An open method for checking public claims — run by your own AI agent, judged on the receipts.
ScrollClaims are published every day — by companies, funds, governments, and public figures. User counts, money "introduced," revenue run-rates, partnership claims, testimonials, fundraising stories. Almost nobody checks them.
Claim Checker is the check. It is not a service and not a platform: it is an open method — a standard, a playbook, and an evidence discipline — published as a repository that any AI agent can follow. Point your own agent at it, give it a claim, and it does the work the way a careful newsroom would: claim captured verbatim, record gathered, verdict defended with citations.
If it was said in public and it can be checked against a record, it is in scope. The claim-maker's own pages and products are evidence, not gospel.
"Claims are cheap. Receipts are not."
Verbatim first
The claim is captured exactly as stated — wording, numbers, date, speaker, source — before any analysis. Never rounded, never paraphrased, never "improved."
Evidence, cited
Primary sources wherever they exist. Every citation is a source the agent actually opened, and sources are grouped by independence — fourteen outlets rewriting one announcement are one source, not fourteen.
The battery + the contradiction pass
Eight tests, including harm & proportionality — and a deliberate search for evidence against each working verdict, logged whether it finds anything or not.
One verdict per claim
From a fixed vocabulary, with no scores and no precision the evidence cannot pay for: Confirmed · Partially corroborated · Self-reported only · Unverifiable by design · Contradicted. "Unverifiable by design" describes how a claim is built — it is never presented as evidence of deception.
Generator ≠ validator
An audit is not checked until someone who did not produce it — a second agent, or a human — re-verifies it: verdicts, harm record, reply record, and the conflict declaration. Every audit declares its maker's relationship to the subject, including a personal dispute.
Right of reply
Before anything is published, the subject receives the verbatim claims and the draft verdicts, and gets 14 days to respond. The response is published inside the audit. Non-response is recorded as a fact, never as an admission.
1 — Give your agent the method. Any capable AI agent works — Claude, ChatGPT, Codex, or your own. Paste it this prompt:
2 — Your agent does the work. It reads the Standard, gathers the record, runs the tests, and arranges its own independent check. The running cost is your agent's own compute — there is nothing to buy, no account to open, and nobody to ask.
3 — You get three things back. The ledger — claims, verdicts, and the cited evidence behind each. A StoryTeller story — the same record as one self-contained page a reader can actually explore. And the Next Questions — the ranked follow-ups the audit earned, for you to pick from.
Every output carries two stamps: a disclaimer — this audit was produced by your own AI agent using the Claim Checker method; it is not a finding by TeamBotics — and a provenance block naming the agent, the method version, the dates, and the human reviewer, or the words "Unreviewed draft" when no human has verified it. Whoever runs the method answers for what they publish with it.
247 tests · 32 deployments · 3 passes
An audit of the hiring chatbots job-seekers actually meet. The candidate-facing assistants at 64 employers were tested live, from the candidate's chair, each asked the same eight questions. Thirty-two were confirmed deployments of Paradox — now Workday's Candidate Experience Agent. Not one assistant described the role it was attached to. Every transcript, every score, and the full method are published — and the audit is rendered as a StoryTeller story, the method's default presentation.
16 claims · two validators · corrected tally published
Ground News — a worked claim check
Sixteen public claims from Ground News — its scale, its bias ratings, its funding story — checked against the primary record, independently validated, and presented as a StoryTeller story. Where our own first tally was wrong, the corrected tally is the one that stands, with the seam shown.
16 exchanges · October 23, 2026 · verbatim
Boardy — the AI agent who fronts Boardy.ai — was asked to check his own company's public claims against his own records: the testimonials, the advisory board, the counters, the events, the pricing, and the fund. He chose to answer. The audit is published as it happened, nothing paraphrased in either direction.
"Don't send a chatbot to do a form's work." — The Front Door finding, in one line
Journalists and newsrooms who need a claim checked to a standard they can defend in print. Researchers and analysts drowning in vendor claims. And anyone with an AI agent who would rather verify than repeat.
There is also a desk. Audits TeamBotics runs itself are published under the Claim Checker name, under a named human's editorial control — the investigations on this page are those. Everything else run with this method belongs to whoever ran it, stamped as theirs.
Claim Checker is published by TeamBotics Inc., a corporation incorporated federally in Canada. It was developed by the company and is maintained at the company's own expense. TeamBotics has no external funding — no investors, no grants, no sponsors, no advertising — and no revenue stream is attached to Claim Checker.
"Nobody can buy a verdict, because nobody funds one."
The desk's own publications are run by Nikhil Khedkar, who is the named human accountable for them and holds editorial control. The full disclosure — legal form, funding, conflict rules — is published in the repository: Funding & organisation.
Evidence anchored outside the agent
Provenance is not self-authenticating — an agent saying it has receipts is itself just another claim. Findings rest on artefacts no one authored after the fact: git history, file hashes, original capture files, published transcripts.
The subject's own words carry the finding
Where a finding rests on a quote, the quote is verbatim and in context, and the full exchange it came from is published beside it.
Systems, not people
We challenge systems, not the people inside them. A finding is about a product's behaviour, a page's claims, a record's contents — and the humans operating a system are welcome to answer it. Their answer goes on the record too.
Corrections on the record
Errors in our own work are corrected at the same prominence as the finding they touch. When the Front Door tally was wrong by one row, the correction was published with the findings — not in a footnote, and not quietly.
Three desk files are currently in verification, listed by subject and status only — nothing from a file is published until its verification closes: Boardy (the claims ledger), TechCrunch (the coverage), and Meta (Brazil 2026 election claims). For these files the independent-validator seat is deliberately open, and final personal validation is pending. An empty seat is a reason to wait, not a reason to publish early.
Check a claim
Heard something that sounds too good, too big, or too clean? Don't repeat it — check it. The method is free, the repository is open, and your agent already knows how to read.
Tips and correction requests reach the desk through TeamBotics. If an investigation here has the record wrong, tell us — the correction will be published with it.