Glossary
What is AI truth?
Updated September 14, 2026
AI truth is whether an answer from a model or an agent can be traced to a record that was true when it was used, and whether the agent behaved while producing it. It has two parts. The record part is data truth: the record was sourced, consented, current and scored before the model saw it, which SuperTruth measures with the Data Trust Index (DTI), 0 to 100 on any record. The agent part is behavior: VIGIL, run by SuperTruth, scores every agent decision on a Behavioral Integrity Index from 0.0 to 1.0, lands it in one of four gates (pass, hold, alert, collapse) before it propagates, and seals it in a hash-chained ledger nobody can quietly edit. SuperTruth's line "Data truth is AI truth" means a model or agent is only as true as the records it acts on.
AI truth has two failure modes, and they need two different instruments. The first is an answer that looks true and has no row behind it: a figure produced from memory, with no source and no year. The second is an agent that reports well on itself. The record test catches the first. The agent test catches the second.
AI truth is not AI detection. Tools that guess whether a text was written by a machine answer a different question. AI truth asks whether what the model or agent did can be traced and whether the agent behaved, and it is proved with a score and a receipt, not a guess about authorship.
The two failure modes
The first failure is an answer with nothing behind it. On the homepage Gauntlet, the same six county-health questions go to four systems; SuperTruth answers live from DataSpine with a source and a year on every figure, and the ChatGPT, Gemini and Grok answers are recordings, captured 26 Aug and 1 Sep 2026, that answer from memory with no figure traceable to a row. Some of the recorded numbers are close. None can be checked.
The second failure is an agent that grades itself. Agents in production talk to each other, some can change their own configuration, and all of them report on themselves. An agent's self-report is a fine input and a poor authority. An agent cannot mark itself authorized, and a system that lets it has no way to tell a good agent from one that has learned to look like one.
The record test
Traceable to a row means a figure can be followed back to a source that published it, for the place and the period it claims, with the vintage kept on the value. The Truth Agent on this site is bound by a written rule: "If no DataSpine block is present, you must not state any specific numeric statistic for a geography. Not from memory, not an estimate." For a record rather than a place, the test is the Data Trust Index: 0 to 100 across eight dimensions, provenance the heaviest at 25 percent of the default weight (DTI white paper, DOI 10.5281/zenodo.19601616, 2026). The full definition is Data truth; the engine is the DTI Engine.
The agent test
VIGIL, run by SuperTruth, sits beside a fleet of agents, not inside it. Every event is scored on the Behavioral Integrity Index (BII), one number from 0.0 to 1.0 recomputed on every event over windows of the agent's recent behavior. Four weighted signal families feed it today: score inflation, config tamper, alignment faking and asset movement. A fifth, declared-behavior drift, is computed and written to every record at weight zero until a calibration study supports a weight. The weights and thresholds are a versioned policy per organization, not constants in code, and the digest of the policy in force rides on every score, so this page does not publish one set of numbers for them.
The API answers every event with one of four gates before anything propagates. Pass goes through. Hold, alert and collapse wait for a named human; collapse sets composite trust to zero. The service that scores an agent cannot release a hold on it: release, reject and escalate live in a second service with its own key and its own database role. Every scored event and every review decision appends one record to a hash-chained ledger that rejects UPDATE and DELETE for every role, SuperTruth's included. Daily anchoring of the chain head to an external timestamp is planned and not yet built; until then, the export you take today is your anchor.
The proof you can check
Two things on this site can be checked by a reader with no account. The Gauntlet: six county-health questions, four systems, SuperTruth live and the other three recorded 26 Aug and 1 Sep 2026 through each vendor's own API, with a "How we ran this" note and the non-affiliation line under the table. Row seven is the reader's own county. The ledger verifier: three commands and a free pilot key export a VIGIL ledger as a bundle whose verifier is forty lines of Python with no dependency on VIGIL; it recomputes the chain from the first record and prints OK with the record count and head hash, or BROKEN at the first bad record.
AI truth is not AI detection
Search for the phrase and the results are tools that guess whether a paragraph was written by a machine. That is a question about authorship. AI truth is a question about evidence and conduct: can the answer be traced to a record that was true when it was used, and did the agent behave while producing it. A human can write an untraceable figure; an agent can produce a traceable one. SuperTruth does not detect machine writing and makes no claim about it.
Where it runs on this site
The Truth Agent in the corner of this site is registered with VIGIL; its public record showed a BII of 0.9375, gate pass and 29 ledger records when read on 14 September 2026. Vera, who answers (215) 918-4140, runs under the same gate: every reply is an event, every event is scored, and a hold or worse pauses the agent for a human.
Where this runs