A provenance record for what AI says is true: what it said, when it changed, and who moved it.

A buyer asks an engine who is best, a patient asks about a clinic, a voter asks about a candidate. The answer leaves no copy, and it moves. In the simulator's control arm, when we repeated a prompt with nothing changed, a business's coverage varied by 7.37 percentage points (SD, provisional). Whether production answers are that noisy, and how much more they move when a model is replaced, is the first thing the record has to show.

The record keeps those answers. CTRL's product tracks eleven engines; the signed record measures four. For each category, a fixed, hashed prompt set goes to those four, ChatGPT, Claude, Gemini and Grok, twice per prompt. Every answer is stored verbatim with the businesses it names, any sources it cites (3% of answers are search-grounded) and, where the API reports it, the model id (about two-thirds of answers). After the first 74, each is signed with an ed25519 key and hash-chained at collection. Nothing is rewritten.

So far the record is broad rather than deep. Most industry-metro cells have been measured once; a few questions have been asked daily. In the sweeps, each engine has answered under a single model (gpt-5, claude-sonnet-5, gemini-2.5-flash, grok-4.5), so the record has not yet seen a release. It is built to catch the first one.

In 2025, before CTRL existed, I created two frameworks to make what generative engines say about an entity measurable, and coined the term Generative Information Engine for the systems being measured. The papers are co-authored.

The AI Visibility Index scores how prominently an entity appears in an engine's answers across structured prompt sets: whether it is named, where, how often, and in what role. The AI Reputation Score scores what is said once the entity appears: sentiment, the authority of the sources the engine leans on, and the consistency of the claims across prompts and versions.

The record says an answer changed. It cannot say why. Engines are non-deterministic and a release can move everything at once, so the causal work happens in instruments built around the record. The simulator places a piece of candidate content in an engine's context and compares the answer against a control arm with nothing injected, on the same prompts and engines (model ids logged per run). It states its own limits on the page: it measures how an engine responds to content it already has, not whether the engine would ever retrieve that content from the open web.

In 2026 two political action committees asked CTRL Reality to change what AI systems say about their candidates. We had the methods to try. We said no, and the refusal became the rule every client is held to: amplify what is true, never fabricate.

Then I went looking for the law, the standard or the record that would stop the next company from saying yes, and found audits, but no standing record. The proposal that followed sets out what a provenance standard for AI answers would have to contain: a version-stamped public record, a change log, source attribution at the moment of the answer, and a contestation path through which a person or institution can see what changed about them and dispute it.

The test I hold myself to: by the time I am done, a person can look up what an AI system said about an entity, see when it changed, and find out who moved it.

For a year I was a research assistant at Harvard Business School, applying NLP to map how leaders win and lose public trust: which words and behaviors signal competence or warmth, the two judgments people make of anyone first (Fiske, Cuddy and Glick, 2007), and what those perceptions led to. Trust signals can be performed, and so can what a system says about someone; both need a record of who shaped them.