validation/PROTOCOL.md

Validation protocol

Precision of the claimed breaking changes was measured by independent audits. The audit agent never had access to the extraction pipeline or the dataset files. It only saw a CSV of claims.

For each claim (project, from_version, to_version, path, change_type, param, detail) the auditor:

1. created isolated virtual environments and installed project==from_version and project==to_version from PyPI, along with any optional dependency needed to import the module;
2. imported the module part of path, walked attributes (including inherited and lazily served ones) and compared inspect.signature where the claim concerned parameters;
3. if installation was impossible (platform, heavy dependencies), read both versions' released source instead;
4. recorded a verdict (confirmed, false_positive, partially, unverifiable), the method used and one sentence of evidence.

Round	Files	Sample	Notes
1	sample.csv, results.csv	60 claims, 32 libraries	first pipeline version: 36 confirmed, 22 false positives, 2 partially
2	sample2.csv, results2.csv	60 claims, 32 other libraries	after fixing round-1 patterns: 43 / 15 / 2
3	sample3.csv, results3.csv	100 claims (80 high, 20 medium confidence), 53 libraries not used in rounds 1–2	released pipeline: high 71 confirmed / 6 false positive / 3 partially; medium 9 / 9 / 2

audits.json re-scores every audited claim against the released tables. A claim counts only if the released data still makes it, at its released confidence tier. Two confidence rules were added after round 3 and should be read as in-sample for that round: removing *args/**kwargs and turning a class into a plain attribute both move to medium. The out-of-sample estimate for high-confidence breaking changes is therefore the as-audited round-3 figure (71/80 = 88.8%, or 92.5% counting partially correct claims). After the post-hoc rules, 67 of the 71 high-confidence round-3 claims that remain are confirmed (94.4%).

checker_false_alarm_test.json records a separate test of apidrift-check. It ran over 1,823 source files of 14 widely used packages, each checked against the dependency versions actually installed with them.

Further studies (all by independent AI agents, no access to the pipeline or data)

Folder	Study	Out-of-sample result (before the fixes it prompted)
benchmark_check/	200 VersionProbe pairs (one per library, libraries not in the precision audits) resolved at runtime in both releases	172/200 labels correct (86.0%, 95% CI 80.5–90.1%)
recall/	Runtime introspection of both releases of 30 random release pairs; every breaking change listed (99)	14/16 (87.5%) within this dataset's public-API definition; 16/27 (59.3%) excluding trio; 16/99 (16.2%) overall
downstream/	Test suites of 12 real projects run before/after a dependency upgrade (9 breaking, 3 controls); API-caused failures located	checker: 7/7 findings correct, 0 on old environments or controls, 4/12 failures predicted

The files ending in _before_fixes.json hold those out-of-sample scores. The files without the suffix hold the scores re-computed on the released data after the fixes: TYPE_CHECKING and Python-2 branches skipped, later redefinitions win, X as X re-exports, imported expired-attribute tables, alias-aware parameter history, and instance attributes excluded from benchmark snippets. Those after-fix scores are in-sample. Downstream source trees are not redistributed; downstream/<case>/freeze_*.txt pins the exact environments, and cases.csv names the project releases.

Second review (all pipeline-isolated AI agents; no human adjudication)

Folder / file	Study	Design
benchmark_check2/	Fresh VersionProbe label audit	400 pairs from the rebuilt benchmark, restricted to 265 libraries never used in any earlier audit, study or fix; at most two pairs per library, signature changes over-sampled. Both releases installed; every name resolved and every keyword bound with inspect.signature; snippets never executed.
downstream2/	Unseen upgrades	Checker and data frozen first (FROZEN_sha256.txt; scored with that checker). Two agents, who never ran the checker, collected 35 cases (25 expected breaks, 10 controls) over libraries split between them, including compiled packages and lazily loaded APIs; test suites run before/after, API-caused failures located to file and line, and masked or unexercised incompatible uses listed separately. comparison_frozen.json is the out-of-sample score; comparison_after_fixes.json is in-sample. Source trees are not redistributed; cases.csv names each release and <case>/freeze_*.txt pins the environments.
tool_comparison/	griffe and AexPy	Both tools run on the 30 recall pairs (commands in NOTES.md, scripts in scripts/); every method scored with one matching rule against recall/ground_truth.csv (comparison.json).
versioning_policy.csv, semver_split.json	Stated versioning policy	Policy sentences matched in each project's PyPI description; every match reviewed by hand (manual_override marks the corrections).
dataset_checks.json	Integrity	pipeline/validate_dataset.py over the released tables.
checker_performance.json	Checker cost	pipeline/bench_checker.py: wall time, CPU time and peak memory on four workloads.
patch_log.json	Corrections	Counts of the corrections applied by pipeline/patch_release.py.

Fourth review

Folder	Study	Design
audit4/	Fresh precision audit of v1.2	160 high-confidence representative breaking changes from 134 libraries never used in any earlier audit or study, stratified 80/80 into claims new in v1.2 (8,274 in the population) and claims unchanged since v1.1; four agents, same protocol as round 3. summary.json gives per-stratum and stratum-weighted precision.
code_repair/	Preregistered code-repair experiment	PREREGISTRATION.md (task list hash recorded before any model output); 150 fresh VersionProbe pairs; three conditions (closed book, griffe-generated documentation of the new release, PyPI API Drift history); runtime check with only the new release installed (pipeline/repair_resolver.py); blind faithfulness judge; exact McNemar tests.
../benchmark/versionprobe_verified_evidence.csv	Verified split evidence	Per-pair runtime verdicts from the two benchmark audits for every pair in versionprobe_verified.parquet.

Sixth review: full-binding sensitivity check and clean-install test (2026-10-21)

• Code repair, post hoc and not preregistered: every task was re-verified in a fresh environment, judging each answer twice, with the preregistered bind_partial and with full inspect.signature(...).bind (pipeline/verify_repair.py ... --strict, scored by pipeline/score_repair_strict.py). The AI-judge verdicts were reused unchanged. Output: code_repair/verify_strict.jsonl, results_strict_binding.json, scored_tasks_strict.csv.
• Clean install: smoke_test.py installs the release non-editably into a fresh venv, runs the tests (including data-level regressions against the downloaded data), verifies the data hashes and runs the installed CLI on the bundled example. Output: clean_install.json. CI has a matching clean-install job on the fixture data.

Third review: placeholder releases, container shims, model replication, human-labelling kit (2026-10-22)

• Placeholder releases (pipeline/drop_placeholders.py, log placeholder_releases.json): an all-zero version uploaded after the release it precedes is removed with its diffs and dependent benchmark pairs. Selection (select_releases.py) and validate_dataset.py now exclude/check them.
• Module __getattr__ container shims (modparse.py: _served_containers, _getattr_container_lookups): the affected libraries were found by re-parsing every dynamic module at the ends of its recorded ranges with the old and new extractor (pipeline/scan_getattr_shims.py) and only those were re-extracted; the rest of the extraction is unchanged. Summary of the data change: v13_changes.json.
• Model replication (multi_model/): VersionProbe-verified and the code-repair experiment re-run unchanged with three more models, judged by the same blind judge, re-verified under both binding rules.
• Human-labelling kit (human_labeling/): 200 items (all 160 v1.2 audit claims, 40 random benchmark-audit pairs), blind sheets for two labellers, an AI key and an agreement script. Not yet labelled.