Rulvar API reference / @rulvar/evals / SweepCellReport
Interface: SweepCellReport
Defined in: packages/evals/src/sweeps.ts:96
Properties
| Property | Type | Description | Defined in |
|---|---|---|---|
caseNames | string[] | - | packages/evals/src/sweeps.ts:110 |
effort? | Effort | - | packages/evals/src/sweeps.ts:98 |
envelopeExhausted? | true | The aggregate envelope refused a TARGET run of this cell before it started; everything measured up to that point stays reported and the cell emits no claim. | packages/evals/src/sweeps.ts:140 |
exhaustedRuns? | number | Count of case results whose TARGET run settled 'exhausted' (its per-run ceiling, not the envelope). A budget-starved measurement must not become a model belief, so any exhausted target suppresses the cell's claim even when the degraded passRate crosses a threshold: the alternative is committing a false weakness that blames the model for the ceiling. | packages/evals/src/sweeps.ts:119 |
incompleteReason? | "judge-exhausted" | "judge-refused" | "envelope-exhausted" | Why the cell is incomplete, when it is. | packages/evals/src/sweeps.ts:142 |
judgeIncompleteRuns? | number | Count of result rows whose grading stopped on a judge budget event (per-run judge ceiling or envelope refusal of a judge run). The paid target evidence stays on those rows; the cell emits no claim. | packages/evals/src/sweeps.ts:134 |
model | `${string}:${string}` | - | packages/evals/src/sweeps.ts:97 |
n | number | Result rows actually measured (completed count). | packages/evals/src/sweeps.ts:102 |
nonOkRuns? | number | Count of case results whose TARGET run settled neither ok nor exhausted ('error', 'cancelled', 'suspended'): measurement artifacts, not model quality. Exactly like exhaustedRuns, any such run suppresses the cell's claim: a passRate deflated by a provider failure or a host cancellation must not become a committed model weakness (cycle 81). | packages/evals/src/sweeps.ts:128 |
passRate | number | - | packages/evals/src/sweeps.ts:100 |
plannedN | number | Cases this cell was asked to measure (v1.17.0 review P1-5). A cell with n < plannedN is incomplete: what ran stays reported, and the cell emits no claim. | packages/evals/src/sweeps.ts:108 |
refusedRunLabel? | string | The refused run's label, when the envelope refused one. | packages/evals/src/sweeps.ts:144 |
taskClass | TaskClass | - | packages/evals/src/sweeps.ts:99 |
totalCostUsd | number | - | packages/evals/src/sweeps.ts:109 |