Meta-analysis (Cochrane-grade): validation

Checks and scope

Writing stage per review (iteration 14, addendum 5 judging).

Render check on every reproduced benchmark review.

Iterations and unchanged limits

Fifteen iterations ran under a frozen contract and five owner-approved addenda; every skill fix was made general (a rule from at least two instances or a Handbook basis) and carries a regression test that fails on the old code. Harness defects (23 numbered issues) were fixed in the scoring harness, not by changing the thresholds. The costs of each iteration are in the log; iteration 15 exceeded its budget by a third when screening cost more than planned, so it stopped before the full-text stage, as the contract requires.

Unchanged limits: six development reviews; only two searched and one screened so far; full-text eligibility not yet measured; the locked runs have not begun; the writing stage still produced four verbatim overlaps; GRADE and eligibility remain drafted for people to decide; databases the scripts cannot reach need a person with access.

Rights and review roles

Original Nexitia-controlled files use Nexitia Skill License 1.1. The three Cochrane data packages bundled as fixtures are CC BY 4.0 works of their authors and are redistributed unchanged with attribution. Of the six development reviews, one (CD016058) is CC BY 4.0 and its abstract sentences appear in the showcase with attribution; the five non-commercial reviews (four CC BY-NC, CD015464 CC BY-NC-ND) are named with their DOIs and agreement statistics only, and none of their text, data or protocol is published. The search and retrieval scripts call public services under their terms and the Cochrane Library was not accessed.

Preparation and technical review were AI-assisted under the owner's authorization. Edition 1.0.0 was the first distribution of the material; this release candidate preserves its grant history, and institutional ownership and enforceability were not reviewed by counsel.

Reproduction commands

PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tests -p 'test_*.py'
PYTHONDONTWRITEBYTECODE=1 python3 tests/run_all.py --round local

python3 scripts/protocol_check.py run/01_protocol/protocol.json
python3 scripts/mesh_lookup.py "term one" "term two" --out run/02_search/mesh.json
python3 scripts/search_build.py run/02_search/strategy.json --out run/02_search
python3 scripts/search_run.py --queries run/02_search/queries.json --out run/02_search --until 2025-01-13

python3 scripts/reproduce.py --package fixtures/CD010216_data_package/CD010216-SUP-07-dataPackage.zip --out ../ma-run-CD010216

Remaining limits

• Release candidate: the 2.0.0 acceptance gate has not been passed; search and screening have run on two of six development reviews, full-text eligibility on none, and no locked run exists.
• Agreement with a published review is a replication result; it is not a re-assessment of the review's judgements and not clinical guidance.
• The review stages were run through Claude Code sub-agents; another host must provide separate fresh contexts for the duplicate passes to be independent.
• Open-access routes only: paywalled full texts and databases the scripts cannot query need a person with access.
• Only Codex CLI discovery of the packaged edition was verified; other hosts and operating systems are untested.

Owner-authorized and AI-assisted. Package identity, a clean-extraction replay of both test suites, the render check, the blind replication record and the showcase are recorded; no independent human statistical review, clinical certification or legal-counsel review is claimed. The 2.0.0 release gate (two consecutive passing iterations on the development reviews, then locked runs on unseen reviews) has not been passed.

Reproduction and package identity

meta-analysis-2.0.0-rc1.zip · 117 files · 1,734,639 bytes unpacked

936632b8e4f4a8af5f05e87c28f2fc925331f2fa261645b3ddff2921a44cc55f