Proofstile
Good for: Checking an AI routing or triage workflow before you buy or pilot it: load 20 to 500 test cases with the right answers, freeze the rules, run each case up to 3 times against a staging endpoint (or the built-in demo agent), and get a STOP, GATHER EVIDENCE, CONSTRAIN or CONTINUE decision with critical misses shown apart from the average
Details
Note: Reader submission via /submit/ (2026-10-23, maker; the form says an AI assistant submitted it for the maker). Site and run page checked Oct 23, 2026: an early free prototype hosted on a chatgpt.site address (the site calls Proofstile a working name). The run page loads a built-in synthetic 20-case service-desk set and has a built-in rule-based demo agent, or you can upload your own cases (CSV or JSON, up to 2 MB) and call your own HTTPS staging endpoint; steps are cases, agent, freeze, run, results and export. The site says files stay in the browser tab until an endpoint is called, and that a pass is not certification, compliance or production approval. Case range (20 to 500) and decision labels corrected Oct 23, 2026 after a note from the maker, rechecked on the run page and the support-routing guide.
Prices are from the official pricing page on Oct 23, 2026 and can change; check the pricing page before you buy.
Real test
PassedOne task from the recorded test on this date. A failing task is shown when the record has one.
- Client
- details not recorded for this run Plan we tested: Free plan.
- Date
- Oct 27, 2026
- The task we gave
Run the built-in 20-case screen against the built-in demo agent
- What it promised
- details not recorded for this run
- What happened
Tool calls: details not recorded for this run
Note recorded with this task
Froze the rules and ran all 20 cases 3 times each: 60 attempts in about 2 seconds, no account. The decision was STOP because 2 cases produced a critical failure (an escalation miss and a forbidden claim), shown apart from the average: 16 of 20 cases fully passed and 17 of 20 were stable. The demo agent has planted mistakes, and the gate caught them as its rules say.Output excerpt: details not recorded for this run
- Verdict
- Passed
Our test
It does what it says with no account: freeze 20 test cases, run each 3 times, and get a clear STOP or go decision where a critical miss can't hide behind a good average. Bad case files are stopped before the run, and the exports keep a matching fingerprint. The catch: it's an early prototype on a chatgpt.site address, and we only tested the built-in demo agent, not a real endpoint.
Tested on free plan · Oct 27, 2026 · verdict Pass (3/3 tasks passed)
What we tested
- Run the built-in 20-case screen against the built-in demo agent PassFroze the rules and ran all 20 cases 3 times each: 60 attempts in about 2 seconds, no account. The decision was STOP because 2 cases produced a critical failure (an escalation miss and a forbidden claim), shown apart from the average: 16 of 20 cases fully passed and 17 of 20 were stable. The demo agent has planted mistakes, and the gate caught them as its rules say.
- Upload our own test cases PassA 5-case file was blocked before any run with clear reasons (it needs 20 to 500 cases and a minimum case mix), and the Configure step stayed locked. Our own 20-case CSV loaded, froze and ran 60 attempts; the rule-based demo agent sent our unfamiliar requests to human review, so the decision was STOP (0 of 20 passed: 17 wrong routes, 3 needless escalations), scored exactly by the frozen rules.
- Export the evidence PassThe evidence ID and the manifest fingerprint stayed the same from freeze to after the run. The internal bundle (JSON) held the manifest, all 20 cases, all 60 attempts and scores matching the screen, plus a bundle fingerprint. The shareable summary kept the counts but left out the case text, as it promises.
What the free plan limited
Fully free in the browser with no account. We used the built-in demo agent only; calling your own staging endpoint needs a public HTTPS endpoint that allows browser (CORS) calls, which we did not try. The $149 setup service was not tried.