The Malayalam editor's round-2 answers are encoded, and exercise blocks in ml are now byte-identical to the source by construction.
Verbatim exercise-family policy (#303, D-2026-09-03)
The editor of record answered lecture-python-programming.ml#12 and reversed the two-day-old exercise-scope ruling: until he has decided each exercise individually, every {exercise}, {exercise-start}…{exercise-end}, {hint}, {solution} and {solution-start}…{solution-end} block stays in English, byte-for-byte. A rule with no judgement boundary is a copy step, so it is enforced in code rather than the prompt:
src/verbatim-directives.tsrestores the blocks from the English source on all three write paths (sync,translate init,translate forward), immediately before the structural-parity guard so parity checks the bytes that are written. Nesting-aware fence walk (backtick, tilde and colon fences; code cells opaque); blocks pair by position; a differing block sequence is left untouched and logged rather than guessed. Policy table is aMapkeyed by language —mlonly.- Review diff-check
verbatimDirectives: a divergent block surfaces as ablockerfinding in the gatingdiff-checkcategory, routing the PR to the editor. Not a fifth published boolean — the verdict schema is unchanged. - Harness scenario 28 adds the suite's first exercise-family directives. Validated with the local bundle before merge: the
.mllane delivered both blocks byte-identical (test-translation-sync.ml#102), and a hand-translated hint line was caught (recommendationeditor) where the v0.27.0 reviewer passed the same head.
The held ml#12 items (#296)
- Rules 24 → 23: the two exercise-scope rules replaced by one verbatim statement; sentence-initial For example, joins the English-retained connectives; the
refercalque is named (never സൂചിപ്പിക്കുന്നു); headings byte-identical including possessives. - Glossary v0.5.0 (77 → 82 terms): already, example / examples, name (the noun; the act of naming stays പേര്), work (light-verb form).
Also
- #299: the ml round-2 second-draw arm archived under
experiments/ml-benchmark/arms/with a closeness measure (experiment data only).
Upgrade notes: none. runDeterministicDiffChecks gains an optional targetLanguage parameter; callers that omit it see no change. Editions other than ml are unaffected.