": "none",
+            "alert_on_drift": False,
+            "check_interval_hours": 24,
+        }
+
+    def test_malformed_literal_url_is_reprompted_without_echo(self, capsys):
+        with patch(
+            "builtins.input",
+            side_effect=_input_returning("y", "private.example/fake-token", "http://[::1]:9090/m", "1", "n", "24"),
+        ):
+            result = wizard._collect_monitoring()
+        assert result["endpoint"] == "http://[::1]:9090/m"
+        output = capsys.readouterr().out
+        assert "Use an http(s)://host URL" in output
+        assert "private.example/fake-token" not in output
+
+    def test_env_prefix_routes_to_endpoint_env(self, capsys):
         with patch(
             "builtins.input",
             side_effect=_input_returning(
@@ -1204,6 +1237,7 @@ def test_env_prefix_routes_to_endpoint_env(self):
             result = wizard._collect_monitoring()
         assert result["endpoint_env"] == "DATADOG_URL"
         assert "endpoint" not in result
+        assert "Metadata only" in capsys.readouterr().out
 
     def test_literal_url_routes_to_endpoint(self):
         with patch(
diff --git a/tools/check_module_size.py b/tools/check_module_size.py
index 64027298..35bb0b0d 100644
--- a/tools/check_module_size.py
+++ b/tools/check_module_size.py
@@ -263,7 +263,7 @@ class _DeferredSplit:
 # _validate_entries() sees a raise without reading the diff.
 _DEFERRED_SPLITS: dict[str, _DeferredSplit] = {
     "forgelm/compliance.py": _DeferredSplit(
-        budget=2492,
+        budget=2514,
         deferred_at_loc=2147,
         reason=(
             "EU AI Act Art. 9-17 + Annex IV builder + hash-chained audit log + "
@@ -271,6 +271,12 @@ class _DeferredSplit:
             "_annex_iv, _provenance, _gdpr."
         ),
         budget_history=(
+            "2026-10-15 (#93 re-review, 2505 -> 2514, +9): Project validated monitoring URLs with stdlib parsing so evidence preserves ports and IPv6 brackets without importing the HTTP/requests stack; keep log masking unchanged.",
+            "2026-10-15 (#93 review, 2493 -> 2505, +12): Persist monitoring default provenance and masked endpoint metadata at the existing writer; an unrelated compliance split would obscure this evidence-contract correction.",
+            "2026-10-15 (#93, 2492 -> 2493, +1): label the persisted monitoring plan "
+            "declarative_only so neither the manifest nor its Annex IV projection implies "
+            "runtime monitoring. Keep the label with the existing writer; the one-line "
+            "evidence correction does not warrant an unrelated compliance-package split.",
             "2026-10-09 (Phase 16 S1-S3 review, 2473 -> 2492, +19): "
             "`match_annex_iv_manifest_hash`. Hashing the Annex IV / pipeline manifest over "
             "the strict-JSON form (so a fresh artefact holding a NaN verifies) meant an "
@@ -313,9 +319,16 @@ class _DeferredSplit:
         ),
     ),
     "forgelm/config.py": _DeferredSplit(
-        budget=1855,
+        budget=1903,
         deferred_at_loc=1795,
         budget_history=(
+            "2026-10-15 (#93 re-review, 1887 -> 1903, +16): Reject malformed literal monitoring URLs before evidence export, with static diagnostics and no network access; validation belongs beside the existing endpoint field.",
+            "2026-10-15 (#93 review, 1870 -> 1887, +17): Normalize and validate monitoring endpoints beside their schema and emit the exporter warning at load_config, avoiding pipeline revalidation duplicates without mutable state.",
+            "2026-10-15 (#93, 1855 -> 1870, +15): describe MonitoringConfig as external-plan "
+            "metadata and warn when an enabled plan declares an exporter ForgeLM does not run. "
+            "The validator belongs beside its schema; the remaining growth is required field "
+            "description wrapping. Keep this contract repair separate from the deferred "
+            "schema split, whose import-time validation order needs its own review.",
             "2026-10-09 (Phase 16 S1-S3 review, 1846 -> 1855, +9): the `FiniteFloat` alias "
             "(`Annotated[float, allow_inf_nan=False]`, 7 lines with its rationale) and its "
             "application to the twelve float fields S2 left open — `learning_rate`, "