Repository navigation
WO52 M0: corpus/recipe contracts + isxml validate - #294
Merged
Merged
Conversation
added 4 commits
October 17, 2026 18:15
L0 contracts for the unified framework (TODO.sota/52): pydantic schemas for corpora and recipes, cross-registry validation, and the isxml CLI validate verb. Standing laws enforced at validation time: mandatory test split, pinned corpus refs (date/version, never moving refs), implementable stages (route or params), and the no-LLM-teacher ban for diacritization recipes. Shipped urd-diac + fas-diac recipes and corpora with route:legacy pointers to the proven training scripts; UNVERIFIED licenses block publishing by design until the provenance review lands. Also records WO51: parity gates must evaluate in the runtime decode mode — single-row ORT matched torch exactly on all 300 rows (CER 0.0352, agreement 1.0); the maskless batched harness was the entire "export defect".
Runner.launch/build_argv encode the campaign's operational laws: always detach, explicit --timeout floored at 4h, literal argv job commands (never shell-joined), dry-run prints the exact argv and spawns nothing, and launches are refused at the 2-concurrent-training budget (run-* names, RUNNING state) unless forced. 8 tests written first and watched failing; full suite 221 green.
plan_stage translates a recipe's legacy-routed stage into the exact budget-guarded JobSpec (params map to the legacy U_* env convention, ckpt bucket rw, corpus sources ro at declared mounts); stages without a route are metadata — decode policy travels with the artifact. Gates evaluate in the runtime's decode mode (WO51) and fail closed on missing results. Proof lane against the real run-066 artifacts: quality PASS (0.0352<=0.045), parity PASS (single-row agree 1.0, three-way 1.0), license FAIL (UNVERIFIED) — the spine refuses to publish incomplete work. 7 new tests, suite 228 green.
4 tasks done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
WO52 milestone M0 of the unified framework (spec:
TODO.sota/52-unified-framework.md):src/pipeline/contracts.py): pydantic schemas forCorpusSpecandRecipeSpecplus cross-registry validation.Enforced at validation time: mandatory test split, corpus pins are
dates or versions (never moving refs like
latest), every stage isimplementable (legacy route or params), metric gates use known keys
(cer/wer/per), and diacritization recipes may never declare an
LLM-teacher stage (standing law, validator-enforced).
isxml validateCLI verb (src/pipeline/cli.py, console scriptregistered) — validates
corpora/+recipes/, exits nonzero onany problem.
recipes/urd-diac.yaml,recipes/fas-diac.yamlwith
route: legacy:pointers to the proven interscript-trainscripts, and
corpora/urd-dia.yaml,corpora/fas-dia.yaml. Sourceswith
UNVERIFIEDlicenses are blocking problems by design until theprovenance review resolves them (the test asserts exactly this open
set, so no silent drift).
TODO.sota/51-artifact-parity-gates.md): paritygates must evaluate in the runtime's actual decode mode — single-row
ORT matched torch exactly on 300/300 rows (CER 0.0352, agreement
1.0); the maskless batched harness was the entire "export defect".
fp16 single-row recheck in flight.
Test plan
pytest tests/test_pipeline_contracts.py— 10 tests, writtenfirst and watched failing (ModuleNotFoundError) before implementation
pytest -m "not slow and not gpu"— greenruff check src/pipeline tests/test_pipeline_contracts.py— cleanisxml validatesmoke — exits 1 with exactly the two expectedUNVERIFIED-license problems