Proofline

Proofline turns a coding agent's "done" into an evidence-backed decision. Instead of trusting the model's own account, it applies allowlisted edits, runs protected Java/Maven acceptance checks in isolated, network-disabled Docker containers, and records scoped verdicts in a SHA-256 hash-chained evidence log. Verdicts are PASS, FAIL, INCONCLUSIVE and NEEDS_CLARIFICATION; PASS means the declared checks passed, not that the code is universally correct. It now also works on an external Maven repository. A person can inspect the per-requirement gaps and check results, record an approval or rejection in the same evidence chain, and use `proofline gate` to require approval for the exact current change, base, manifest and protected tests; any change after approval invalidates the gate. From a clean clone I reproduced the full flow: FAIL on an intentionally incomplete implementation, a fix, a fully verified PASS, human approval, gate exit code 0. The repo includes `examples/inventory`, a Maven project outside the original fixtures, with an end-to-end test on real Maven. The 2:37 video shows a real Muse Spark run on a defect we seeded (verification fails, the model's fix is verified independently), followed by the review and gate flow on an external Maven project, which uses no model: FAIL, human rejection with a reason, fix, fully verified PASS, approval, gate exit 0, and gate exit 12 after one line is edited post-approval. Model evidence and software tests are separate. Discount-service: 102 real Muse Spark runs, which are the same 10-task battery repeated three times, not 102 distinct tasks; the first two rounds asked for JSON in the prompt, only the last used native tool calling; only summaries and one trace per task are published. Retry-policy: three real RP1 runs from an empty skeleton (3/3 PASS, one model call each, no repair) and one RP2 run repairing a boundary defect we seeded in an existing implementation: one repair case, not spontaneous bug detection. The new review workflow was built and tested with a simulated model and used no additional real calls. The suite has 295 automated tests. Limitations: Java/Maven only; few fixtures and one external example; results depend on the declared checks; the reviewer's name is typed, not authenticated; the chain is unsigned and unanchored; the gate recomputes hashes but does not rerun checks; the sandbox is hardened but not a complete security boundary. Runs used Meta's contributor tier (inputs and outputs may be used for training), so only synthetic code was sent. A prototype for auditable, human-controlled agent verification, not a proof of correctness. Open source (MIT). Built solo between Oct 27 and Nov 4, 2026; the first commit is dated Oct 27. Repo: https://github.com/valentinlineiro/proofline Demo: https://youtu.be/-1lO7nwPswk