Housekeeping: skill guide published, jev plan on main, openspec tree un-ignored and archived, one-clock fix

.gitignore

@@ -534,6 +534,24 @@ docs/architecture/plan-integration/*
 !docs/architecture/plan-integration/proposals/
 !docs/architecture/plan-integration/*.architecture.json
 !docs/architecture/plan-integration/*.md
+# Un-ignored 2026-09-19 — jev-judge programme record (Jev / TypeSafe System One as an
+# optional judgement provider): stage receipts, council briefs and seat reviews.
+# Same shape as plan-integration above: small Markdown/JSON only.
+!docs/architecture/jev-judge/
+docs/architecture/jev-judge/*
+!docs/architecture/jev-judge/council/
+!docs/architecture/jev-judge/receipts/
+!docs/architecture/jev-judge/*.architecture.json
+!docs/architecture/jev-judge/*.md
+# Un-ignored 2026-09-19 — learning-tier programme record (feed the mentor's writers on all
+# six harnesses, then measure one judgement model on the struggle ruler): council brief,
+# seat reviews, arbitration, pre-registration and stage receipts. Small Markdown/JSON only.
+!docs/architecture/learning-tier/
+docs/architecture/learning-tier/*
+!docs/architecture/learning-tier/council/
+!docs/architecture/learning-tier/receipts/
+!docs/architecture/learning-tier/*.architecture.json
+!docs/architecture/learning-tier/*.md
 
 # Demo recordings (large, local-only)
 demos/
@@ -621,9 +639,6 @@ MagicMock/
 .pi/
 agent/
 skills-lock.json
-#
-# openspec files
-openspec/
 # dotfiles/dirs
 .graphify-labels.json
 .graphifyignore

.pre-commit-config.yaml

@@ -41,11 +41,15 @@ repos:
     # sha256 and the UAT redacted summary carries the rubric hash and the
     # private bundle's manifest digest (council D-22) -- every one a digest the
     # redaction rules allow out precisely because it identifies without revealing.
+    # Learning-tier council manifests and receipts (2026-09-19) are the same two
+    # classes: brief/system-prompt digests and pre-registration fingerprints.
     exclude: |
       (?x)^(
         packages/studyloop/tests/acceptance/uat/data/.*_registry\.json|
         docs/architecture/plan-integration/council/.*/manifest.*\.json|
-        docs/architecture/plan-integration/receipts/.*\.json
+        docs/architecture/plan-integration/receipts/.*\.json|
+        docs/architecture/learning-tier/council/.*/manifest.*\.json|
+        docs/architecture/learning-tier/receipts/.*\.json
       )$
 - repo: https://github.com/PyCQA/bandit
   rev: 1.8.3

CHANGELOG.md

@@ -29,6 +29,39 @@ experience may change before `1.0.0`.
   a hand-off during a live session, or rejoining another session never shows
   it under different material. Proposal only — the companion says nothing
   about it. The no-plan payload is byte-identical.
+- A standalone `studyloop-study-notes` Agent Skill with per-lesson Markdown and
+  section-overview templates, source/enrichment attribution, and explicit
+  Obsidian-only, xTiles-only, and linked dual-destination workflows. Installed
+  separately through the skills CLI, not by `studyloop install agents`; its
+  guide is published on the docs site as *Lesson Study Notes*
+  (`docs/study-notes-skill.md`).
+
+### Changed
+
+- The `openspec/` tree is no longer listed in `.gitignore`. Seventy-eight
+  tracked, load-bearing files lived under an ignored path, so every new spec
+  or archive file was invisible to `git status` and skipped by `git add -A`.
+  The `herdr-ghostty-multiplexer-transport` change is archived as deferred
+  (the owner's 2026-09-05 decision, unchanged: tmux stays the production
+  default, herdr an experimental opt-in) with its five open tasks closed as
+  not built and its spec deltas deliberately not merged — they modified
+  requirements no main spec contains and describe a wterm selector and a ttyd
+  fallback the tree has since retired. The July e2e/MCP archive's four open
+  tasks are reconciled against the tree (all four were built the same day
+  their "not done" notes were written), so `openspec validate --archived
+  --all` is green for the first time since the release guard was added.
+
+### Fixed
+
+- The `now` engine counts every "last seen N day(s) ago" from the one clock
+  it reads per plan. The due-progress collector took its day count from a
+  second clock inside `history/progress.py`, so under a frozen test clock the
+  count — and the score built from it — drifted with the real date (the
+  medium-energy screen in receipt `now-rubric-2026-09-16` printed 8 days for
+  a struggle planted 3 days back). Production always read one wall clock, so
+  no learner saw the split; `spaced_repetition_due()` gains a keyword-only
+  `now` for callers that already hold an instant, defaulting to the wall
+  clock for everyone else.
 
 ## [0.5.0] - 2026-09-21
 

Justfile

@@ -218,9 +218,10 @@ release-consistency:
 # change with commits since the last tag must be archived or carry a
 # `deferred: <reason>`, and archive entries ADDED since the last tag must pass
 # `openspec validate` (soft-skipped when the CLI is absent, same convention as
-# spec-check; not `--archived --all`, because a July archive predating this
-# guard has unticked tasks nobody has evidence to reconcile, and re-failing
-# every future release on it would teach people to ignore the gate).
+# spec-check; not `--archived --all`, so that a historical archive nobody is
+# working on can never re-fail a future release and teach people to ignore
+# the gate — the July archive that motivated this was reconciled against the
+# tree on 2026-09-23 and `--archived --all` is green today).
 # Deliberately NOT part of preflight: open changes are legal during a cycle;
 # only shipping one is not. Both guards would have fired on the 0.2.0 cut
 # (2026-09-04 review, Q5).

docs/architecture/jev-judge/receipts/stage0-access-2026-09-19.md

@@ -0,0 +1,100 @@
+# Stage 0 — Jev access spike (2026-09-19)
+
+**Programme:** `feat/jev-judge` — evaluate TypeSafe's Jev (a "System One" judgement model:
+typed questions against a text state, returns typed answers with probabilities and
+confidence, no text generation) as an *optional judgement provider* for StudyLoop's
+unfilled judgement slots. Out of scope by design: the decision engine and completion
+review (`learning/decision.py`, `planning/views.py`) — those count and compare dates,
+which the vendor's own jaggedness page says Jev cannot do.
+
+**Question this stage answers:** does the pre-release account answer at all, does the SDK
+match its docs, and do the five 1–4 teach-back dimensions map onto Jev's Score legend?
+It is an access and shape check, **not** a measurement of accuracy (n = 1 synthetic state).
+
+## Method
+
+- Script: `scripts/eval/jev_stage0_spike.py` (disposable; reads `TYPESAFE_API_KEY` from
+  the environment, never prints it). SDK `typesafe-sdk==0.7.0`, Python 3.12.8.
+- Model pinned to `jev-1.13.0` — never the `jev-latest` alias, because any threshold
+  tuned against an alias silently moves when the alias does (vendor's advice, adopted).
+- State: one synthetic teach-back (1,873 chars) — a networking-background learner
+  explaining Python decorators with a middlebox/NAT analogy and a retry-decorator
+  transfer example. **One factual error planted on purpose**: "the wrapping happens every
+  time you call the decorated function, not when it's defined". A human marks that
+  Accuracy 2 ("mostly correct, minor gaps") at best.
+- Questions in one call: five `Score` questions whose four criteria are the verbatim level
+  descriptors from `agents/shared/teach-back-protocol.md` (Recitation → Teaching), plus
+  two `Noul` controls — positive (`has_factual_error`) and negative (`not_english`).
+- Three identical calls to measure spread.
+
+## Results
+
+| dimension | jev score ×3 (0–3) | StudyLoop 1–4 (mean) | confidence ×3 | range | human expectation |
+|---|---|---|---|---|---|
+| accuracy | 2.09, 2.10, 2.15 | **3.11** | 0.52, 0.52, 0.54 | 0.06 | **2** (planted error) |
+| own_words | 2.98, 2.98, 2.98 | 3.98 | 0.98, 0.98, 0.98 | 0.00 | 4 (novel analogies) |
+| structure | 2.90, 2.88, 2.87 | 3.88 | 0.90, 0.88, 0.87 | 0.03 | 3–4 |
+| depth | 2.03, 2.03, 2.03 | 3.03 | 0.96, 0.96, 0.97 | 0.00 | 3 (WHAT/HOW/WHY, no tradeoffs) |
+| transfer | 2.67, 2.65, 2.68 | 3.67 | 0.67, 0.65, 0.68 | 0.03 | 3–4 |
+| noul: has_factual_error | 0.48, 0.48, 0.43 | – | – | 0.05 | high |
+| noul: not_english | 0.01, 0.01, 0.01 | – | – | 0.00 | ≈ 0 |
+
+Per-level probabilities are returned for every Score (accuracy call 1:
+`{0: 0.02, 1: 0.16, 2: 0.54, 3: 0.28}`), so a caller can take the argmax level and gate on
+its mass instead of rounding the expected-value float — the docs say score levels are weak
+in numerical calibration and warn against interpolating between levels.
+
+Access: `model_answered = jev-1.13.0`. Latency 1,779 ms cold, then 521 / 534 ms. Usage
+973 input / 107 output tokens per call → about $0.00004 per teach-back at $0.042 / Mtok.
+Cost is not a factor in any later decision.
+
+## Findings
+
+1. **Access works and the SDK matches its docs.** `TypeSafeClient()`, `Score`, `Noul`,
+   `client.system_one(state=, questions=, model=)`; answers carry `score`, `confidence`,
+   `probabilities`, `legend`.
+2. **Consistent, not deterministic — confirmed by probe, not by prose.** Across identical
+   calls the Score range was ≤ 0.06 on a 0–3 scale (two dimensions identical to three
+   decimals), Noul range ≤ 0.05. Any Stage 1/2 test must assert a band or a level, never
+   an exact float, and CI must replay recorded fixtures rather than call live.
+3. **The pedagogical dimensions landed where a human would put them.** Own words, structure
+   and depth scored within the human expectation with confidence ≥ 0.87; depth 3.03 is
+   exactly the "WHAT, HOW and WHY, but no tradeoffs" reading of the text.
+4. **Accuracy was blind to the planted error.** 82 % of the mass sat on the two "accurate"
+   levels and the error Noul stayed at 0.43–0.48, i.e. "don't know". This is the vendor's
+   documented *literal reading* / *no technical precision* edge: Jev is a common-sense
+   judge, not a Python-semantics checker. **But the confidence signal worked** — accuracy
+   was the least-confident dimension (0.52 vs ≥ 0.65 elsewhere), so a confidence gate
+   would have routed it to "ask the learner" rather than recording a 3.
+5. Negative control held at 0.01: the Noul is not agreeing with everything.
+
+## Design consequences carried into Stage 1 / Stage 2
+
+- **Split by strength.** Correctness is domain knowledge; pedagogy is calibrated
+  judgement. Stage 2 should have the mentor agent (harness LLM, which does know Python
+  semantics) supply the *accuracy* judgement, and Jev score the four pedagogical
+  dimensions — own words, structure, depth, transfer — where it excelled and where a
+  generative model is weakest at calibration. Do **not** ship Jev as a sole accuracy judge.
+- **Argmax + gate, not rounding.** Map a Score to StudyLoop's 1–4 as `argmax(probabilities)
+  + 1`, and record it only when the argmax mass clears a threshold pinned to `jev-1.13.0`;
+  below it, the dimension is "not scored — ask the learner" (the same never-fabricate rule
+  the parked Bedrock extractor taught: `history/teachback.py` must never receive a guess).
+- **Hypothesis for the Stage 2 gold set, not a result:** the two least-confident dimensions
+  here (accuracy 0.52, transfer 0.67) were the two a human would hesitate on. Whether
+  confidence tracks human disagreement is the first thing the labelled set should test.
+
+## Limits
+
+One synthetic state, three repeats, no gold labels. Nothing above is an accuracy figure;
+it is a shape check that surfaced one hard constraint (finding 4) early enough to design
+around it.
+
+## Reproduce
+
+```bash
+set -a; . /path/to/studyloop/.env; set +a   # TYPESAFE_API_KEY only; never committed
+uv run --no-project --python 3.12 --with typesafe-sdk==0.7.0 python scripts/eval/jev_stage0_spike.py
+```
+
+Writes `stage0-receipt.json` beside this file (the committed copy is the run described
+above). Numbers will differ slightly on re-run — see finding 2.