Figure captions ship untranslated: 74 of 237 on lecture-python.zh-cn across 18 lectures (re-measured 2026-08-20; was 85 of 222)

[mmcky]

85 of 222 mystnb figure captions (38%) on QuantEcon/lecture-python.zh-cn are byte-identical to the English source — untranslated, and reader-visible on pages whose prose is fully Chinese. Example live text: "Relaxation algorithm paths, undamped and damped".

Found by sweeping the site after the _toc.yml caption repair in QuantEcon/lecture-python.zh-cn#245. The ToC captions turned out to be the visible half of a larger localisation hole.

What is affected

Verified by comparing every caption: value against the same file's caption in QuantEcon/lecture-python.myst — byte-identical means untranslated, which avoids the false positives a script-detection test produces.

Lecture	Untranslated captions
phillips_drifts_volatilities	26
merging_of_opinions	8
phillips_two_stories	7
olg_adaptive_money, pricing_information	5 each
exchange_rate_learning, genetic_classifier, lq_robust_smoothing, theil_2, var_subsets	4 each
bounded_rationality, market_diffusion	3 each
prospects_bounded_rationality	2
lq_robust_bewley, misspecified_recovery, phillips_escaping_nash, phillips_learning, phillips_self_confirming, sargent_surico	1 each

Which path introduced them

Tracing the commit that first introduced each English caption:

Introduced by	Lectures	Captions
sync (organic [translation-sync] PRs)	15	68
bulk seed (init, the 2026-07-22 37-lecture seed)	4	17

So it is predominantly the sync path — consistent with #178's thesis — but not exclusively. init left 17 as well, in merging_of_opinions, lq_robust_smoothing, theil_2 and misspecified_recovery. Any fix scoped to sync alone would leave those.

The current release is affected. Delivery tool-versions across the 19 lectures are 0.20.0, 0.24.0 and 0.25.0 — pricing_information and market_diffusion were both delivered by v0.25.0 on 2026-08-07 and both carry English captions.

This is a gap, not an absence

The engine clearly can translate captions: 137 captions on the same edition are correctly in Chinese, arriving via init, resync and hand translation. And QuantEcon/lecture-intro.zh-cn is completely clean — 115 captions, zero byte-identical to its source. So the capability exists and something about the sync path (and part of init) skips it.

Blast radius is one edition, for a structural reason

Swept all six editions:

Edition	Captions	Untranslated
lecture-python.zh-cn	222	85
lecture-intro.zh-cn	115	0
lecture-python-programming.zh-cn / .fa / .fr / .ml	0	—

The whole lecture-python-programming series has no mystnb captions in the English source at all, so those four editions are unaffected by construction rather than by correctness. If captions are ever added upstream there, they inherit this bug.

One false-positive class worth recording for anyone writing a checker: lecture-intro.zh-cn's inequality.md and long_run_growth.md carry 19 captions written as "洛伦..." — double-quoted \u escapes that YAML decodes correctly and that render as Chinese on the published page. A naive "does this line contain CJK codepoints" test flags them; comparing against the source does not.

Suggested fix

Same family as #178 (sync applies no localisation rules) but a different surface — YAML caption: metadata inside {figure} / code-cell directives, not plot labels or font config inside code. #178's evidence and likely fix do not reach this.

Worth deciding alongside #254, which proposes a deterministic target-script assertion for _toc.yml captions: one check could cover both surfaces, since the predicate is the same — a target-language file should not contain a caption byte-identical to its source. Deciding that before #254 is built narrowly would avoid writing the same check twice.

Repair of the existing 85 is a separate, hand-fixable pass on the edition; it is prose-only, so it publishes without a cache run.

[mmcky]

Ruled by @mmcky, 2026-08-18: one shared deterministic check with #254 — see the ruling recorded there. Consequence for the held set this issue gates: the 16 python.zh-cn lectures (lq_robust_smoothing + 15 sync-delivered) stay held until the shared check and the accompanying engine fix land, at which point re-delivery repairs them without overwriting hand work. The seed-delivered residual was already hand-fixed and merged (lecture-python.zh-cn#252, theil_2).

[mmcky]

The held set grew because the 2026-08-19 drain shipped two fresh regressions — re-measured 2026-08-20

Re-measured lecture-python.zh-cn@origin/main (1965827) against lecture-python.myst@origin/main while re-verifying project-translation#37. The held set has been carried as 16 lectures. It is now 18 lectures / 74 captions, and the two additions are not a counting correction — they are new damage.

The two new ones arrived on 2026-08-19, in merged sync PRs

Both lectures had zero English captions before the drain and one after:

Lecture	Merged by	New English caption
imp_sample	zh-cn#232 (e7e518c, 2026-08-19)	caption: 'Monte Carlo and importance sampling estimates'
util_rand_resp	zh-cn#237 (b88dbb9, 2026-08-19)	caption: Iso-variance curves

So this class is actively recurring on the sync path, not a static backlog awaiting repair — every lecture-touching sync is a fresh opportunity to introduce one, and the drain that took the backlog to zero introduced two. That is a stronger argument for the deterministic check than the existing held set, because it bounds the growth rate rather than the stock.

Current measurement, and a caveat on how it is taken

Resolving YAML block scalars before comparing: 74 byte-identical captions across 18 lectures. A naive line-based scan reports 83 across 20.

The nine-caption gap is entirely false positives on the block-scalar indicator itself: merging_of_opinions.md matches on | eight times and misspecified_recovery.md once, while the caption bodies on the following lines are fully Chinese in both. Two further lectures — lq_robust_bewley.md (1) and lq_robust_smoothing.md (4) — also use block scalars, and there the naive scan reaches the right verdict for the wrong reason: it compares the indicator, not the body, and the bodies happen to be English.

That matters for #254's design and is recorded there in full: a line-based predicate over caption: has a false-positive class and a blind spot at the same time, so the check should compare parsed caption values.

The 18, with English-caption counts

bounded_rationality 3/3 · exchange_rate_learning 4/4 · genetic_classifier 4/4 · imp_sample 1/7 · lq_robust_bewley 1/1 · lq_robust_smoothing 4/4 · market_diffusion 3/3 · olg_adaptive_money 5/5 · phillips_drifts_volatilities 26/26 · phillips_escaping_nash 1/5 · phillips_learning 1/5 · phillips_self_confirming 1/4 · phillips_two_stories 7/7 · pricing_information 5/5 · prospects_bounded_rationality 2/2 · sargent_surico 1/1 · util_rand_resp 1/5 · var_subsets 4/4

phillips_drifts_volatilities alone accounts for 26 of the 74. Reconciling against this issue's title figure of 85: minus theil_2's repairs merged in zh-cn#252, minus the 9 block-scalar false positives, plus the 2 new regressions above, against a corpus that has itself grown to 237 captions since the 2026-08-10 measurement.

[mmcky]

Field evidence, 2026-08-20: caption localisation is unreliable, not absent — which strengthens the case for the deterministic check

Four lectures were regenerated today at engine v0.26.0 while recovering the estate's sync-failure backlog. All four had captions in source. The engine's raw output:

Lecture	Edition	Path	Captions localised
prob_dist	lecture-intro.zh-cn	forward	19 / 19
hoist_failure	lecture-python.zh-cn	forward	7 / 7
learning_approximation	lecture-python.zh-cn	init	4 / 4
mobility	lecture-intro.zh-cn	init	0 / 4

mobility's four captions came through byte-identical to English — the class this issue tracks — and were localised by hand before merge (lecture-intro.zh-cn#299). Its nine matplotlib axis labels, one legend label and one Series index were also byte-identical to source, and no i18n font block was emitted despite the lecture plotting figures.

The point for this issue is the inconsistency. init produced correct captions on learning_approximation and untranslated captions on mobility, same session, same engine version, same edition family. I first read that as path-specific — forward works, init does not — and that generalisation is wrong.

It sits alongside other non-determinism observed the same day on the same runs: learning_approximation refused its structural parity check on four of six attempts and passed on two, always with the identical signature (the output drops the leading ```{raw} header block, so directive #1 shifts {raw} → {index} and the count lands one short). hoist_failure refused once, then passed.

Why this matters for the design. An unreliable rule is harder to live with than a missing one. If captions were never localised you could special-case them; if they are localised most of the time, every individual run looks fine and nobody can tell from the output whether the rule fired. That is precisely the case a deterministic post-generation assertion exists to cover, and it argues against ever treating "the localisation rules handle it" as sufficient.

It also means the assertion needs to run on every write path, not just sync — init produced the defect here, and forward produced clean output on both of its runs but has no guarantee of doing so next time.

Related: the predicate defect recorded on #254 (a line-based caption: comparison has both a false-positive class and a blind spot on YAML block scalars) applies to whatever assertion lands here.

[mmcky]

Design draft for the shared check posted on #254 (#254 (comment)). For this issue's surface: the predicate reads the mystnb figure: caption: value by parsing the code-cell option header as YAML (so block scalars and \u escapes compare by value), pairs cells by position, exempts identical values with no Latin word of ≥3 letters, and runs beside checkStructuralParity on sync, init and forward — with one bounded retry before refusing, given the 2026-08-20 unreliable-not-absent finding. The 74-caption held set repairs after the release ships.