# Evidence model

## Purpose

The evidence model describes how public report findings stay connected to their sources. It does not disclose the proprietary numeric method.

## Evidence unit

A public evidence unit can contain:

- The role and interview identifiers.
- The approved question identifier.
- The configured criterion identifier.
- The transcript source location.
- The attributable speaker.
- A short evidence excerpt or source reference.
- A support state.
- An answer opportunity state.
- An ownership state when relevant.
- A reviewer note and review status.

## Support states

The public vocabulary uses these descriptive states:

- **Supported:** the supplied interview contains relevant evidence for the criterion.
- **Partially supported:** the evidence addresses part of the criterion but leaves material elements unresolved.
- **Contradicted:** the candidate explicitly provides evidence that conflicts with the criterion.
- **Not observed:** the supplied interview does not contain enough relevant evidence for the criterion.
- **Not evaluable:** the available material cannot support a responsible determination, for example because the question was not asked, the transcript is unusable, or attribution is unclear.

These states describe evidence coverage. They do not reveal a person's total capability and should not be treated as permanent labels.

## Answer opportunity states

- **Valid opportunity:** the candidate received a clear chance to address the relevant criterion.
- **Limited opportunity:** the candidate received only a partial or interrupted chance.
- **No opportunity:** the interview did not create a usable chance to address the criterion.
- **Unknown opportunity:** the supplied record is not sufficient to determine the opportunity.

Opportunity is important because absence of evidence can result from interview design, not candidate capability.

## Ownership states

- **Personal ownership stated:** the candidate clearly identifies work they personally performed or decided.
- **Shared ownership stated:** the candidate describes team work and distinguishes their contribution.
- **Team outcome only:** a team result is described without enough evidence to assign personal ownership.
- **Ownership unclear:** the transcript does not establish who performed the work.
- **Not applicable:** ownership is not relevant to the criterion or finding.

Ownership states prevent a team outcome from being silently converted into personal evidence.

## Contradiction handling

A contradiction requires explicit conflicting evidence. Silence, brevity, missing detail, or an unasked question is not enough. When two transcript segments conflict, the report should preserve both references and send the conflict to human review.

## Source traceability

Every material finding should be traceable to the configured criterion and transcript source. A reviewer should be able to answer:

1. Which role criterion was evaluated?
2. Which question created the evidence opportunity?
3. Which candidate words support the finding?
4. What evidence is still missing?
5. Was ownership established?
6. What must a human verify?

## What the evidence model does not include publicly

The public model does not expose formulas, weights, thresholds, score anchors, calibration logic, private prompts, internal harness instructions, hidden reasoning, candidate records, customer records, or production implementation details.

## Related resources

- [Evaluation workflow](https://axiomcx.dev/knowledge/evaluation-workflow.md)
- [Public report contract](https://axiomcx.dev/knowledge/report-contract.md)
- [Evidence model JSON](https://axiomcx.dev/data/evidence-model.json)
