RepoMod-Bench: a Parity-audited porting workflow on all 21 cases

The complete evidence of one experiment: Bosun's Parity slices workflow (survey → plan → per slice: Parity proof, implement, verify, repair → conform → finalize) run on all 21 cases of RepoMod-Bench, three repetitions each, one model (GPT-6.1 Sol), every proof in audit mode, a 4-hour limit per trial as in the paper. 63 trials, 7 to 9 November 2026.

Headline, from cases: mean of the 21 per-case pass rates 85.4 % of hidden tests (the paper's best agent: 48.2 %; its best agent per case: 57.6 %); on the 12 cases of 50K lines and more 74.4 % against 26.7 %. Cost 781.85 credits, about $687 of inference, 2.44 B prompt tokens (91.5 % cached), 20.1 M completion tokens, 7,754 minutes.

The dataset also carries what the score cannot: every behaviour difference the Parity proofs surfaced, classified, and graded against the benchmark's own hidden tests (coverage): 205 of the 411 behaviour fixes are behaviour the hidden suites never exercise, 151 more are reached only by an exit-code check.

The workflow

workflow.yaml at the root is the Parity slices manifest every one of the 63 trials ran, byte for byte (sha256 f4314c56548b8a96…, frozen before the first case started and recorded in the run's experiment.json). It holds every step's prompt, the typed stop schemas, the fixed audit policy block and the finalize checks, including the policy-integrity gate. Each trial's archive carries the same file as inputs/workflow.yaml next to the case prompt (task.md) and the model catalog (models.toml). The harness that ran it is referenced in harness.json.

Tables

Load any table with datasets: load_dataset("bosun-ai/repomod", "cases"). Nested values are JSON text. Trial ids are <case>--<workflow>--<model>--<repeat>.

Config	Rows	What a row is
cases	21	one case: paper line count and test count, the four paper agents' pass rates, our mean/min/max pass, build, minutes, runs, cost and Parity status counts per workflow (slices is this run; baseline, staged, audited are the earlier control tiers on 12 cases)
trials	147	one trial of this run or an earlier tier: grade (passed/eligible), wall minutes, tokens, credits, USD
tests	34,812	one hidden test of one trial: source and target outcome
steps	1,253	one workflow step of one trial: minutes, tokens, credits
slices	188	one slice of one trial: minutes, tokens, credits
parity	147	one trial's Parity record: slices, proof cases, verify verdicts, repairs, fixes, exclusions (JSON list), proof changes, conform result, final verify counts, finalize status
usage, stages, artifacts	63 / 1,750 / 514,168	credit ledger per trial; per-step usage and log records; every evidence file with size and sha256
paper	84	the paper's Table 4: one case × agent pass rate, with model, lines and test count
diagnostic	6	pdfcpu and uncrustify regraded with their fixtures copied in (not part of the score)
fixed_items	918	every fix the repair and conform steps reported, with a hand label: behaviour, side-effect, runtime-noise, build, process, unclear
exclusion_items	1,052	every proof-scoped exclusion the agents added, labelled runtime-noise, nondeterminism, behaviour-hiding or unclear, with broad
coverage	644	each behaviour fix and each difference left open in a 100 %-scoring port, graded against the hidden tests: covered / partial / uncovered / not-behaviour, a behaviour class, the tests consulted and a one-line example
events	193,969	the full agent transcripts: one event per row (kind is the event type, response its JSON)

Evidence

evidence/<trial_id>.tar.zst (63 archives, 1.3 GB) holds trials/<trial_id>/** exactly as the run produced it: inputs/ (workflow manifest, task prompt, model catalog), agent/ (the Parity project with specs and lockfiles, verify reports, per-slice and per-stage records, finalize logs), evaluation/ (junit and pytest logs of the hidden suite), usage/events.jsonl (the raw transcript), final/ (the port as submitted, without build output or .git), diagnostic-0001/ where a diagnostic regrade ran, and the trial's result and log files. Extract with tar --zstd -xf. MANIFEST.json lists every published file with size and sha256 and, under omitted, the agent scratch files that were left out (329,035 files, 61 GB) with their hashes.

Provenance and method

• Workflow manifest frozen at sha256 f4314c56… before the run; harness commit in harness.json. Config: 3 repetitions, seed 42, 40 credits per trial, 1,500 total, 4 h per trial.
• Grading: the upstream pytest suites, run from a writable copy; eligible tests are those the source itself passes (tests skipped on the source are outside the denominator). Paper numbers are copied from the paper's tables, not rerun.
• Controls: baseline (the upstream prompt in one coding-agent step), staged and audited rows come from earlier tiers on 12 of the cases; the 9 largest cases have no control.
• Classifications and the coverage grading were done by model reviewers from the fix text and the hidden test files, with a written rubric; the evidence column names the tests consulted.

Privacy and secrets

The package was scanned for credential patterns before export. All matches are the fake credentials that gitleaks and httpie-xh ship as test fixtures and taskwarrior's sync.aws.* configuration names; the hosting machine's real credentials were searched for by value and are absent. Local filesystem paths were replaced by <home> and <repo> in the tables; the raw transcripts inside the archives are unmodified. No personal data beyond what the upstream repositories themselves contain.

Licence

Everything the run produced (tables, workflow manifest, Parity proofs and reports, transcripts, logs, classifications) is under the Apache License 2.0. The ported code in final/ inside each evidence archive is a derivative of its source repository and stays under that repository's licence, which for bcal, charcoal-cli, uncrustify, qalculate and lightningcss is GPL, AGPL, LGPL or MPL. The per-case list is in LICENSE.md.

Caveats

Sol is a newer model than the paper's agents ran, and 9 cases have no baseline control, so the gap to the paper is "our workflow on our model", not Parity's gain. Three runs against the paper's one. Our grader is not the paper's. pdfcpu and uncrustify score 5.7 % and 0 % because their hidden tests read fixtures from inside the port, which no agent copied; see diagnostic.