Citation resolution rate
The share of tested identifiers that resolve to the expected source and provision.
Assurance scorecard
Measure the evidence chain, including its failures.
Ansvar will publish separate measures for citation resolution, claim support, unresolved findings, freshness and practitioner review. The scorecard will show the test corpus, method, exclusions and prior results beside each release.
- No composite accuracy claim
- No public benchmark value yet
- Versioned results planned
Measures
Publish the failure classes separately.
A single accuracy percentage would hide the difference between a broken citation, weak support and a correct refusal to conclude.
Claim-support rate
The share of cited factual claims where the cited text supports the statement made.
Unsupported-claim rate
The share of factual conclusions attempted without enough served evidence.
Correct unresolved rate
The share of known evidence-gap cases where the workflow leaves the conclusion unresolved.
Freshness and correction latency
The elapsed time from a source change or confirmed record defect to the served update.
Reviewer acceptance
The share of findings that pass practitioner review without a substantive correction.
Reproducibility
The stability of evidence selection and finding status under the same pinned inputs and sources.
Time to reviewed deliverable
The elapsed time from an accepted input set to practitioner approval of the assessment.
Publication status
Definitions are public. Results remain pending.
The table states what a buyer can inspect today. A pending artifact carries no implied score.
| Artifact | Status | Current evidence |
|---|---|---|
| Measure definitions | Draft published | The definitions on this page establish the intended evaluation units. |
| Versioned test corpus | Not published | No downloadable claim, citation and evidence-gap set is public yet. |
| Baseline results | Not published | Ansvar makes no numerical score claim on this page. |
| Historical series | Not published | No version-to-version public result series exists yet. |
| Independent rerun | Not commissioned | No third party has rerun the planned public test corpus. |
Release bar
A result needs enough context to be rerun.
Ansvar will publish a result only with the inputs, adjudication rules and version information needed to interpret it.
Pin the evidence
- Source snapshot and workflow version
- Input fixture and expected evidence state
- Model and client configuration used for the run
Publish the adjudication
- Claim and citation sampling rules
- Reviewer instructions and disagreement handling
- Known exclusions and failure categories
Keep the history
- Versioned results and corrected records
- Method changes beside the affected series
- Independent rerun material when commissioned
Ask for the evidence behind an assurance claim.
Until baseline results ship, use the methodology status and current product boundaries in your evaluation.