Attempt 1
FAIL
Kane confirmed the add action accepted the name but did not place the task in the list or update the count. AC-001, AC-002, and AC-003 mapped to FAIL.
Every attempt leaves a record.
Runs live under .elenchos/runs. The public evidence page is a sanitized FAIL to repair to PASS loop. Private Kane session bundles stay local.
Where evidence lives
.elenchos/
└── runs/
└── <run-id>/
├── run.json
├── attempts/
│ ├── 01/
│ │ ├── application.stdout.log
│ │ ├── application.stderr.log
│ │ ├── kane.stdout.ndjson
│ │ └── kane.stderr.log
│ └── ...
└── workspace-evidence/Elenchos records state transitions, attempt metadata, Kane structured events, mapped criteria, screenshot or dashboard references when Kane reports them, and a workspace patch after a detached worktree run.
Closed-loop proof
Verified locally on August 20, 2026 with Kane CLI 0.8.4 and AGY 1.1.16 using Gemini.
Repair 1
Gemini
Elenchos sent that structured product-failure evidence to AGY using Gemini. The agent changed only demo/target-app.mjs in the detached worktree.
Attempt 2
PASS
Kane reran the unchanged _test.md contract. All three criteria mapped to PASS. The detached worktree was cleaned after evidence capture.
Run status: VERIFIED
Attempts: 2
Changed file: demo/target-app.mjs
Baseline commit: eaa25ba6d727a45aa09194e4f13fdb9f7986af8c
Kane contract SHA-256: 3e6a018b105155ee470451b74e17ae74080f14dbcdfb8d6da4140835c84ab739
Verified diff SHA-256: 61fb47f7ccd6b50bc8b5157a501e957341f95e5ec5dfa2f50b3617477aeaf184What is public, what is not
- EVIDENCE.md in the repository is a sanitized summary with stable hashes.
- `.elenchos` and `.testmuai` stay out of Git and out of the npm package.
- Human summaries omit raw evidence paths.
- MCP status responses are redacted in the same way.
Included demo
The repository ships a small Proofboard-style app. From a clean checkout:
npm install
npm run build
npm test
npx elenchos init --force --start "node demo/target-app.mjs" --url http://127.0.0.1:3000 --agent agy
npx elenchos run demo/tasks/add-task.json