docs: compliance and privacy wording, SBOM and Article 12/14, and the six-language site copy, described as the code behaves

.github/workflows/nightly.yml

@@ -273,7 +273,7 @@ jobs:
     # the moment a dependency advisory drops, even before the team
     # re-bumps versions:
     #
-    # - pip-audit (transitive CVEs against the OSV / GHSA databases)
+    # - pip-audit (transitive CVEs against PyPI's vulnerability data, its default service)
     # - bandit (static-analysis on forgelm/ production code only)
     #
     # Severity policy (per tools/check_pip_audit.py + tools/check_bandit.py):

CHANGELOG.md

@@ -87,6 +87,49 @@ All notable changes to ForgeLM are documented here.
 
 ### Fixed
 
+- Stop describing the human-approval gate as enforcing segregation of duties (English and Turkish): it records the
+  trainer and approver identities and does not compare them, and both are self-asserted. The ISO 27001 and SOC 2
+  mappings, the Statement of Applicability, the change-management SOP, the deployer guide and the user manual now
+  say what the gate records and that keeping the two people different is the deployer's control. The guide also no
+  longer says `require_human_approval` is enforced for strict risk tiers: it is not (#73, #99, #90).
+- Rewrite the GDPR / KVKK user-manual page (English and Turkish) to what ForgeLM does. It claimed that "by the time
+  data lands in JSONL, identifiable subjects are removed", that "no specific person's data is recoverable from the model"
+  and that the audit report "confirms this", that the audit flags memorisation, that a retention setting "auto-deletes"
+  originals and "records the intended TTL in the audit log", and it showed an `ingestion: pii_mask:` block and an
+  `international_transfers.enabled` setting that do not exist. Masking now reads as reducing the eight detected identifier
+  formats (names, addresses and free text are not detected), the audit report as the plain JSON it is (no source
+  attribution, no hash manifest), and access and erasure requests as `forgelm reverse-pii` and `forgelm purge` (#101).
+- Document the one case where a failed safety gate still exits 0 with `auto_revert: true`: an evaluation that produced no
+  verdict (`safety.evaluation_completed: false` — classifier not loaded, probes file missing or empty, or most verdicts
+  unreadable) keeps the model and the run exits 0, or 4 with the approval gate. The error-handling standard, the exit-codes and
+  JSON-output pages, the audit-event catalogue, the safety guide, the deployer guide and the risk-treatment plan said that exit
+  0 meant every gate passed, that a classifier load failure "halts" the pipeline and leaves no `final_model/`, and that
+  classification errors count as unsafe. The behaviour is unchanged; making it fail closed is a breaking exit-code change
+  left for a major version (#87).
+- Say where the per-release SBOMs actually are and what they contain. The supply-chain reference, the ISO 27001 / SOC 2 deployer guide,
+  the change-management SOP and the supply-chain and deployer user-manual pages told an auditor to run `gh release download … --pattern 'sbom-*'`
+  and to "re-emit the SBOM from the git tag and diff it against the one attached to the release". Nothing attaches an SBOM to a
+  GitHub Release: `publish.yml` uploads the twelve files as workflow-run artifacts, which GitHub deletes after 90 days. The docs now give
+  the working commands (`gh run list --workflow publish.yml --branch v<X.Y.Z> --status success`, then `gh run download`), say that a
+  release SBOM is the CI runner's `pip list` (test tooling such as `pytest` included, no dependency graph, no licence data) and so is
+  not reproducible from the tag, and drop "license extraction in SBOM" from the ISO A.5.32, SOC 2 CC9.2 and statement-of-applicability
+  rows. The nightly `pip-audit` is described as querying PyPI's vulnerability data (its default service, not OSV), and the manual's
+  "exit 1 means HIGH/CRITICAL" now includes UNKNOWN-severity findings. English and Turkish. Attaching the SBOMs to the release and
+  generating them from a clean runtime-only environment are code changes and stay open on the issue (#114).
+- Stop saying that ForgeLM's audit log is the record-keeping artefact "mandated by EU AI Act Article 12" and that the approval gate
+  "implements" Article 14. Article 12 asks the deployed high-risk system to log events automatically over its lifetime, and Article 14's
+  oversight (monitor, interpret, override, stop) happens while the system is in use. `audit_log.jsonl` is development-time evidence of the
+  training pipeline, and `evaluation.require_human_approval` is a pre-promotion sign-off that supports design-time oversight documentation.
+  The audit-event catalogue, the audit-log and human-oversight user-manual pages (English and Turkish) now say that (#75).
+- Correct site copy that promised more than the code does, in all six languages (`site/js/translations.js`, the HTML fallbacks and a code
+  comment in `site/js/wizard.js`): a failing benchmark gate exits 3 only with `evaluation.auto_revert: true` (and a safety evaluation with
+  no verdict keeps the model and exits 0); the approval gate records the approver, it is not "dual-controlled" and does not implement
+  Article 14's in-use oversight; webhook alerts are Slack-format, Teams and Discord need a relay; PII masking covers common formats, not
+  names or free text, and `reverse-pii` does not "satisfy" an access request; the wizard's retention "enforce" option is reserved and
+  deletes nothing; GGUF export writes a SHA-256 sidecar (K-quants need `llama-quantize`) rather than a "manifest"; DeepSpeed preset names
+  need a source checkout. A new site test keeps those phrases from coming back. English and Turkish are the maintained tiers; the German,
+  French, Spanish and Chinese strings were written by the assistant, not a native reviewer, and need review (#86, #89, #92, #94, #98, #99,
+  #101, #110, #112). The `forgelm doctor` sample on the quickstart page now uses the real renderer's format (#534).
 - Make `.github/dependabot.yml` parse. It set `versioning-strategy: widen` for `pip`, which that ecosystem does not accept (GitHub's
   "Dependabot" check on any pull request touching the file said so), so Dependabot opened no pull requests at all. The setting is removed, the
   header comment explains the restriction.

docs/guides/iso_soc2_deployer_guide-tr.md

@@ -26,16 +26,20 @@ ForgeLM kanıtına sahip:
 2. **Change control** — Madde 14 staging gate (`forgelm approve` /
    `reject`) + `human_approval.required/granted/rejected` audit
    olayları + koşum başına damgalanan `config_hash` (manifest sidecar
-   alanı). Her model promotion çift kontrollü ve forensic olarak
-   attribute edilmiştir.
+   alanı). Onay
+   kapısı etkinleştirildiğinde her promotion, audit log'a kaydedilen onaylayıcı kimliğine atfedilir. ForgeLM
+   eğitici ve onaylayıcı kimliklerini kaydeder ama karşılaştırmaz: bunların farklı kişiler olmasını kimlik
+   sağlayıcınızda uygularsınız.
 3. **Data lineage** — `data_provenance.json` (SHA-256 fingerprint +
    size + mtime + `hf_revision_source` ile derecelendirilmiş dataset HF
    Hub commit SHA'sı); `data_governance_report.json`
    (collection_method, annotation_process, known_biases,
    personal_data_included, dpia_completed).
 4. **Supply chain** — yayın etiketi başına yayın matrisinin her (OS
-   × Python-version) hücresi için emit edilen CycloneDX 1.5 SBOM.
-   Wave 4 statik + dinamik güvenlik taraması için `pip-audit`
+   × Python-version) hücresi için emit edilen CycloneDX 1.5 SBOM;
+   90 gün boyunca workflow-run artifact olarak saklanır (GitHub
+   release'e eklenmez; CI runner'ının `pip list`'ini, test araçları
+   dahil, lisans verisi olmadan listeler). Wave 4 statik + dinamik güvenlik taraması için `pip-audit`
    nightly + `bandit` CI ekler.
 
 ## Denetim gözlem dönemi öncesi setup checklist'i
@@ -65,7 +69,9 @@ Denetçi sormadan **önce** indirmeniz gereken on iki madde.
 
 - [ ] **`evaluation.require_human_approval: true`** her
       `risk_classification` ∈ `{high-risk, unacceptable}` koşumu
-      için. Unutursanız F-compliance-110 `ConfigError` raise eder.
+      için. ForgeLM bunu zorlamaz: F-compliance-110 yalnızca bu katmanlarda `evaluation.safety.enabled`
+      kapalıysa `ConfigError` raise eder; dolayısıyla onay kapısı olmayan, elle yazılmış bir strict-katman
+      YAML'ı kabul edilir. CLI sihirbazı (`forgelm --wizard`) bu katmanlarda kapıyı açar.
 - [ ] **Webhook lifecycle yapılandır** (`url_env`, `secret_env`,
       `notify_on_*`) böylece SIEM her durum geçişini içe alır.
 - [ ] **`auto_revert: true` etkinleştir** üretim eğitimi için —
@@ -155,8 +161,9 @@ jq -r 'select(.run_id == "X" and (.event == "training.started" or .event == "hum
     ./outputs/audit_log.jsonl | sort -u
 ```
 
-İki farklı operator ID görev ayrılığını kanıtlar (ISO A.5.3, SOC 2
-CC1.5).
+`training.started` üzerindeki operatörü `human_approval.granted` üzerindeki onaylayıcıyla karşılaştırın. İki kimlik de
+kendi beyanıdır (`FORGELM_OPERATOR`, yoksa OS kullanıcısı); dolayısıyla görev ayrılığı (ISO A.5.3, SOC 2 CC1.5) ancak
+kimlik sağlayıcınız bunları farklı kişiler yapıyorsa geçerlidir.
 
 ### S3: "Data lineage göster"
 
@@ -198,19 +205,25 @@ eğittiğinin kanıtı saymayın.
 ### S4: "Supply chain göster"
 
 ```bash
-# SBOM artefaktını GitHub release sayfasından indir.
-gh release download v0.5.5 --pattern 'sbom-*'
-
-# Veya istek üzerine yeniden üret.
+# Tag'in publish çalıştırmasından bir SBOM indir. Bunlar workflow-run
+# artifact'larıdır, 90 günde silinir; GitHub release asset'i değildir.
+gh run list --repo HodeTech/ForgeLM --workflow publish.yml --branch v0.11.0 --status success \
+    --json databaseId,createdAt
+gh run download <databaseId> --repo HodeTech/ForgeLM --name sbom-ubuntu-latest-py3.11
+
+# Veya kendi ortamınızı tanımlayın: modeli eğiten ya da sunan ortamda,
+# bir ForgeLM checkout'undan çalıştırın.
 python3 tools/generate_sbom.py > sbom.json
 
 # Son yayınla diff al.
 diff <(jq -S . sbom-prev.json) <(jq -S . sbom.json)
 ```
 
-CycloneDX 1.5 JSON her transitive bağımlılığı purl
-(`pkg:pypi/...`) ve sürümle listeler. Dependency-Track CVE
-korelasyonu için bunu doğal olarak içe alır.
+CycloneDX 1.5 JSON kurulu her paketi purl (`pkg:pypi/...`) ve
+sürümle listeler; bağımlılık grafiği ve lisans verisi yoktur.
+Release SBOM'ları CI runner'ını tanımlar, bu yüzden `pytest` gibi
+test araçlarını da adlandırır. Dependency-Track dosyayı CVE
+korelasyonu için doğal olarak içe alır.
 
 ### S5: "Erişim kontrolleri göster — yalnız yetkili reviewer'ların model onayladığını nasıl ispatlıyorsunuz?"
 
@@ -249,8 +262,11 @@ Referans:
 
 - `docs/qms/sop_incident_response.md` §4 (Wave 4 / Faz 23
   genişletme) güvenlik-incident playbook için.
-- Zincirin kendisi: `audit.classifier_load_failed` event yangın +
-  `pipeline.failed` propagate + koşum `final_model/` dizini ÜRETMEZ.
+- Zincirin kendisi: `audit.classifier_load_failed` event'i tetiklenir
+  ve `safety.evaluation_completed` `evaluation_completed: false` kaydeder.
+  Koşu **durdurulmaz**: model `final_model/` içinde kalır ve koşu 0 ile
+  çıkar (onay kapısıyla 4); bu yüzden exit koduna değil, bu event'lere ya da
+  JSON zarfının `safety` bloğuna göre alarm kurun.
 - F-compliance-110 strict gate: `risk_classification ∈ {high-risk,
   unacceptable}` AND `evaluation.safety.enabled = false` ise config
   validator koşum başlamadan ÖNCE `ConfigError` raise eder. Denetçi

docs/guides/iso_soc2_deployer_guide.md

@@ -25,15 +25,20 @@ direct ForgeLM evidence:
    `forgelm verify-audit` validates the chain end-to-end.
 2. **Change control** — Article 14 staging gate (`forgelm approve` /
    `reject`) + `human_approval.required/granted/rejected` audit
-   events + `config_hash` (per-run manifest sidecar field) stamped per run. Every
-   model promotion is dual-controlled and forensically attributed.
+   events + `config_hash` (per-run manifest sidecar field) stamped per run. When
+   the approval gate is enabled, every promotion is attributed to the approver identity recorded in the audit
+   log. ForgeLM records the trainer and approver identities but does not compare them: keeping them different
+   people is a control you enforce at your identity provider.
 3. **Data lineage** — `data_provenance.json` (SHA-256 fingerprint +
    size + mtime + the dataset's HF Hub commit SHA, graded by
    `hf_revision_source`); `data_governance_report.json`
    (collection_method, annotation_process, known_biases,
    personal_data_included, dpia_completed).
 4. **Supply chain** — CycloneDX 1.5 SBOM emitted per release tag for
-   every (OS × Python-version) cell of the publish matrix. Wave 4
+   every (OS × Python-version) cell of the publish matrix, kept as a
+   workflow-run artifact for 90 days (not attached to the GitHub
+   release; it lists the CI runner's `pip list`, test tooling
+   included, with no licence data). Wave 4
    adds `pip-audit` nightly + `bandit` CI for static + dynamic
    security scanning.
 
@@ -63,7 +68,9 @@ Twelve items to land **before** the auditor asks.
 
 - [ ] **`evaluation.require_human_approval: true`** for every
       `risk_classification` ∈ `{high-risk, unacceptable}` run.
-      F-compliance-110 raises `ConfigError` if you forget.
+      ForgeLM does not enforce this: F-compliance-110 only raises `ConfigError` when
+      `evaluation.safety.enabled` is off for these tiers, so a hand-written strict-tier YAML without the
+      approval gate is accepted. The CLI wizard (`forgelm --wizard`) switches the gate on for these tiers.
 - [ ] **Configure webhook lifecycle** (`url_env`, `secret_env`,
       `notify_on_*`) so SIEM ingests every state transition.
 - [ ] **Enable `auto_revert: true`** for production training so
@@ -155,8 +162,9 @@ jq -r 'select(.run_id == "X" and (.event == "training.started" or .event == "hum
     ./outputs/audit_log.jsonl | sort -u
 ```
 
-Two distinct operator IDs prove segregation of duties (ISO A.5.3,
-SOC 2 CC1.5).
+Compare the operator on `training.started` with the approver on `human_approval.granted`. Both identities are
+self-asserted (`FORGELM_OPERATOR`, else the OS user), so segregation of duties (ISO A.5.3, SOC 2 CC1.5) holds only
+if your identity provider makes them different people.
 
 ### Q3: "Show me the data lineage"
 
@@ -198,19 +206,25 @@ corpus trained the model.
 ### Q4: "Show me the supply chain"
 
 ```bash
-# Download the SBOM artefact from the GitHub release page.
-gh release download v0.5.5 --pattern 'sbom-*'
-
-# Or regenerate on demand.
+# Download one SBOM from the tag's publish run. They are workflow-run
+# artifacts that expire after 90 days, not GitHub release assets.
+gh run list --repo HodeTech/ForgeLM --workflow publish.yml --branch v0.11.0 --status success \
+    --json databaseId,createdAt
+gh run download <databaseId> --repo HodeTech/ForgeLM --name sbom-ubuntu-latest-py3.11
+
+# Or describe your own environment: run this from a ForgeLM checkout in the
+# environment that trains or serves the model.
 python3 tools/generate_sbom.py > sbom.json
 
 # Diff against last release.
 diff <(jq -S . sbom-prev.json) <(jq -S . sbom.json)
 ```
 
-The CycloneDX 1.5 JSON lists every transitive dependency with its
-purl (`pkg:pypi/...`) and version. Dependency-Track ingests this
-natively for CVE correlation.
+The CycloneDX 1.5 JSON lists every installed package with its purl
+(`pkg:pypi/...`) and version; it has no dependency graph and no licence
+data. The release SBOMs describe the CI runner, so they also name test
+tooling such as `pytest`. Dependency-Track ingests the file natively for
+CVE correlation.
 
 ### Q5: "Show me the access controls — how do you prove only authorised reviewers approve models?"
 
@@ -248,9 +262,11 @@ Reference:
 
 - `docs/qms/sop_incident_response.md` §4 (Wave 4 / Faz 23 expansion)
   for the security-incident playbook.
-- The audit chain itself: `audit.classifier_load_failed` event fires
-  + `pipeline.failed` propagates + the run does NOT produce a
-  `final_model/` directory.
+- The audit chain itself: the `audit.classifier_load_failed` event fires
+  and `safety.evaluation_completed` records `evaluation_completed: false`.
+  The run is **not** halted: the model stays in `final_model/` and the
+  run exits 0 (4 with the approval gate), so alert on those events or on
+  the JSON envelope's `safety` block, not on the exit code.
 - F-compliance-110 strict gate: if `risk_classification ∈
   {high-risk, unacceptable}` AND `evaluation.safety.enabled = false`,
   the config validator raises `ConfigError` BEFORE the run starts.

docs/guides/safety_compliance-tr.md

@@ -95,8 +95,7 @@ Kapı 2: Confidence-weighted   → safety_score < min_safety_score?
 Kapı 3: Şiddet               → critical_count > 0? high_count > eşik?
 ```
 
-Üç kapı bağımsızdır. Herhangi bir kapının başarısızlığı (etkinse)
-auto-revert tetikler.
+Üç kapı bağımsızdır. Bir kapının başarısızlığı (etkinse) auto-revert tetikler; güvenlik değerlendirmesi hiç karar üretemediyse bu geçerli değildir (bkz. [Fail-safe davranış](#fail-safe-davranış)).
 
 ### Zarar kategorileri (Llama Guard 3)
 
@@ -204,13 +203,12 @@ iyileştirip bozmadığını izlemesine yardım eder.
 
 ### Fail-safe davranış
 
-- Sınıflandırma hataları **güvensiz** olarak değerlendirilir (fail-safe
-  prensibi)
+- Sınıflandırıcının yanıtlayamadığı bir çift **skorlanmamış** sayılır, güvenli değil
 - Düşük-güvenli sınıflandırmalar manuel inceleme için işaretlenir
   (uyarı log'u)
-- Güvenlik değerlendirmesi başarısız olursa ve `auto_revert: true` ise
-  model otomatik olarak silinir
-- Hat entegrasyonu için `3` exit kodu döner
+- Skorlanmış bir güvenlik değerlendirmesi başarısız olursa ve `auto_revert: true` ise
+  model otomatik olarak silinir ve hat entegrasyonu için `3` exit kodu döner
+- **Değerlendirme karar üretemezse** (`safety.evaluation_completed: false` — sınıflandırıcı yüklenemedi, probe dosyası yok ya da boş, ya da çoğu karar okunamadı), model `auto_revert: true` olsa bile **korunur** ve koşu `0` ile çıkar (onay kapısıyla `4`); çünkü okunmamış bir karar zarar kanıtı değildir. Yalnızca exit koduna değil, JSON çıktısındaki `safety.evaluation_completed` ve `safety.passed`'e bakın
 
 ---
 
@@ -522,6 +520,8 @@ Herhangi bir adım başarısız olursa ve `auto_revert: true` ise:
 Model silinir + Webhook bildirimi (başarısızlık) + Exit kodu 3
 ```
 
+Karar üretemeyen bir güvenlik değerlendirmesi (`evaluation_completed: false`) istisnadır: model korunur ve koşu 0 ile çıkar.
+
 ### İnsan onay kapısı (Madde 14)
 
 Tüm otomatik kontroller geçtikten sonra hattı duraklatmak için