# Proof Rack

> A claim-evidence audit for crowdfunding campaigns and product launch pages, drawn as a rack of test tubes. Paste the page; every sentence that claims something becomes a tube. The audit fills each tube to the rung of evidence the claim really reaches, colours it by the flaw that weakens it most, and writes the questions to send the creator before any money moves.

Live: https://proof-rack.skillsafe.ai/
API: https://proof-rack.skillsafe.ai/api.html

## Why it exists

In late October 2026 a famous blood-testing startup story drew renewed attention: the Wikipedia article on its founder rose from roughly 35,000 views a day in early October to about 253,000 on 20 October (Wikimedia pageviews). Its lesson applies to any launch page - a result is only as good as the evidence behind it - and the holiday season is when most people back a campaign or preorder a gadget.

## What runs free in the browser

- The claim finder: the page is split into sentences; a sentence is a claim when it carries a figure, a comparison, a superlative, a proof word, a demo, a render or a result verb. Up to 12 claims, each marked with what the page noticed: a figure, "up to", a relative figure, an unnamed baseline ("than ordinary..."), outside proof (independent lab, delivered units), in-house tests, a demo, a render or projection, a test size, hype words, health words.
- The page's own quick reading of each claim's rung and flaw, and the verdict those readings lean towards.
- The rack: one test tube per claim; liquid height is the rung, colour is the flaw.
- The loop: paste the creator's answer or the updated page and audit again; the page lines the two audits up by claim text and lists which claims rose or fell a rung, which changed flaw, and whether the verdict moved.
- Exports: questions for the creator, Markdown, claims CSV, JSON, print.

## The metered lane: audit

Input fields (all strings): task ("audit"), product, asking, claims (JSON array of {id, text, marks}), optional clip_note, retry_note.

Rungs, lowest to highest: asserted (only said; renders and projections count here), shown (a demo or working prototype, no measurement), measured (the makers' own tests), checked (an outside party named: independent lab, certification, published study, units delivered).

Flaws, checked in order: phantom (rests on a render, projection or unbuilt product, or a health, accuracy or breakthrough claim that is only asserted), strawman (an unnamed baseline), cherry-pick (best case as typical: "up to", ideal conditions, relative change without the absolute), thin-test (fewer than 30 people or units, under a week, or lab-only for a daily-life claim), none.

Verdicts, checked in order:
1. needs-claims - fewer than two claims; the reply asks two or three questions instead.
2. walk-away - two or more phantom claims.
3. back-it (shown as "Holds up") - every claim flaw-free and at least one checked.
4. ask-first - everything else.

The page re-checks every reply: one audit per claim, the verdict recomputed from the reply's own rungs and flaws, top-rung claims that name no outside checker, renders called "shown", numbers the claims never state, and accusing words. The audit judges evidence, never intent, and is not advice to buy.

## Sources

Built on the SkillSafe skill @wanshuiyin/experiment-audit (https://skillsafe.ai/skill/@wanshuiyin/experiment-audit): auditing results before they are claimed - phantom results, fake ground truth, normalisation tricks and insufficient scope. Runs on gpt-terra.