TRACKING: work plan 2026-07-22 — rubric v2 quick-wins, merge + email, skill wiring toward the #654 re-test

[mmcky]

Coordination issue for tomorrow's session, sequenced. Detail lives in the linked issues; this tracks the day.

Milestone this points at: the skill runs a full evaluation of lecture-python.myst#654 under the v2 rubric, @xuanguang-li adjudicates it via the expert-parallel protocol (#4), and that becomes both the skill's acceptance test and the one authoritative comment that PR thread receives.

1. Rubric v2 engine quick-wins (#7, items 1–4)

• Enforce the safety couplings in score_all (derive the logic-design cap from builds/x64-divergence; verdict gates: correctness 1 → at most "net regression", correctness 2 → at most "mixed/wash")
• Add the "no-conversion" verdict output (reconciles review mode with triage)
• Sensitivity stamp in score.py (perturb booleans + band-adjacent values; mark scorecards robust/fragile with deciding flips)
• K-repeat as-used measurement (median of ≥3 fresh-process runs; contested-band annotation)
• Re-validate: both committed evidence files re-scored; confirm ge_arrow/markov_asset verdicts and document any band movements as deliberate v2 changes

~50 lines total, each validatable in seconds. Doing this first means Kenko reviews a system already hardened by the design review.

2. Merge order + the Kenko email

• Decide: merge Land the lecture evaluation system (benchmark plugin 0.3.0: rubric v2, skill wired) #5 + Docs: skill usage, repo setup, and validated triage mode #6 as the v1 record with v2 as a follow-up PR (lean), or fold v2 into the stack first
• Final pass on the email draft (add the PLAN: evaluation rubric v2 — enforce couplings, verdict vocabulary, instrument fixes (from the three-way design review) #7 design-conversation invitation; asks: PR reviews, the three design calls, readability-instrument input, expert-parallel test) — then send

3. Skill wiring (#4 item 3)

• ${CLAUDE_PLUGIN_ROOT} anchoring + evaluation working directory in the user's workspace
• Operational hardening of the SKILL.md procedure (drive scaffold → measure → evidence-with-citations → score → report end-to-end)
• Dry run on a checkout to surface adaptation failure modes before the real #654 run

4. Parallel track (as time allows)

• Ping @jstac on QuantEcon/project-style-guide#6 (Stage 0 ratification gates the style flagship)
• Start the style preflight engine (PLAN: qe style skills — from scaffold to operational /qe:check-style #3 item 1: MyST context splitter + rule loader + fixtures runner against the 5 POC rules)

Standing context

Design review: reviews/benchmark-design-2026-07-21-merged.md. lecture-python.myst#654 has a deliberately clean slate — nothing gets posted there until the full v2 evaluation is done and adjudicated.

🤖 Generated with Claude Code

[mmcky]

Dry run complete (§3 done): ge_arrow re-evaluated end-to-end from a fresh checkout of lecture-python.myst#717 (base 8cfba4c, head 8c2d0d7) via the wired workspace procedure, on jax 0.10.1 vs the reference 0.4.35 — reproduces 2.85 / no-conversion / fragile with the same three deciding flips; every measured quantity moved only within its band. Cross-comparison table and the three adaptation findings (x64 clobber in the ge_arrow equivalence template — fixed; undisclosed whitespace normalisation in the baseline extraction — now disclosed; SHA provenance added to both evidence files) in reviews/validation-run-ge_arrow-2026-07-22.md. #654 remains untouched and the acceptance run can proceed on this procedure. All on PR #5 (a56cc1d).

[mmcky]

Email sent to @xuanguang-li 2026-07-22 — asks: PR #5 review, the three design calls on #7, and the #654 expert-parallel test; full protocol in #10. Sections 1–3 of this plan are now complete; remaining items are the parallel track (§4) and everything gated on the review.

[mmcky]

Closing: this tracked the 2026-07-22 session, and its three main sections are done. Sections 1–3 landed with #5 — the four rubric v2 engine changes are in benchmark/scripts/scoring/ (couplings enforced in score_all, no-conversion verdict, sensitivity stamp, K-repeat as-used), both evidence files re-scored, the merge-order decision taken and executed, and the skill wired for workspace runs with ${CLAUDE_PLUGIN_ROOT} anchoring.

Two items from section 4 were never started, and are re-homed here rather than lost with this issue:

• Ping @jstac on QuantEcon/project-style-guide#6. Stage 0 ratification still gates the style flagship. This is the one item with no other home — it is a dependency of PLAN: qe style skills — from scaffold to operational /qe:check-style #3 rather than a task on it.
• Start the style preflight engine. Already tracked as item 1 of PLAN: qe style skills — from scaffold to operational /qe:check-style #3, so nothing is lost by closing here.

The milestone this pointed at — the skill running a full v2 evaluation of lecture-python.myst#654, adjudicated via the expert-parallel protocol — has not happened and is tracked on #4 and #10. lecture-python.myst#654 still has its deliberately clean slate; nothing has been posted there.