{
  "schemaVersion": "blobfish.benchmark-realism-standard.v5",
  "name": "Blobfish Enterprise Benchmark Realism Standard",
  "effectiveDate": "2026-08-30",
  "scope": "Every Blobfish-authored first-party task and every released companion task shown on /benchmarks.",
  "definition": "A realistic task is one a real employee answers by working a dependent chain of evidence: check the schedule for a slot, check whether the materials for that slot exist, read the specification for the required count, find what is missing, read the vendor's lead time, derive the realistic timeline, and weigh the alternatives (add a shift, buy more) before giving the options, the alternatives, and the exact answer from all the information. The deterministic verifier must recognise and grade that whole chain, not just the final value.",
  "principle": "The employee asks for an outcome. The agent must discover the necessary investigation, compare plausible options, make only supported changes, verify persisted state, and communicate the result. Public evidence must never contain the oracle plan or precomputed outcome.",
  "requirements": {
    "employeeRequest": {
      "minimumWords": 45,
      "maximumWords": 220,
      "exactDuplicatesAllowed": 0,
      "maximumPairwiseFiveShingleJaccard": 0.72,
      "proceduralRecipeAllowed": false,
      "toolNamesAllowed": false,
      "harnessSyntaxAllowed": false,
      "gradingLanguageAllowed": false,
      "mustDescribeBusinessOutcomeNotProcedure": true,
      "singleCoherentOperationalObjectiveRequired": true,
      "taskScopedExecutionAuthorityRequired": true
    },
    "workflow": {
      "distinctReferenceToolSequences": true,
      "distinctSemanticActionGraphs": true,
      "maximumPairwiseRawSequenceSimilarity": 0.95,
      "maximumPairwiseSemanticGraphSimilarity": 0.85,
      "structuralCloneGate": {
        "rawSequenceSimilarityThreshold": 0.9,
        "identifierNeutralSemanticSimilarityThreshold": 0.7,
        "comparison": "strictly-greater-than-both",
        "pairsAllowed": 0,
        "rationale": "Shared enterprise controls may legitimately reuse tools. A pair is a structural clone only when both the ordered tool trajectory and the identifier-neutral employee job remain above their respective thresholds."
      },
      "semanticIdentityInputsMayEstablishUniqueness": false,
      "derivedVariantsCountAsDistinctEmployeeJobs": false,
      "minimumEvidenceReadsPerTask": 20,
      "minimumProviderEvidenceReadsPerTask": 26,
      "minimumLiveDomainReadsPerTask": 17,
      "layerSpecificReadMinimumsApplyWhenThatEvidenceLayerExists": true,
      "minimumIndependentEvidenceSources": 5,
      "minimumStateChangingCallsPerTask": 1,
      "exactCallOrderGraded": false,
      "causalConstraintsGraded": [
        "required exact evidence selectors succeed",
        "causal evidence is read before the first authorized write",
        "the supported state transition precedes communication",
        "the exact persisted result is read back after the write"
      ]
    },
    "writeContract": {
      "taskScopedExecutionAuthorityRequired": true,
      "completionDestinationMustBeDiscoverable": true,
      "providerInputSchemaRequired": true,
      "freeTextWritesMustUseSemanticGradingOrAgentVisibleTemplate": true,
      "hiddenReferenceTextAllowed": false,
      "hiddenSerializationAllowed": false,
      "postWriteReadbackRequired": true,
      "writeScopeContainmentRequired": true,
      "wrongExistingTargetMustFail": true,
      "keywordStuffingMustFailWhenFreeTextIsGraded": true
    },
    "assetRoom": {
      "minimumAgentVisibleFilesPerTask": 28,
      "minimumNativeFormatsPerTask": 7,
      "requiresStructuredData": true,
      "requiresSpreadsheet": true,
      "requiresDocument": true,
      "requiresCommunicationRecord": true,
      "requiresCurrentAndStaleOrConflictingEvidence": true,
      "requiresCrossFileIdentityAndRevisionCorrelation": true,
      "allAgentVisibleFilesGloballyUnique": true,
      "nativeMagicAndParserValidation": true,
      "visibleAndVerifierOnlyArtifactsSeparated": true,
      "precomputedOutcomeAgentVisible": false,
      "referenceTrajectoryAgentVisible": false,
      "verifierContractAgentVisible": false
    },
    "rubric": {
      "minimumSpecificCriteriaPerTask": 40,
      "requiredCategories": [
        "identity",
        "authority",
        "investigation",
        "correlation",
        "alternatives",
        "decision",
        "procedure",
        "state",
        "answer",
        "containment"
      ],
      "minimumDecisionOptions": 3,
      "genericToolCalledCriteriaAllowed": false,
      "mustExplainCausalDecisionPath": true,
      "exactArgumentsOrStateSelectorsRequired": true,
      "deterministic": true,
      "llmJudgeCalls": 0
    },
    "reasoningChain": {
      "hopClasses": {
        "H1": "scope_identity: the subject record is located through search over immutable identifiers, never guessed from a title or filename.",
        "H2": "requirement_derivation: the required quantity, amount, capacity, or match key is derived from a specification, bill of material, contract, or policy rather than read off a header.",
        "H3": "availability_reconciliation: observed coverage, the excluded or ineligible portion, and the resulting usable coverage are each graded.",
        "H4": "gap_or_bound: the requirement is netted against usable coverage into an uncovered, bounded, or threshold-tested quantity.",
        "H5": "external_constraint: an independently confirmed third-party input (vendor lead time, supplier confirmation, counterparty timing) is read from its own source and applied.",
        "H6": "internal_constraint: capacity, schedule, calendar, or control-window evidence is read from its own source and applied.",
        "H7": "alternatives_with_outcomes: at least three alternatives, each carrying an exact outcome, an incremental cost, and an authority status; at least one needs approval beyond current authority and at least one is feasible-but-inferior or unsupported.",
        "H8": "recommendation: the selected alternative and its outcome are graded.",
        "H9": "control_comparison: the selected outcome is compared with an independently documented business need or control date into a signed variance and an honest timing status.",
        "H10": "authority: the approval record is applied to the selected scope and the unauthorized alternative is not executed.",
        "H11": "controlled_execution: the primary state change carries exact provider-critical values, is read back after the write, and every write stays inside the declared scope.",
        "H12": "communication: a stakeholder artifact carrying the decision, outcome, and constraint is produced for review.",
        "H13": "exact_answer: the answer contract grades the intermediate values as well as the final ones."
      },
      "mandatoryHopClasses": ["H1", "H2", "H3", "H7", "H8", "H9", "H10", "H11", "H12", "H13"],
      "constraintHopClasses": ["H4", "H5", "H6"],
      "minimumConstraintHops": 2,
      "spineHopClasses": ["H2", "H3", "H4", "H5", "H6", "H7", "H8", "H9"],
      "minimumChainDepth": 7,
      "minimumDependentDerivations": 8,
      "minimumSourceSystemsBeforeDecision": 3,
      "minimumEvidenceReadsBeforeDecision": 8,
      "minimumAlternatives": 3,
      "alternativeMustCarry": ["outcome", "incrementalCost", "authorityStatus"],
      "requiresUnauthorizedAlternative": true,
      "requiresInferiorOrUnsupportedAlternative": true,
      "allAlternativeOutcomesGraded": true,
      "minimumGradedAnswerFields": 12,
      "intermediateValuesGraded": true,
      "llmJudgeCalls": 0,
      "auditReport": "benchmark/reports/reasoning-chain-audit.json",
      "auditCommand": "python3 benchmark/reasoning_chain_audit.py --write"
    },
    "sandbox": {
      "isolatedStartingState": true,
      "statefulMcp": true,
      "closedProviderWorldState": true,
      "liveProviderCallsAllowed": false,
      "osLevelNoEgressRequired": false,
      "osLevelNoEgressMustNotBeClaimedWithoutEnforcement": true,
      "verifierHiddenFromAgent": true,
      "referenceSolutionHiddenFromAgent": true,
      "providerContractDisclosureRequired": true,
      "unsupportedOneToOneProviderClaimsAllowed": false,
      "writesMustBeTaskScoped": true
    },
    "qualification": {
      "oracleCoverage": "all released tasks",
      "deterministicReplayCoverage": "all released tasks",
      "negativeControlsPerTask": 10,
      "requiredNegativeControls": [
        "noop",
        "shortcut",
        "state_only",
        "incomplete_read",
        "write_before_read",
        "missing_readback",
        "unauthorized_write",
        "wrong_value",
        "wrong_decision",
        "wrong_evidence"
      ],
      "allNegativeControlsApplicableToEveryTask": true,
      "executionsPerTask": 12,
      "falseAcceptsAllowed": 0,
      "harborOracleCoverage": "all released tasks before publication",
      "publishedTreeRoundTrip": "exact file and digest equality"
    },
    "leaderboard": {
      "partialRunRankedAsReleaseResult": false,
      "staleReleaseRunRankedAsCurrent": false,
      "modelRowRequiresExactReleaseAndFullTaskCoverage": true,
      "modelRowRequiresInspectableTrajectories": true,
      "referenceControlsRankedAsModels": false,
      "missingRunRenderedAsScore": false
    }
  }
}