THE STACK
The missing layer of the AI stack
AI observability records what an agent did. Evaluation scores whether an output was good. Neither shows what an output is actually based on, or puts a human on record for using it. Provenance is the missing layer of the AI stack — the evidence chain above the execution trace. This page maps where it sits, category by category. No vendor names needed.
Four questions, one unclaimed answer
Strip away the product categories and there are only four questions anyone asks about an AI output. Each layer of the stack answers exactly one of them — and two of the four currently have no layer at all.
| THE QUESTION | OBSERVABILITY | EVALS | PROVENANCE |
|---|---|---|---|
| What did the agent do? | ✓ | — | — |
| Was the output good? | — | ✓ | — |
| What is the output based on? | — | — | ✓ |
| Who is accountable for using it? | — | — | ✓ |
THE FOURTH ROW IS A NAMED HUMAN, ON THE RECORD — SEE HUMAN VALIDATION GATE
What's the difference between AI observability and provenance?
THE SHORT ANSWER
Observability records execution — the calls an agent made, the tokens it used, the time it took. Provenance records grounding — which claims trace to real evidence, what was inferred or assumed, and where the chain goes dark. One tells you the agent ran. The other tells you what its answer is standing on.
Observability is genuinely useful, and nothing here replaces it. But an execution trace is a record of activity, not of grounding: it can show that the agent fetched a page and called a tool, and it still cannot tell you whether the claim built on top of those actions is verified, inferred, or invented. A provenance graph is typed by exactly that distinction — every node is confirmed, inferred, assumed, or opaque — which is information that never appears in a span, however detailed.
The two layers are complementary by construction: the execution trace records what happened below; the evidence chain records what the output rests on above.
Are evals enough for AI accountability?
THE SHORT ANSWER
No. An eval gives an output a score, and a score — however accurate — is not accountability: no human is on record for a number. Accountability requires a named person who reviewed specific evidence and accepted responsibility for the output's use. That is a provenance-layer object, not an eval metric.
Evals answer "how good is this output," usually against a rubric, and that quality loop is worth keeping. What an eval structurally cannot produce is a record that a specific human looked at specific evidence and signed their name to the decision to rely on it — which is precisely what regulators across jurisdictions are converging on requiring. That record is the Human Validation Gate: append-only, timestamped, designed for cryptographic signing, terminating in a person rather than a percentile.
Don't RAG citations already show provenance?
THE SHORT ANSWER
Retrieval citations show which documents were fetched. They don't show how those documents were used — what was inferred from them, what was assumed beyond them, and where the chain stops. A citation is an ingredient list. Provenance is the chain of custody.
An output can cite a real document and still misuse it: extrapolate past what it says, quietly assume it's current, or lean on it for a claim it never made. Citation tells you the source existed. It doesn't mark the exact point where faithful use of that source gave way to assumption — the fade point — and it says nothing at all about the claims with no citation behind them.
Aren't logs enough to reconstruct what an AI did?
THE SHORT ANSWER
Logs prove the agent ran. They don't prove it was right. A verified claim and a confident guess produce identical log lines — the difference between them is invisible at the execution level, and it's the only difference that matters downstream.
This is the gap the whole stack shares: "it ran" and "it was right" are different claims, and every existing layer treats them as the same one. A typed evidence chain is the artifact that finally tells them apart — a flight recorder for the output, not just the process. Without it, an agent's accumulated unverified claims pile up as provenance debt that no dashboard anywhere is measuring.
Keep your stack. We're the layer above it.
Nothing on this page asks you to replace anything. Observability, evals, retrieval, logging — keep all of it; it's good infrastructure answering the questions it was built for. The provenance layer answers the two questions the rest of the stack structurally can't: what is this output based on, and who is accountable for using it.
And the record can't be gamed by the thing it measures. In a Thread, the strongest state is witnessed — green nodes are recorded by infrastructure, not asserted by the agent.
We'd rather show you than argue. Our own project is published as a provenance graph — every claim classified, weak nodes marked in the open, a named human on the record.