See it work

The intelligence behind shipping

The OpenAI agent (gpt-5.4-mini) drafts a post from the app's real page and a Kane Explore map, and turns every promise into a check. Kane CLI runs headless with --agent in real Chrome and stores what it sees; my code judges those values. On FAIL the agent repairs the promise from what Kane saw and Kane checks again. Publish stays locked until Kane's latest result passed that exact version. After a release, Re-verify replays Kane's recording with kane-cli testrun, and Audit past posts checks old posts against today's site.

1. 01Agent drafts post + checks
2. →
3. 02Kane checks in real Chrome (--agent)
4. →
5. 03Code judges Kane's stored values
6. →
7. 04Agent repairs, Kane re-checks
8. →
9. 05Publish gate (PASS + human confirm)
10. →
11. 06Re-verify and audit after release

Components: kane-cli run --agent (headless Chrome) · kane-cli testrun (_test.md replay, Re-verify) · NDJSON output (run_end verdict, stored values) · explore run (site map of scribo.cc) · context ingest (campaign as spec) · design tests (assurance loop)

What Kane actually verified

• ○ Team-reportedOn scribo.cc/?ct=threads-oct24 Kane read the App Store button before clicking: ct=site-en-hero, not the campaign's ct=threads-oct24. FAIL (KR-63). The agent rebuilt the link from what Kane saw: PASS (KR-68). Same FAIL then PASS in all 5 full loops.run 63 · RUNS.md rows KR-63 and KR-68 (Kane run_end: passed; recorder shows ERRORED, ticket H-0007). Other loops: KR-13/14, KR-15/16, KR-27/28, KR-74/75.
• ✓ VerifiedAssurance loop: Kane designed 2 tests from the campaign copy and both passed on the repaired link (excludes ct=site-en-hero, reaches Scribo's App Store page). Coverage 75%; the other 25% (installs credited to the campaign) Kane marked unmeasurable from a browser.run 46 · commit 414c939395 · Kane run KR-46 (kanehack report -- kane-cli testrun t-2 t-3), source-verified pass on commit 414c939.
• ✓ VerifiedRe-verify replayed Kane's own recording of the passing campaign and passed (KR-85; earlier KR-77). Only then did Publish to Threads unlock, and a real post went out through the gate: threads.com/@scribo.voca/post/De4-yPnATF5run 85 · Kane run KR-85 (kane-cli testrun replay of c-2026-10-24T19-41-41Z-a2_test.md), source-verified pass. Logged in RUNS.md.
• ✓ VerifiedDrift 2: Kane checked promises from 3 old @scribo.voca posts against today's scribo.cc; the handwriting promises are gone: FAIL (KR-126). Replaying Kane's recording confirmed promises #1 and #6 again (KR-130). The agent proposes a rewrite or takedown, never edits Threads itself.run 130 · commit 4f0329d847 · Kane run KR-130 (kane-cli testrun replay of a-2026-10-24T20-46-49Z-en_test.md), source-verified. First audit KR-126 in RUNS.md; #5 was not reported on replay.
• ○ Team-reportedSupporting example: on TestMu AI's free JSON tool I seeded a false 'convert JSON to YAML' claim. Kane got JSON back: FAIL (KR-100). The agent swapped in Minify: PASS (KR-102).run 100 · RUNS.md rows KR-100 and KR-102 (Kane run_end: passed; recorder shows ERRORED, ticket H-0007).

Not yet verified: Checks stop at the App Store page; they don't follow the user into the app. · A PASS means the promises are true when the post ships. It doesn't mean installs went up. · A sentence is backed by a phrase Kane finds on the page, which can cover only part of a long sentence.

What improved