Rulvar API reference / @rulvar/core / EvidenceContract
Interface: EvidenceContract
Defined in: packages/core/src/engine/ctx.ts:242
A declared evidence floor (RV303): preflight judges tool caps against it, and under enforce: 'refuse' the runtime refuses an ok settle below it (RV507); see AgentProfile.evidenceContract.
Properties
| Property | Type | Description | Defined in |
|---|---|---|---|
calibration? | { callsPerEntry: number; source?: string; } | A journal observed prior for the per-entry call estimate (RV3309): the figure toolCalibrationFromJournal folds from a prior run of the same profile (aggregate or a p90 over several), fractional on purpose. Preflight uses the HIGHER of the declared estimate and this prior when it computes the evidence call floor, never the lower, so a stale generous declaration still holds and an optimistic one stops hiding the observed reality: the 2026-08-12 comparison run observed 4.211 calls per entry where the default estimate says 3. When the prior raises the floor, preflight names it in an evidence-estimate-below-observed finding beside the usual floor arithmetic. source is echoed in that finding so a reader knows which journal spoke. | packages/core/src/engine/ctx.ts:263 |
calibration.callsPerEntry | number | - | packages/core/src/engine/ctx.ts:263 |
calibration.source? | string | - | packages/core/src/engine/ctx.ts:263 |
classify? | (file) => EvidenceCategory | Replaces the default path classifier; must return one of the four categories. | packages/core/src/engine/ctx.ts:300 |
distribution? | Partial<Record<EvidenceCategory, number>> | The required spread of entries over source categories (RV4908, the tenth comparison experiment): the task demanded citations across implementation, tests, documentation, and examples, the contract could say only minEntries, the specialists cited documentation alone, and the judge took the points. Each declared category must reach its count of successful record_evidence executions whose file classifies into it (classifyEvidenceFile by default: test files and test directories are tests, an examples directory is examples, a docs directory or a prose file is docs, everything else implementation; classify replaces it). The effective floor is the larger of minEntries and the categories' sum: the finalization window's deficit and widened reserve, the RV809 proactive grant, the surplus gate, the terminal verdict (evidence.byCategory, met only when every category is met), and the enforce: 'refuse' refusal all read the per category shortfall, and the window notice names it ("record 3 more evidence entries first (implementation: 2 more, tests: 1 more)"). Preflight sizes the evidence call floor to the effective floor. Absent, every surface keeps its bytes. Policy, never identity. | packages/core/src/engine/ctx.ts:298 |
enforce? | "warn" | "refuse" | What the floor does at the child's terminal settle (RV507). The default 'warn' keeps the historical behavior: the contract is a preflight signal only. 'refuse' turns an ok finish whose message window carries fewer successful record_evidence executions (result recorded: true; duplicates and verification errors never count) than minEntries into a typed error terminal (kind 'terminal') whose journaled error data carries the machine-readable evidenceFloor: { recordedEntries, minEntries }; the outcome is memoized, so a resume rolls the refusal forward instead of re-paying the invocation. Non-ok terminals are never re-judged. | packages/core/src/engine/ctx.ts:276 |
estCallsPerEntry? | number | Estimated executed calls per recorded entry; default 3. | packages/core/src/engine/ctx.ts:246 |
minEntries | number | Evidence entries the task must record; positive integer. | packages/core/src/engine/ctx.ts:244 |
overheadCalls? | number | Estimated non-evidence overhead calls; default 8. | packages/core/src/engine/ctx.ts:248 |