Decision evidence: why the agent chose this
When an agent acts, Provantix already records what it did. Decision evidence adds, when the agent provides it, why it chose that action: the hypothesis, the evidence it cited, the actions it considered, the one it selected, and what it rejected and why.
It is kept as evidence, not as truth. A decision is what the agent said it considered. It does not show the reasoning was sound, that it was the real reason, or that the action happened as chosen.
In the product
Product capture, 15–16 November 2026. Open the image for full size. Counts and dates are examples from the captured workspace.
Read Why the agent chose this on the run page. This example explicitly says no decision records were supplied. It illustrates the missing-evidence state, not a populated reasoning trace or proof that the action was correct.
What a reviewer sees
On the run page, a card called Why the agent chose this lists each decision:
- the hypothesis and the evidence it cites (references, not raw data);
- each candidate, marked Chosen, Rejected (with the reason) or Considered;
- who recorded it: the agent, a person or a system;
- the instruction it answers (a ticket, a request id, a hash) and the model behind it (model, version, invocation), when the agent gave them. When it didn't, the card says so: the instruction it answers is not recorded, the model behind it is not recorded;
- whether the action it selected is recorded, and whether the two records agree.
Next to each action in the run, a badge says decision linked, decision differs (the decision selects a different action or candidate than the one that cites it), decision id claimed twice, or decision not recorded here.
Three layers stay apart, and the page never merges them:
| Layer | What it is | Where you see it |
|---|---|---|
| Rationale | What the agent said it considered and chose. Asserted by the agent. | The decision card. |
| Activity | The action itself, as recorded and sealed. | The action in the run. |
| Destination result | What the destination system's own record shows, from destination confirmation. | The "Said versus landed" card. |
An action with no decision says so: No decision records in this run: why the agent chose these actions was not recorded. Nothing is written in its place.
How an agent records it
The SDK emits a decision record before the action, and the action's record carries a pointer back. See the SDK README ("Decision evidence"). In short:
buildDecisionAAR({ hypothesis, candidates, selected, rejected, ... })builds the decision record.- The action's record carries
ext: { decision_ref: decisionRef(decision.decision) }.
Neither record is changed after it is sealed. The link is two records that cite each other, so anyone holding both can check that they agree.
What is kept, and what is refused
Kept: short summaries and references (a metric id, a ticket, a hash, a model name and invocation id), capped in length. The instruction itself and the prompt are never kept, only a reference to them. Refused by the SDK: anything that looks like a credential, and any field the contract does not name, so a raw prompt or a reasoning trace cannot ride along by accident. Leave out what the agent did not say, and it reads as not recorded.
Erasure works as for any sealed record. If a decision's extension is erased, the run no longer shows it, the action says its decision is not recorded here, and the export still verifies.
Checking it offline
An export carries the decision and action records as they were sealed, and the verifiers compute the links from them with no extra file. Two checks appear when an export has a decision record or a pointer:
- decision evidence: every decision agrees with the action that cites it
- decision evidence: decisions are well formed
The verifier also reports how many actions cite a decision that is not in the export, and how many decisions selected an action that nothing points back at. Neither is a failure on its own, and the second is not evidence the action did not happen. Removing a decision from an export is caught by the chain, the roots and the emitter proof. The exact rules are in the format specification in the verification kit.
What this is not
- It is not a record of the agent's reasoning. Provantix does not keep raw prompts or chain-of-thought, and refuses to accept them as decision evidence.
- It is not a judgement of the decision. It does not say the choice was right, or that the agent considered everything it could have.
- It is not a block. Provantix reports; whether an action should have been allowed is for the systems that enforce.
