This is a small family of narrow, deterministic verifiers for AI-adjacent claims. Each one answers one specific question and gives you an artifact — a receipt, a diff, a proof — that anyone can recompute independently and get the exact same byte-for-byte answer.
No model is asked "does this look right?" None of these tools judge, grade, or vibe-check. They check a specific, falsifiable thing against a specific, disclosed source, and they show their work.
Is: a curated set of narrow verifiers. Each one checks a specific, well-defined claim against a source you provide, and produces a deterministic result plus a reproducible receipt or hash.
Isn't: a fact-checker, a truth oracle, or a general AI auditor. None of these tools evaluate whether a claim is true in the world — only whether it holds up against the specific evidence you gave it. They don't grade quality, safety, or intent. Scope stays deliberately small so the answer stays provable.
Verify a claimed quote is genuinely present, verbatim, in a source you paste. Tells a clean quote apart from an altered paraphrase or a fabrication outright, with a reproducible receipt.
Prove a redacted document is structurally faithful to a committed original. Catches an edited, forged, dropped, or reordered line — without ever seeing the hidden content.
Recompute a claimed "random" eval sample or train/test split from a disclosed population, seed, and size, then diff it byte-for-byte — catching train/test leakage a claim alone would hide.
Deterministic JSON Schema conformance for AI structured output. PASS/FAIL with exact violation paths, plus a canonical hash that tells reformatting apart from real content drift.
Check an AI component's observed state against a declared warrant — config, capabilities, files, invariants. One severe break is enough to block; nothing needs a human to spot it.
Commit a batch of content to one Merkle root you can publish anywhere, then hand out per-item inclusion proofs that anyone can re-check offline, with no shared ledger required.
Check an AI's claims about a public GitHub repo — files exist, a gate passed, phrases are present — against GitHub's own public API. Catches an agent claiming work it never did.
Scan text — a prompt, an output, a pasted doc — for the tricks that hide in characters: invisible/zero-width chars, homoglyph lookalikes, bidi overrides, and hidden Unicode-tag payloads (ASCII smuggling). Deterministic, byte-for-byte re-runnable report.
Find near-duplicate documents that exact-hash dedup misses — padded or lightly reworded copies. Word-shingling + MinHash gives a re-derivable IDENTICAL / NEAR-DUPLICATE / OVERLAPPING / DISTINCT verdict plus a recomputable receipt.
Prove two differently-formatted JSON documents are byte-identical after canonicalization (RFC 8785-aligned: sorted keys, normalized numbers, minimal escaping) — or pinpoint the exact path where they diverge. Reproducible in any language.
Recompute an AI agent's spend log — cap minus cumulative cost — to prove the balances are consistent, gap-free, and the budget cap was never exceeded, even if a later balance was edited to hide an overspend.
A source document, a population, a schema, a warrant, a repo — something disclosed and fixed, not a vibe.
A deterministic function, not a model call. Same inputs always produce the same output, every time.
A plain result plus the artifact behind it — a hash, a diff, a proof path — so the claim is checkable, not just stated.
A third party with the same inputs gets the same byte-for-byte answer. That's the whole proof — no trust required.