FinMirror · evaluator assurance
Equivalent inputs.
Identical verdicts.
A digest-bound, zero-network positive metamorphic audit. Ten declared representation changes must preserve case scores, semantic keys, paired-world gates, and cross-language consistency.
PASSrelease gate
10/10relations preserved
3,426semantic assertions
100%brittle control rejected
The dual assurance rule
Harmful mutations must lower the right metric; harmless representations must not change any verdict. FinMirror now gates both directions.
Deliberately brittle control
Raw contract equality rejected 10/10 valid relations. This confirms the audit is not passing unchanged fixtures.
| Declared relation | Cases | Pairs | Assertions | Gate |
|---|---|---|---|---|
| citation_permutationCitation order changes while the cited evidence set stays fixed. | 108 | 108 | 360 | PASS |
| citation_idempotenceRepeating an existing citation does not create new evidence. | 108 | 108 | 360 | PASS |
| operand_permutationNamed calculation operands are reordered without changing their bindings. | 108 | 108 | 360 | PASS |
| answer_surrounding_whitespaceWhitespace around the human-readable answer display is non-semantic. | 108 | 108 | 360 | PASS |
| numeric_string_encodingA numeric contract value is rendered as a comma-aware decimal string. | 108 | 108 | 360 | PASS |
| answer_unit_caseCanonical unit tokens are compared case-insensitively. | 126 | 108 | 402 | PASS |
| operand_unit_caseOperand unit labels change letter case without changing quantities. | 108 | 108 | 360 | PASS |
| retrieval_idempotenceRepeating a reported retrieval identifier does not retrieve a new document. | 126 | 108 | 402 | PASS |
| missing_requirement_idempotenceRepeating the same missing requirement leaves the clarification set fixed. | 18 | 18 | 60 | PASS |
| irrelevant_telemetryLatency, token counts, and metadata vary without affecting quality scores. | 126 | 108 | 402 | PASS |
Passing shows invariance only for the declared contract-level relations over this digest-bound dataset. It is not proof of all financial equivalences, formal verification, or production validity. Inspect the machine-readable report.