Not another AI wrapper.
Every Pigfox tool is a purpose-built verification pipeline. Where AI is involved, it's the analyst — never the source of the facts.
Evidence before inference.
Every check hits real infrastructure first: live DNS, WHOIS, TLS chains, raw EXIF binary structures, reverse-image search across the open web. The facts exist before any model sees them.
The contract demos enforce that ordering rather than promising it. In the parametric-insurance and deposit-dispute contracts on Base Sepolia, the hash of the evidence is committed on chain before any verdict exists, and each model’s identifier, prompt hash and evidence hash are published with its answer. The sequence is checkable by a stranger, which is a different claim from being asserted here.
AI interprets, it doesn't invent.
Models only reason over evidence the pipeline already collected. Results cite the raw signals, so you can verify every claim yourself.
Most Pigfox tools don't use AI at all — deterministic checks need no model. Of the 28 tools on this site, only a handful call one.
Purpose-built pipelines, not prompt templates.
Each tool is a dedicated Go pipeline — queued jobs, retries, timeouts, rate limits — engineered per problem, not a text box in front of one API.
The same machinery runs the 32 demos and publishes a researched article every day through a job queue with nobody at the keyboard. It is one Go binary, and you are reading its output.
Deterministic where it matters.
Header parsing, record lookups, certificate validation, and metadata extraction are computed, never generated. No invented WHOIS records, ever.
Privacy engineered in.
No accounts warehousing your data: magic-link sign-in, photo storage that deletes itself, and analysis that runs on what you submit — nothing more.
Concretely: an uploaded photo is held in memory for at most two minutes, and is deleted explicitly the moment the reverse-image lookup returns rather than waiting for that limit. Sign-in is a magic link, so there is no password here to lose.
What didn't work.
The labs on this site publish their misses, because a lab that only ever reports success is not reporting. These are the results as they came out:
- bench-lab registered six predictions before the run and had three refuted. The run also invalidated one of its own metrics — the tokens-per-second figure credits work done before the first token to the interval after it, which flatters reasoning models enormously. That erratum is published beside the numbers rather than quietly fixed.
- judge-lab discarded 815 of one judge’s 1,032 replies for formatting, 761 of them because the model wrapped valid JSON in a Markdown fence. The parsing contract was not amended afterwards to rescue the run. Five of its six predictions were refuted.
- sba-lab found that the federal 7(a) loan files spell paid-in-full with embedded spaces, across 1,179,360 rows. A decoder that compares against the documented spelling reports a default rate near 100% as its headline, and nothing about that output looks wrong.
- gas-optimization deploys the same contract twice, and the optimized one costs about 10% more to deploy than the naive one. Cheaper to run, dearer to ship.
- reconcile-lab had a headline claim cut during production verification: the effect is real in the simulation and could not be reproduced on the live page at any setting a visitor can reach, across seven runs.
- agent-gateway scored 150 of 150 on its blind run, then 38 of 38 on a harder set built to break it, and both are published as failed experiments — the model being graded had also written the engine grading it.
See the pipelines yourself → browse the tools, or drive the demos.