Showcase / Made at Kane CLI Hack Day
Validation harness
ProofForge
A browser test can pass while a booking disappears after sign-out. ProofForge uses OpenAI and Kane CLI to show what its evidence actually establishes.
ProofForge is a working, contract-anchored verification calibration harness built with OpenAI and Kane CLI. Its Inspect → Calibrate → Prove workflow compares an independent product requirement with actual test evidence, uses OpenAI to propose stronger checks, and runs those checks through Kane against controlled co…
1 builder credited · submitted Oct 24, 6:00 PM PDT
Kane · proof view
- Define the contract
- Inspect the passing baseline
- Calibrate with OpenAI + Kane
- Review the evidence receipt
- ProofForge baseline; actual Kane session fa1a01a3, pack execution a7bdf778. Corrects incomplete KR-11 reporting; this is the same browser run, not additional evidence.Verified · passed
- Navigate to [local fixture] Sign in as alex using Sign in. Click Confirm booking. Click Sign out. Sign in again as alex using Sign in again. Click Retrieve booking once. Wait exactly 7 seconds once. Finish. Finish after the fixed wait; a separate Kane condition check will evaluate retrieval. Do not assert or analyze the retrieval result in this navigation phase.Verified · passed
- Navigate to [local fixture] Sign in as alex using Sign in. Click Confirm booking. Click Sign out. Sign in again as alex using Sign in again. Click Retrieve booking once. Wait exactly 7 seconds once. Finish. Finish after the fixed wait; a separate Kane condition check will evaluate retrieval. Do not assert or analyze the retrieval result in this navigation phase.Verified · passed
- Navigate to [local fixture] Sign in as alex using Sign in. Click Confirm booking. Click Sign out. Sign in again as alex using Sign in again. Click Retrieve booking once. Wait exactly 7 seconds once. Finish. Finish after the fixed wait; a separate Kane condition check will evaluate retrieval. Do not assert or analyze the retrieval result in this navigation phase.Verified · passed
Product
Validation harness
Run command in repo
Harness
2 Kane parts
kane-cli run · sealed evidence packs
Evidence
5 runs
5 verified · 0 detected
Repo
312c922
github.com/beyond-time/proofforge
One-pager · by the team
Open the one-pager ↗01 / The creation
What it is, and why.
A browser test can correctly pass while verifying only part of the actual product requirement. For example, a booking confirmation can appear successfully while the booking itself disappears after the customer signs out. The original test may have executed exactly as intended, but it never checked whether that booking remained accessible after fresh authentication. This creates a dangerous gap between a passing test and the product behavior a team believes it has verified.
ProofForge is a working, contract-anchored verification calibration harness built with OpenAI and Kane CLI. Its Inspect → Calibrate → Prove workflow compares an independent product requirement with actual test evidence, uses OpenAI to propose stronger checks, and runs those checks through Kane against controlled correct, violating, and admissibly delayed behaviors. A deterministic evidence classifier preserves the original Kane verdicts, evaluates the observed results, and produces traceable receipts distinguishing established claims, relevant failures, and inconclusive observations. In the completed booking-persistence demonstration, the calibrated checks distinguished one violating behavior from two admissible behaviors.
- 01Start with an independently versioned product contract.
- 02Run a baseline Kane test and inspect what its passing evidence actually establishes.
- 03Use OpenAI to identify missing observations and propose stronger checks.
- 04Execute Kane browser journeys against violating, correct, and admissibly delayed behavior.
- 05Evaluate separate DOM conditions against session, booking-identity, and timing observations.
02 / The harness
How Kane checks it.
ProofForge uses an independently versioned contract, BOOKING-PERSISTENCE 1.0.0. After sign-out and fresh authentication, the same customer must be able to retrieve the original booking ID and matching details within six seconds of the retrieval request. The previous session must be invalidated. The baseline Kane CLI test verifies booking confirmation and passes, but does not establish persistence. OpenAI analyzes the missing requirement-level observations and proposes stronger browser actions and seven atomic DOM conditions. Kane executes isolated browser journeys against three controlled behaviors: correct server-side persistence, session-only booking loss, and an admissible retrieval delay. The original Kane navigation verdicts are preserved. Separate condition evaluations on the retained browser pages provide actual boolean results. ProofForge correlates these with independent fixture observations, authentication state, booking identity, timing, and validated navigation evidence packs. The deterministic classifier accepts the correct and admissibly delayed cases, rejects the corroborated session-loss violation, and retains unrelated automation failures or insufficient observations as inconclusive. Actual outcomes return to OpenAI for feedback. Each evidence receipt preserves source identity, execution provenance, original verdicts, and the limits of the supported conclusion.
- kane-cli run
- sealed evidence packs
- 01Define the contract
- 02Inspect the passing baseline
- 03
03 / Evidence
What actually ran.
Verified means a passing run reported through a paired, verified channel. A failing run is a detection, not a confirmed bug. Everything else is as the team reported it.
- Verified · passedOct 24, 11:53 AM PDT
ProofForge baseline; actual Kane session fa1a01a3, pack execution a7bdf778. Corrects incomplete KR-11 reporting; this is the same browser run, not additional evidence.
mcp · fa1a01a3-2986-4cb3-b0ec-866840e027d6 · exit 0 · 94ef689
- Verified · passedOct 24, 12:52 PM PDT
Navigate to [local fixture] Sign in as alex using Sign in. Click Confirm booking. Click Sign out. Sign in again as alex using Sign in again. Click Retrieve booking once. Wait exactly 7 seconds once. Finish. Finish after the fixed wait; a separate Kane condition check will evaluate retrieval. Do not assert or analyze the retrieval result in this navigation phase.
mcp · 1bb24f01-1ef7-461e-b32d-a0c7c5879c5c · exit 0 · e7afc66
04 / Outcome
What changed.
No measured outcomes were claimed.
Known limitations
- The current calibration covers one independently versioned booking-persistence contract and three controlled application behaviors, not arbitrary production systems.
- Authentication uses a local demonstration identity and instrumented fixtures rather than production identity infrastructure.
- The system runs locally with a single-operator execution queue and requires available Kane CLI and OpenAI services.
- Kane DOM condition evaluations do not produce separate sealed evidence packs. Their results are correlated with validated navigation evidence, retained-page identity, and independent fixture observations.
- The observed success-prefix comparison used a contains operator. Strict starts-with behavior outside the recorded examples remains unverified.
- The Codex model backend snapshot is not pinned. Earlier timeouts, unsupported conclusions, and classifier corrections remain preserved as part of the evidence history.
- This demonstration does not establish universal software correctness, production security, broad requirement coverage, or measured developer productivity improvements.
05 / Ship it forward
The code. The proof.
The story.
- Mykyta Kharkevych@mkharkevych
Also submitted
Run it
git clone https://github.com/beyond-time/proofforge.git && cd proofforge && git checkout 312c9222fd202f03ed1403b15829417b426483ea && npm start