ml round-3 arm: matplotlib regenerated at v0.28.0 before review, vs the v0.27.0 seed

.dev/log/2026-09-03-ml-round2-answers.md

@@ -16,3 +16,5 @@
 **Next**: local-bundle validation against a scenario-28 harness PR on the `.ml` lane (expect the blocks byte-identical and the diff-check green), then PR, then release before round 4 is generated.
 
 **Addendum, same day — released.** Copilot's three comments on #303 addressed (`8af4de5`); #303 merged `07e7c64`; release PR #304 (`9284fbc`, Copilot's one comment adopted: an empty `[Unreleased]` header now stays above the released section, recorded in the AGENTS.md checklist). §4a gate on `v0.28.0`: 84/84 sync runs, 28/28 delivery + 28/28 verdicts per lane at `engineVersion: 0.28.0`; scenario 28 on `.ml` delivered both blocks byte-identical (test-translation-sync.ml#126), PASS/auto-merge. `v0.28` + `v0` moved to `9284fbc`; alias smoke on scenario 01 reports `engineRef: v0` / `0.28.0` on all lanes. GitHub release published. W1 (#259) retargeted to v0.29.0. Gate-tally recipe: `gh api …/actions/runs?created=>=<gate-start>` grouped by workflow name and conclusion, then the last verdict comment's `engineVersion` per open target PR — the `created` filter is loose, so key the tally on the runs' own timestamps if the numbers look inflated.
+
+**Addendum 2 — round 3 regenerated at v0.28.0.** The editor had not started ml#13, so the seed was regenerated at the release (ml#13 `9e8616b`, v0.27.0 draw kept as the previous commit) and the pair archived as `experiments/ml-benchmark/arms/2026-09-03-round3-matplotlib-v0.28.0/` (#306). Under glossary v0.5.0 the v0.27.0 draw fails the *already* / *name* pins and needed a heading hand-fix; the v0.28.0 draw is clean on both, lints 0/0/0/1, exercise blocks verbatim by construction; draw-to-draw prose 14% identical lines / 0.683 similarity, the round-2 order of variance. #302 and #305 merged (main `37374d7`); #306 open.

experiments/ml-benchmark/arms/2026-09-03-round3-matplotlib-v0.28.0/README.md

@@ -0,0 +1,49 @@
+# Arm: round 3 (`matplotlib`) regenerated at v0.28.0 — before review, 2026-09-03
+
+The round-3 calibration seed on lecture-python-programming.ml#13 was generated at
+v0.27.0 on 2026-09-01. The editor had not started reviewing it when v0.28.0 shipped
+on 2026-09-03 with his ml#12 answers (glossary v0.5.0, rules 24 → 23) and the
+verbatim exercise-family policy, so the lecture was regenerated at the new release
+and the PR head replaced (ml#13 `9e8616b`; the v0.27.0 draw stays as the previous
+commit). Both draws are archived here because the pair is a clean before/after of
+one release on one lecture with nothing else varied — the first arm whose "reviewed"
+column is still empty, to be filled in when round 3 is applied.
+
+## Provenance
+
+| Field | v0.27.0 draw (ml#13 `1f726ff`) | v0.28.0 draw (ml#13 `9e8616b`) |
+|---|---|---|
+| Model | `claude-sonnet-5` | `claude-sonnet-5` |
+| Command | `init … --target-language ml -f matplotlib.md --localize none` | same |
+| Source | `QuantEcon/lecture-python-programming@4980d62` | same (unchanged) |
+| Engine | v0.27.0 (`125801c`) | v0.28.0 (`9284fbc`), built at main `37374d7` (dev-only diff) |
+| Tokens / wall time | 25.8k / 3.4 min (as logged on ml#13) | 13,168 / 124 s |
+| Hand edits | one heading restored (`Matplotlib-യുടെ` → `Matplotlib's`) | none |
+
+## Scores (`scripts/ml_metrics.py`, glossary v0.5.0 for both)
+
+| Check | v0.27.0 draw | v0.28.0 draw |
+|---|---|---|
+| headings byte-identical | 15/15 **after** a hand restoration | 15/15, no hand edit |
+| pinned-term retention (17 checked) | FAIL: *already* → ഇതിനകം, *name* → പേര് (+ `returns`, the #240 false positive) | clean apart from `returns` |
+| round-2 lints (bare ending / lowercase-initial / banned / hortative-watch) | 0 / 1 / 0 / 1 | 0 / 0 / 0 / 1 |
+| casing variants / transliterations | 0 / 0 | 0 / 0 |
+| exercise + solution blocks byte-identical | yes (by content — the v0.27.0 rule allowed translation) | yes (by construction — verbatim policy) |
+| script ratio (mean / median) | 0.517 / 0.539 | 0.520 / 0.534 |
+| നമുക്ക് (hortative subject) | 8 | 10 |
+| *For example* kept English | 1/1 | 1/1 |
+
+## Closeness between the two draws
+
+Same length (509 lines) and the same 66 Malayalam prose lines, but only 9 of those
+lines are byte-identical across the draws (14%); character-level similarity of the
+prose 0.683. That is the same order of draw-to-draw variance the round-2 arms showed
+(exact-line match 11–21% against the reviewed text), so the gate improvements above
+are attributable to the release, the wording differences are not.
+
+## What to read off it when round 3 is reviewed
+
+Fill in a "reviewed" column from the editor's suggestions on ml#13 and compute the
+closeness of each draw to it. If the v0.28.0 draw is not closer than the v0.27.0
+draw despite the cleaner gates, the ml#12 pins are not where the remaining review
+effort goes — the prose classes (#260's post-processing) are.