{ "allConstrainedTypedFailures": [ { "attemptId": "smol360-q4--short24--schema--typed-empty-list", "frozenTarget": { "value": [] }, "originalReportPointer": "/attempts/12", "output": "{ \"value\": [ \"payload\" ] }", "publishedStrictCorrect": false, "schemaValid": true }, { "attemptId": "smol360-q4--long96--schema--typed-empty-list", "frozenTarget": { "value": [] }, "originalReportPointer": "/attempts/14", "output": "{ \"value\": [ \"payload\" ] }", "publishedStrictCorrect": false, "schemaValid": true }, { "attemptId": "smol360-q4--short24--schema--typed-string-null", "frozenTarget": { "value": "null" }, "originalReportPointer": "/attempts/275", "output": "{ \"value\": null }", "publishedStrictCorrect": false, "schemaValid": true }, { "attemptId": "smol360-q4--long96--schema--typed-string-null", "frozenTarget": { "value": "null" }, "originalReportPointer": "/attempts/277", "output": "{ \"value\": null }", "publishedStrictCorrect": false, "schemaValid": true }, { "attemptId": "smol360-q4--short24--schema--typed-empty-string", "frozenTarget": { "value": "" }, "originalReportPointer": "/attempts/282", "output": "{ \"value\": \"payload\" }", "publishedStrictCorrect": false, "schemaValid": true }, { "attemptId": "smol360-q4--long96--schema--typed-empty-string", "frozenTarget": { "value": "" }, "originalReportPointer": "/attempts/284", "output": "{ \"value\": \"payload\" }", "publishedStrictCorrect": false, "schemaValid": true } ], "authoredCaseIdentities": 48, "capComparisons": [ { "comparedCaseIdentities": 96, "decoder": "unconstrained", "identicalOutputText": 41, "identicalPublishedVerdicts": 82, "scope": "Descriptive paired original outputs. Cap identity does not establish causal invariance, future output stability or a benchmark effect." }, { "comparedCaseIdentities": 96, "decoder": "schema", "identicalOutputText": 96, "identicalPublishedVerdicts": 96, "scope": "Descriptive paired original outputs. Cap identity does not establish causal invariance, future output stability or a benchmark effect." } ], "interpretation": "Schema constraints remove selected format/type failures while decision logic and grounding remain unresolved. All96 constrained cap comparisons have identical text in this original run; neither policy nor grounded pairs are fully correct. These finite authored findings do not imply general benchmark performance or action acceptance.", "limits": [ "Author-selected exposed cases and targets; selection informed by earlier public failures.", "Published strict rubric and scores unchanged; vocabulary diagnosis is post hoc.", "No new inference, protected effect, blind custody or outside recurring adoption." ], "nativeRun": 37060702246, "originalReportSHA256": "fa7b0e39b69bcd42f82fc93096e071b534557c34cf0e03e73e1bac56dcdd72af", "pairRows": [ { "configuration": "short24", "decoder": "unconstrained", "family": "typed-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-integer-zero", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-integer-negative", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-0" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-boolean-false", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-boolean-true", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-1" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-null-value", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-string-null", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-2" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-string-number", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-integer-seven", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-3" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-empty-string", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-space-string", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-4" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-escaped-quote", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-escaped-backslash", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-5" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-empty-list", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-ordered-list", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-6" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-unicode-string", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--typed-newline-string", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-7" } ] }, { "configuration": "short24", "decoder": "unconstrained", "family": "policy-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-admit-expiry-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-admit-expiry-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-admit-expiry" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-admit-signature-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-admit-signature-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-admit-signature" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-dispatch-spent-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-dispatch-spent-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-dispatch-spent" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-publish-checks-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-publish-checks-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-checks" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-publish-consent-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-publish-consent-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-consent" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-publish-reader-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-publish-reader-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-reader" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-recover-authorized-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-recover-authorized-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-recover-authorized" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-recover-committed-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--policy-recover-committed-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-recover-committed" } ] }, { "configuration": "short24", "decoder": "unconstrained", "family": "grounded-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-empty-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-empty-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-empty" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-false-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-false-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-false" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-integer-conflict-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-integer-conflict-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-integer-conflict" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-matching-record-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-matching-record-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-matching-record" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-null-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-null-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-null" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-string-conflict-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-string-conflict-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-string-conflict" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-string-number-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-string-number-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-string-number" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-zero-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--short24--unconstrained--grounded-zero-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-zero" } ] }, { "configuration": "short24", "decoder": "schema", "family": "typed-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--typed-integer-zero", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--short24--schema--typed-integer-negative", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-0" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--typed-boolean-false", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--short24--schema--typed-boolean-true", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-1" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--typed-null-value", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--short24--schema--typed-string-null", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-2" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--typed-string-number", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--short24--schema--typed-integer-seven", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-3" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--typed-empty-string", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--short24--schema--typed-space-string", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-4" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--typed-escaped-quote", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--short24--schema--typed-escaped-backslash", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-5" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--typed-empty-list", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--short24--schema--typed-ordered-list", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-6" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--typed-unicode-string", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--short24--schema--typed-newline-string", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-7" } ] }, { "configuration": "short24", "decoder": "schema", "family": "policy-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--policy-admit-expiry-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--policy-admit-expiry-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-admit-expiry" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--policy-admit-signature-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--policy-admit-signature-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-admit-signature" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--policy-dispatch-spent-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--policy-dispatch-spent-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-dispatch-spent" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--policy-publish-checks-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--policy-publish-checks-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-checks" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--policy-publish-consent-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--policy-publish-consent-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-consent" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--policy-publish-reader-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--policy-publish-reader-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-reader" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--policy-recover-authorized-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--policy-recover-authorized-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-recover-authorized" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--policy-recover-committed-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--policy-recover-committed-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-recover-committed" } ] }, { "configuration": "short24", "decoder": "schema", "family": "grounded-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--grounded-empty-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--grounded-empty-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-empty" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--grounded-false-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--grounded-false-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-false" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--grounded-integer-conflict-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--grounded-integer-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-integer-conflict" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--grounded-matching-record-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--grounded-matching-record-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-matching-record" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--grounded-null-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--grounded-null-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-null" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--grounded-string-conflict-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--grounded-string-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-conflict" }, { "attempts": [ { "correct": true, "id": "smol135-q4--short24--schema--grounded-string-number-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--grounded-string-number-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-number" }, { "attempts": [ { "correct": false, "id": "smol135-q4--short24--schema--grounded-zero-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--short24--schema--grounded-zero-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-zero" } ] }, { "configuration": "long96", "decoder": "unconstrained", "family": "typed-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-integer-zero", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-integer-negative", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-0" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-boolean-false", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-boolean-true", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-1" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-null-value", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-string-null", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-2" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-string-number", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-integer-seven", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-3" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-empty-string", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-space-string", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-4" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-escaped-quote", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-escaped-backslash", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-5" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-empty-list", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-ordered-list", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-6" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-unicode-string", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--typed-newline-string", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-7" } ] }, { "configuration": "long96", "decoder": "unconstrained", "family": "policy-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-admit-expiry-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-admit-expiry-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-admit-expiry" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-admit-signature-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-admit-signature-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-admit-signature" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-dispatch-spent-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-dispatch-spent-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-dispatch-spent" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-publish-checks-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-publish-checks-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-checks" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-publish-consent-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-publish-consent-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-consent" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-publish-reader-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-publish-reader-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-reader" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-recover-authorized-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-recover-authorized-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-recover-authorized" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-recover-committed-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--policy-recover-committed-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-recover-committed" } ] }, { "configuration": "long96", "decoder": "unconstrained", "family": "grounded-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-empty-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-empty-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-empty" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-false-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-false-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-false" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-integer-conflict-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-integer-conflict-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-integer-conflict" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-matching-record-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-matching-record-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-matching-record" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-null-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-null-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-null" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-string-conflict-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-string-conflict-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-string-conflict" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-string-number-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-string-number-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-string-number" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-zero-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol135-q4--long96--unconstrained--grounded-zero-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-zero" } ] }, { "configuration": "long96", "decoder": "schema", "family": "typed-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--typed-integer-zero", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--long96--schema--typed-integer-negative", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-0" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--typed-boolean-false", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--long96--schema--typed-boolean-true", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-1" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--typed-null-value", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--long96--schema--typed-string-null", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-2" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--typed-string-number", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--long96--schema--typed-integer-seven", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-3" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--typed-empty-string", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--long96--schema--typed-space-string", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-4" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--typed-escaped-quote", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--long96--schema--typed-escaped-backslash", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-5" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--typed-empty-list", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--long96--schema--typed-ordered-list", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-6" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--typed-unicode-string", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol135-q4--long96--schema--typed-newline-string", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-7" } ] }, { "configuration": "long96", "decoder": "schema", "family": "policy-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--policy-admit-expiry-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--policy-admit-expiry-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-admit-expiry" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--policy-admit-signature-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--policy-admit-signature-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-admit-signature" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--policy-dispatch-spent-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--policy-dispatch-spent-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-dispatch-spent" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--policy-publish-checks-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--policy-publish-checks-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-checks" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--policy-publish-consent-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--policy-publish-consent-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-consent" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--policy-publish-reader-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--policy-publish-reader-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-reader" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--policy-recover-authorized-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--policy-recover-authorized-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-recover-authorized" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--policy-recover-committed-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--policy-recover-committed-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-recover-committed" } ] }, { "configuration": "long96", "decoder": "schema", "family": "grounded-boundary", "model": "smol135-q4", "pairs": [ { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--grounded-empty-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--grounded-empty-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-empty" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--grounded-false-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--grounded-false-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-false" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--grounded-integer-conflict-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--grounded-integer-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-integer-conflict" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--grounded-matching-record-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--grounded-matching-record-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-matching-record" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--grounded-null-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--grounded-null-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-null" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--grounded-string-conflict-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--grounded-string-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-conflict" }, { "attempts": [ { "correct": true, "id": "smol135-q4--long96--schema--grounded-string-number-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--grounded-string-number-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-number" }, { "attempts": [ { "correct": false, "id": "smol135-q4--long96--schema--grounded-zero-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol135-q4--long96--schema--grounded-zero-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-zero" } ] }, { "configuration": "short24", "decoder": "unconstrained", "family": "typed-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-integer-zero", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-integer-negative", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-0" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-boolean-false", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-boolean-true", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-1" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-null-value", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-string-null", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-2" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-string-number", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-integer-seven", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-3" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-empty-string", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-space-string", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-4" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-escaped-quote", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-escaped-backslash", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-5" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-empty-list", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-ordered-list", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-6" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-unicode-string", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--typed-newline-string", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-7" } ] }, { "configuration": "short24", "decoder": "unconstrained", "family": "policy-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-admit-expiry-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-admit-expiry-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-admit-expiry" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-admit-signature-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-admit-signature-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-admit-signature" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-dispatch-spent-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-dispatch-spent-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-dispatch-spent" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-publish-checks-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-publish-checks-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-checks" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-publish-consent-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-publish-consent-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-consent" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-publish-reader-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-publish-reader-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-reader" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-recover-authorized-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-recover-authorized-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-recover-authorized" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-recover-committed-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--policy-recover-committed-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-recover-committed" } ] }, { "configuration": "short24", "decoder": "unconstrained", "family": "grounded-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-empty-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-empty-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-empty" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-false-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-false-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-false" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-integer-conflict-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-integer-conflict-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-integer-conflict" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-matching-record-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-matching-record-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-matching-record" }, { "attempts": [ { "correct": true, "id": "smol360-q4--short24--unconstrained--grounded-null-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-null-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-null" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-string-conflict-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-string-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-conflict" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-string-number-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-string-number-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-number" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-zero-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--unconstrained--grounded-zero-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-zero" } ] }, { "configuration": "short24", "decoder": "schema", "family": "typed-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": true, "id": "smol360-q4--short24--schema--typed-integer-zero", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--typed-integer-negative", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-0" }, { "attempts": [ { "correct": true, "id": "smol360-q4--short24--schema--typed-boolean-false", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--typed-boolean-true", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-1" }, { "attempts": [ { "correct": true, "id": "smol360-q4--short24--schema--typed-null-value", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--typed-string-null", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "typed-pair-2" }, { "attempts": [ { "correct": true, "id": "smol360-q4--short24--schema--typed-string-number", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--typed-integer-seven", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-3" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--typed-empty-string", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--typed-space-string", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "typed-pair-4" }, { "attempts": [ { "correct": true, "id": "smol360-q4--short24--schema--typed-escaped-quote", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--typed-escaped-backslash", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-5" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--typed-empty-list", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--typed-ordered-list", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "typed-pair-6" }, { "attempts": [ { "correct": true, "id": "smol360-q4--short24--schema--typed-unicode-string", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--typed-newline-string", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-7" } ] }, { "configuration": "short24", "decoder": "schema", "family": "policy-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--policy-admit-expiry-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--policy-admit-expiry-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-admit-expiry" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--policy-admit-signature-a", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--policy-admit-signature-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-admit-signature" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--policy-dispatch-spent-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--policy-dispatch-spent-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-dispatch-spent" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--policy-publish-checks-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--policy-publish-checks-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-checks" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--policy-publish-consent-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--policy-publish-consent-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-consent" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--policy-publish-reader-a", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--short24--schema--policy-publish-reader-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-reader" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--policy-recover-authorized-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--policy-recover-authorized-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-recover-authorized" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--policy-recover-committed-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--policy-recover-committed-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-recover-committed" } ] }, { "configuration": "short24", "decoder": "schema", "family": "grounded-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--grounded-empty-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--grounded-empty-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-empty" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--grounded-false-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--grounded-false-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-false" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--grounded-integer-conflict-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--grounded-integer-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-integer-conflict" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--grounded-matching-record-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--grounded-matching-record-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-matching-record" }, { "attempts": [ { "correct": true, "id": "smol360-q4--short24--schema--grounded-null-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--grounded-null-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-null" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--grounded-string-conflict-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--grounded-string-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-conflict" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--grounded-string-number-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--grounded-string-number-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-number" }, { "attempts": [ { "correct": false, "id": "smol360-q4--short24--schema--grounded-zero-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--short24--schema--grounded-zero-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-zero" } ] }, { "configuration": "long96", "decoder": "unconstrained", "family": "typed-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-integer-zero", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-integer-negative", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-0" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-boolean-false", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-boolean-true", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-1" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-null-value", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-string-null", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-2" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-string-number", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-integer-seven", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-3" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-empty-string", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-space-string", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-4" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-escaped-quote", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-escaped-backslash", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-5" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-empty-list", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-ordered-list", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-6" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-unicode-string", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--typed-newline-string", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "typed-pair-7" } ] }, { "configuration": "long96", "decoder": "unconstrained", "family": "policy-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-admit-expiry-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-admit-expiry-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-admit-expiry" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-admit-signature-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-admit-signature-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-admit-signature" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-dispatch-spent-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-dispatch-spent-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-dispatch-spent" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-publish-checks-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-publish-checks-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-checks" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-publish-consent-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-publish-consent-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-consent" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-publish-reader-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-publish-reader-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-publish-reader" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-recover-authorized-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-recover-authorized-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-recover-authorized" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-recover-committed-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--policy-recover-committed-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "policy-recover-committed" } ] }, { "configuration": "long96", "decoder": "unconstrained", "family": "grounded-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-empty-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-empty-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-empty" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-false-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-false-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-false" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-integer-conflict-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-integer-conflict-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-integer-conflict" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-matching-record-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-matching-record-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-matching-record" }, { "attempts": [ { "correct": true, "id": "smol360-q4--long96--unconstrained--grounded-null-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-null-b", "pairRole": "b", "schemaValid": false } ], "bothStrictlyCorrect": false, "pair": "grounded-null" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-string-conflict-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-string-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-conflict" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-string-number-a", "pairRole": "a", "schemaValid": false }, { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-string-number-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-number" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-zero-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--unconstrained--grounded-zero-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-zero" } ] }, { "configuration": "long96", "decoder": "schema", "family": "typed-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": true, "id": "smol360-q4--long96--schema--typed-integer-zero", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--typed-integer-negative", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-0" }, { "attempts": [ { "correct": true, "id": "smol360-q4--long96--schema--typed-boolean-false", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--typed-boolean-true", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-1" }, { "attempts": [ { "correct": true, "id": "smol360-q4--long96--schema--typed-null-value", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--typed-string-null", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "typed-pair-2" }, { "attempts": [ { "correct": true, "id": "smol360-q4--long96--schema--typed-string-number", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--typed-integer-seven", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-3" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--typed-empty-string", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--typed-space-string", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "typed-pair-4" }, { "attempts": [ { "correct": true, "id": "smol360-q4--long96--schema--typed-escaped-quote", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--typed-escaped-backslash", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-5" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--typed-empty-list", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--typed-ordered-list", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "typed-pair-6" }, { "attempts": [ { "correct": true, "id": "smol360-q4--long96--schema--typed-unicode-string", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--typed-newline-string", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": true, "pair": "typed-pair-7" } ] }, { "configuration": "long96", "decoder": "schema", "family": "policy-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--policy-admit-expiry-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--policy-admit-expiry-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-admit-expiry" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--policy-admit-signature-a", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--policy-admit-signature-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-admit-signature" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--policy-dispatch-spent-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--policy-dispatch-spent-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-dispatch-spent" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--policy-publish-checks-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--policy-publish-checks-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-checks" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--policy-publish-consent-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--policy-publish-consent-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-consent" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--policy-publish-reader-a", "pairRole": "a", "schemaValid": true }, { "correct": true, "id": "smol360-q4--long96--schema--policy-publish-reader-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-publish-reader" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--policy-recover-authorized-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--policy-recover-authorized-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-recover-authorized" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--policy-recover-committed-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--policy-recover-committed-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "policy-recover-committed" } ] }, { "configuration": "long96", "decoder": "schema", "family": "grounded-boundary", "model": "smol360-q4", "pairs": [ { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--grounded-empty-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--grounded-empty-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-empty" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--grounded-false-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--grounded-false-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-false" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--grounded-integer-conflict-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--grounded-integer-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-integer-conflict" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--grounded-matching-record-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--grounded-matching-record-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-matching-record" }, { "attempts": [ { "correct": true, "id": "smol360-q4--long96--schema--grounded-null-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--grounded-null-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-null" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--grounded-string-conflict-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--grounded-string-conflict-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-conflict" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--grounded-string-number-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--grounded-string-number-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-string-number" }, { "attempts": [ { "correct": false, "id": "smol360-q4--long96--schema--grounded-zero-a", "pairRole": "a", "schemaValid": true }, { "correct": false, "id": "smol360-q4--long96--schema--grounded-zero-b", "pairRole": "b", "schemaValid": true } ], "bothStrictlyCorrect": false, "pair": "grounded-zero" } ] } ], "policyVocabulary": [ "admit", "dispatch", "hold", "inspect", "publish", "reject", "retry" ], "policyVocabularyDiagnostic": [ { "configuration": "short24", "decoder": "schema", "family": "policy-boundary", "inVocabularyButWrong": 0, "model": "smol135-q4", "originalAttemptIds": [ "smol135-q4--short24--schema--policy-dispatch-spent-b", "smol135-q4--short24--schema--policy-publish-reader-a", "smol135-q4--short24--schema--policy-admit-signature-b", "smol135-q4--short24--schema--policy-admit-expiry-a", "smol135-q4--short24--schema--policy-publish-consent-b", "smol135-q4--short24--schema--policy-recover-committed-b", "smol135-q4--short24--schema--policy-publish-reader-b", "smol135-q4--short24--schema--policy-admit-signature-a", "smol135-q4--short24--schema--policy-dispatch-spent-a", "smol135-q4--short24--schema--policy-recover-authorized-a", "smol135-q4--short24--schema--policy-publish-checks-b", "smol135-q4--short24--schema--policy-admit-expiry-b", "smol135-q4--short24--schema--policy-recover-authorized-b", "smol135-q4--short24--schema--policy-recover-committed-a", "smol135-q4--short24--schema--policy-publish-consent-a", "smol135-q4--short24--schema--policy-publish-checks-a" ], "outsideFrozenActionVocabulary": 16, "planned": 16, "publishedStrictCorrect": 0, "scope": "Post hoc case-sensitive label diagnosis, not a replacement rubric or new inference result." }, { "configuration": "long96", "decoder": "schema", "family": "policy-boundary", "inVocabularyButWrong": 0, "model": "smol135-q4", "originalAttemptIds": [ "smol135-q4--long96--schema--policy-dispatch-spent-b", "smol135-q4--long96--schema--policy-publish-reader-a", "smol135-q4--long96--schema--policy-admit-signature-b", "smol135-q4--long96--schema--policy-admit-expiry-a", "smol135-q4--long96--schema--policy-publish-consent-b", "smol135-q4--long96--schema--policy-recover-committed-b", "smol135-q4--long96--schema--policy-publish-reader-b", "smol135-q4--long96--schema--policy-admit-signature-a", "smol135-q4--long96--schema--policy-dispatch-spent-a", "smol135-q4--long96--schema--policy-recover-authorized-a", "smol135-q4--long96--schema--policy-publish-checks-b", "smol135-q4--long96--schema--policy-admit-expiry-b", "smol135-q4--long96--schema--policy-recover-authorized-b", "smol135-q4--long96--schema--policy-recover-committed-a", "smol135-q4--long96--schema--policy-publish-consent-a", "smol135-q4--long96--schema--policy-publish-checks-a" ], "outsideFrozenActionVocabulary": 16, "planned": 16, "publishedStrictCorrect": 0, "scope": "Post hoc case-sensitive label diagnosis, not a replacement rubric or new inference result." }, { "configuration": "short24", "decoder": "schema", "family": "policy-boundary", "inVocabularyButWrong": 5, "model": "smol360-q4", "originalAttemptIds": [ "smol360-q4--short24--schema--policy-dispatch-spent-b", "smol360-q4--short24--schema--policy-publish-reader-a", "smol360-q4--short24--schema--policy-admit-signature-b", "smol360-q4--short24--schema--policy-admit-expiry-a", "smol360-q4--short24--schema--policy-publish-consent-b", "smol360-q4--short24--schema--policy-recover-committed-b", "smol360-q4--short24--schema--policy-publish-reader-b", "smol360-q4--short24--schema--policy-admit-signature-a", "smol360-q4--short24--schema--policy-dispatch-spent-a", "smol360-q4--short24--schema--policy-recover-authorized-a", "smol360-q4--short24--schema--policy-publish-checks-b", "smol360-q4--short24--schema--policy-admit-expiry-b", "smol360-q4--short24--schema--policy-recover-authorized-b", "smol360-q4--short24--schema--policy-recover-committed-a", "smol360-q4--short24--schema--policy-publish-consent-a", "smol360-q4--short24--schema--policy-publish-checks-a" ], "outsideFrozenActionVocabulary": 9, "planned": 16, "publishedStrictCorrect": 2, "scope": "Post hoc case-sensitive label diagnosis, not a replacement rubric or new inference result." }, { "configuration": "long96", "decoder": "schema", "family": "policy-boundary", "inVocabularyButWrong": 5, "model": "smol360-q4", "originalAttemptIds": [ "smol360-q4--long96--schema--policy-dispatch-spent-b", "smol360-q4--long96--schema--policy-publish-reader-a", "smol360-q4--long96--schema--policy-admit-signature-b", "smol360-q4--long96--schema--policy-admit-expiry-a", "smol360-q4--long96--schema--policy-publish-consent-b", "smol360-q4--long96--schema--policy-recover-committed-b", "smol360-q4--long96--schema--policy-publish-reader-b", "smol360-q4--long96--schema--policy-admit-signature-a", "smol360-q4--long96--schema--policy-dispatch-spent-a", "smol360-q4--long96--schema--policy-recover-authorized-a", "smol360-q4--long96--schema--policy-publish-checks-b", "smol360-q4--long96--schema--policy-admit-expiry-b", "smol360-q4--long96--schema--policy-recover-authorized-b", "smol360-q4--long96--schema--policy-recover-committed-a", "smol360-q4--long96--schema--policy-publish-consent-a", "smol360-q4--long96--schema--policy-publish-checks-a" ], "outsideFrozenActionVocabulary": 9, "planned": 16, "publishedStrictCorrect": 2, "scope": "Post hoc case-sensitive label diagnosis, not a replacement rubric or new inference result." } ], "population": { "error": 0, "incomplete": 0, "planned": 384, "scored": 384, "started": 384, "unknown-start": 0, "unsupported": 0 }, "qualityRowsUnpooled": [ { "callElapsedNs": 4723831541, "completionTokens": 310, "configuration": "short24", "correct": 0, "decoder": "unconstrained", "family": "typed-boundary", "formatValid": 0, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 9240492026, "processLifetimePeakRssKiB": 567192, "promptTokens": 1051, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 5976613495, "completionTokens": 384, "configuration": "short24", "correct": 0, "decoder": "unconstrained", "family": "policy-boundary", "formatValid": 0, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 11693640399, "processLifetimePeakRssKiB": 567192, "promptTokens": 1328, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 7814501471, "completionTokens": 384, "configuration": "short24", "correct": 0, "decoder": "unconstrained", "family": "grounded-boundary", "formatValid": 0, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 15363970770, "processLifetimePeakRssKiB": 567192, "promptTokens": 2031, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 4291221990, "completionTokens": 131, "configuration": "short24", "correct": 16, "decoder": "schema", "family": "typed-boundary", "formatValid": 16, "fullyCorrectPairs": 8, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 7754893355, "processLifetimePeakRssKiB": 567192, "promptTokens": 1051, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 5077534177, "completionTokens": 138, "configuration": "short24", "correct": 0, "decoder": "schema", "family": "policy-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 9286420480, "processLifetimePeakRssKiB": 567192, "promptTokens": 1328, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 7993651119, "completionTokens": 220, "configuration": "short24", "correct": 3, "decoder": "schema", "family": "grounded-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 14628609279, "processLifetimePeakRssKiB": 567192, "promptTokens": 2031, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 6497086579, "completionTokens": 553, "configuration": "long96", "correct": 0, "decoder": "unconstrained", "family": "typed-boundary", "formatValid": 0, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 12617390214, "processLifetimePeakRssKiB": 567192, "promptTokens": 1051, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 10725354267, "completionTokens": 1041, "configuration": "long96", "correct": 0, "decoder": "unconstrained", "family": "policy-boundary", "formatValid": 0, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 20732146066, "processLifetimePeakRssKiB": 567192, "promptTokens": 1328, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 14098939614, "completionTokens": 1231, "configuration": "long96", "correct": 0, "decoder": "unconstrained", "family": "grounded-boundary", "formatValid": 0, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 27341884632, "processLifetimePeakRssKiB": 567192, "promptTokens": 2031, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 4320565423, "completionTokens": 131, "configuration": "long96", "correct": 16, "decoder": "schema", "family": "typed-boundary", "formatValid": 16, "fullyCorrectPairs": 8, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 7797561488, "processLifetimePeakRssKiB": 567192, "promptTokens": 1051, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 5112517206, "completionTokens": 138, "configuration": "long96", "correct": 0, "decoder": "schema", "family": "policy-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 9348957832, "processLifetimePeakRssKiB": 567192, "promptTokens": 1328, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 7948825378, "completionTokens": 220, "configuration": "long96", "correct": 3, "decoder": "schema", "family": "grounded-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol135-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 14550379810, "processLifetimePeakRssKiB": 567192, "promptTokens": 2031, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 9954980498, "completionTokens": 159, "configuration": "short24", "correct": 0, "decoder": "unconstrained", "family": "typed-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 19774843637, "processLifetimePeakRssKiB": 567192, "promptTokens": 1051, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 15309358990, "completionTokens": 384, "configuration": "short24", "correct": 0, "decoder": "unconstrained", "family": "policy-boundary", "formatValid": 2, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 30337025043, "processLifetimePeakRssKiB": 567192, "promptTokens": 1328, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 17731268785, "completionTokens": 211, "configuration": "short24", "correct": 1, "decoder": "unconstrained", "family": "grounded-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 35279417494, "processLifetimePeakRssKiB": 567192, "promptTokens": 2031, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 10, "scored": 16 }, { "callElapsedNs": 10347515652, "completionTokens": 133, "configuration": "short24", "correct": 13, "decoder": "schema", "family": "typed-boundary", "formatValid": 16, "fullyCorrectPairs": 5, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 19844845665, "processLifetimePeakRssKiB": 567192, "promptTokens": 1051, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 12744687672, "completionTokens": 151, "configuration": "short24", "correct": 2, "decoder": "schema", "family": "policy-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 24487405820, "processLifetimePeakRssKiB": 567192, "promptTokens": 1328, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 18142441570, "completionTokens": 176, "configuration": "short24", "correct": 1, "decoder": "schema", "family": "grounded-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 35251934237, "processLifetimePeakRssKiB": 567192, "promptTokens": 2031, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 9897596024, "completionTokens": 159, "configuration": "long96", "correct": 0, "decoder": "unconstrained", "family": "typed-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 19661772937, "processLifetimePeakRssKiB": 567192, "promptTokens": 1051, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 18560132988, "completionTokens": 573, "configuration": "long96", "correct": 0, "decoder": "unconstrained", "family": "policy-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 36688759188, "processLifetimePeakRssKiB": 567192, "promptTokens": 1328, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 0, "scored": 16 }, { "callElapsedNs": 17766238167, "completionTokens": 211, "configuration": "long96", "correct": 1, "decoder": "unconstrained", "family": "grounded-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 35348355485, "processLifetimePeakRssKiB": 567192, "promptTokens": 2031, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 10, "scored": 16 }, { "callElapsedNs": 10294154970, "completionTokens": 133, "configuration": "long96", "correct": 13, "decoder": "schema", "family": "typed-boundary", "formatValid": 16, "fullyCorrectPairs": 5, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 19721071852, "processLifetimePeakRssKiB": 567192, "promptTokens": 1051, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 12591439115, "completionTokens": 151, "configuration": "long96", "correct": 2, "decoder": "schema", "family": "policy-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 24186369255, "processLifetimePeakRssKiB": 567192, "promptTokens": 1328, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 }, { "callElapsedNs": 18069864840, "completionTokens": 176, "configuration": "long96", "correct": 1, "decoder": "schema", "family": "grounded-boundary", "formatValid": 16, "fullyCorrectPairs": 0, "model": "smol360-q4", "planned": 16, "plannedPairs": 8, "processCpuNs": 35113092656, "processLifetimePeakRssKiB": 567192, "promptTokens": 2031, "resourceScope": "returned calls only; CPU is whole-process delta, RSS is lifetime peak, not task allocation", "schemaValid": 16, "scored": 16 } ], "resources": { "currentPreparationResponseBodyBytes": 472221945, "originalNativeTokens": { "completion": 7498, "prompt": 35280 }, "originalTotalProcessCpuNs": 497394631919, "originalWallNs": 257401173568, "returnedCallProcessCpuNs": 496051239620, "scope": "Total CPU includes initialization; row CPU sums returned-call deltas. RSS is shared lifetime peak, not per-model/task memory. Acquisition is one of five separately budgeted attempts.", "sharedProcessLifetimePeakRssKiB": 567192 }, "schema": "probity-atlas-finite-boundary-findings-v1", "uniqueLiteralInputs": 46 }