Repository navigation
Seed learning_approximation and restore its four local references (blocked on action-translation#118) #210
Description
Activity
Folding in #235, which was filed 2026-08-05 during a planning-register prune without finding this issue — it tracks the same reinstatement and is being closed as a duplicate. Two things it adds are worth keeping here:
- The parity failure reproduced deterministically on two consecutive lecture-adding syncs: lecture-python.myst#1004 (source PR #1001) and lecture-python.myst#1007 (source PR #1006), both carrying the action-translation#118 fence-wrap signature.
- This lecture is the cheapest deterministic reproducer the estate has for action-translation#118 — use it as the validation case when that fix is attempted. An alternative unblock short of the engine fix: hand-seeded
init -fwith manual wrapper repair, as was done for the tenth Phillips lecture.
Status note: #209 merged 2026-08-06, so the four cross-references now point at the English edition — the state this issue's opening paragraph describes is live (though unpublished, like everything since 2025-11-06, pending project-translation#25).
- added a commit that references this issue
on Aug 7, 2026 New data point (2026-08-07): the sync behind #230 re-introduced the dangling
learning_approximationToC entry that the 2026-07-27 #207 repair removed — the sync mirrors the source ToC wholesale, so every lecture-adding sync will re-dangle this entry (and fail the strict preview) until the lecture is actually seeded. Removed again on the #230 branch (6244cbc). This turns the reinstatement from a nice-to-have into a recurring CI cost: each full-ToC sync now needs the same hand edit. Raises the value of the interim hand-seed path (init -f+ manual wrapper repair) if action-translation#118 stays unfixed much longer.- added 5 commits that reference this issue
on Aug 7, 2026 Tally closing out 2026-08-07: the ToC entry was reintroduced four times in one day across three PRs — by #230's sync run, twice on #231 (once by the sync, once by the rebase automation overwriting the hand trim within minutes of #230's merge), and by the automation again on #242's branch. Every reintroduction costs a hand trim plus a 20–60 minute preview cycle, and the automation races repairs after every merge to main. That is the recurring cost curve this issue predicted, now measured rather than projected. The cheapest exits remain action-translation#118's fix, or the interim hand-seed (
init -f+ manual wrapper repair).- added a commit that references this issue
on Aug 20, 2026 Done and closing — both halves of this issue.
The lecture is seeded.
learning_approximation.mdis onmainas of #261 (merged 2026-08-20), delivered withtranslate init -fat engine v0.26.0 against sourcea4fbd865. Coverage moves to 144/145; onlymccall_riskremains (#258).The four local references are restored. They had been rewritten as
{doc}`intermediate:learning_approximation`（英文版）— intersphinx links to the English site with an "English version" annotation, because the lecture did not exist here. All four now resolve locally as{doc}`learning_approximation`: one inbounded_rationality.md, three ingenetic_classifier.md. Verified onmain: zero intersphinx references to it remain, four local ones present.On the blocker
This issue was titled "blocked on action-translation#118", and that framing needs qualifying rather than simply lifting. The parity failure is real and it reproduces on the
initpath, so it is not sync-specific. But it is not deterministic: across six attempts the run refused four times and succeeded twice, always with the same signature — the output drops the leading```{raw}header block, so directive #1 shifts{raw}→{index}and the count lands one short. The guard is fail-closed and wrote nothing on any refusal.So the lecture landed on a run that happened to pass, verified after the fact rather than trusted: 36/36 directives, 17/17 code cells, 15/15 headings, 0 python syntax errors across all 17 cells, 4/4 captions localised with none byte-identical to source, i18n font block present.
action-translation#118 stays open, and the dropped-leading-raw-block signature is recorded there as a narrower target than "wholesale mismatch".
One caveat for anyone re-running
initon this edition: it reverted all 19 localised_toc.ymlpart captions to English and added a dangling- file: mccall_riskentry for a lecture that does not exist here, which would have broken the build. Both were reverted before merge; the ToC carries one added line. That is the action-translation#254 class, on a path that issue did not previously cover.
learning_approximationis the one lecture in the bounded-rationality suite this edition does not have. Four references to it currently point at the English edition via intersphinx (#209); once it is translated here they should become plain local{doc}references and the lecture should join_toc.yml.How it got this way
The sync run behind #207 named
learning_approximationin_toc.ymlbut never wrote the file. It failed the structural parity check — the engine refusing to commit corrupt output, which is the guard working — and the run committed the ToC entry anyway (action-translation#156). The same drop happened silently in #208, which wrote no ToC entry, so nothing dangled and nothing reported it.The parity failure is deterministic and reproduces on demand:
That is action-translation#118 — the
{raw} jupyterheader block and document anchor dropped from above the title.Why it is blocked, and on what
A measurement on 2026-07-27 identified the cause: the localisation rules. Holding model (
claude-sonnet-5), engine (v0.24.0) and glossary fixed and varying only--localize, the structural pass rate is 5/15 with the rules against 14/15 without (Fisher exact p = 0.002), with every failure carrying the signature above. So seeding this lecture today with default settings is roughly a one-in-three proposition, and each failed attempt is a silent no-op.Details on action-translation#118 and action-translation#225.
What to do, once #118 is fixed
translate init -f learning_approximation.md, on the released engine, with the localisation rules on (so it matches the rest of the edition: font config, localised figure labels and code comments).lectures/_toc.ymlin theBounded Rationality in Macroeconomicsblock, betweenolg_adaptive_moneyandexchange_rate_learning, matching the source order.{doc}links and drop the （英文版）markers:lectures/bounded_rationality.mdlectures/genetic_classifier.mdWorth knowing
This lecture is the estate's cheapest deterministic reproducer for #118, so it has value beyond the coverage gap: it is the natural acceptance test for any fix, and is named as day-one corpus in the engine benchmark plan (
PROJECT-PLAN-ENGINE-BENCHMARK.mdin QuantEcon/project-translation, tracked at action-translation#227). Do not seed it opportunistically before the fix lands — a successful roll would remove the reproducer without fixing anything.