eval-quality 0.1.0 → 0.3.0

package/dist/core/score/score.js

@@ -0,0 +1,607 @@
+/**
+ * AD-24's score stage: the one orchestration over the reference functions
+ * `core/score/` already ships, closing owed item 1's remaining half (a
+ * trial-set shape for `validated-observations`) and giving `score.module`
+ * its first real value.
+ *
+ * The body lifts `scripts/worked-example-target.ts:1071-1381`'s order
+ * rather than designing a second one, generalised two ways that file never
+ * needed: over `readonly ValidatedObservations[]` instead of one record, and
+ * without ever throwing on a domain input. Every `fail()` call that script
+ * used to stop on a malformed-but-schema-legal shape (a rejected probe, a
+ * clean-control or canary probe with no seeded defect, an oracle with no
+ * check, a resolution naming a finding the trial does not carry) is replaced
+ * here by a graceful, documented fallback: a `schema-parse-failure` already
+ * happened at the application boundary before any of these values reached
+ * this function, so nothing left to see here is a reason to crash.
+ *
+ * Two of `AssessmentCommon`'s fields have no source among this stage's five
+ * declared inputs and are not among its two caller-supplied parameters
+ * either: `evidenceIntegrity.disclosure` (`EvidenceDisclosure`, which lives
+ * only on `SealedRunRecord`, a field `ingest`'s product does not carry
+ * through) and `remediationState` (`AD-12`'s remediation chain, which needs
+ * a caller-presented ancestor sequence score's five inputs never carry, only
+ * the current contract). Both arrive declared with a neutral value, exactly
+ * the posture `ladder.ts`'s own `EvidenceIntegrityInputs` doc comment
+ * already states for `overTruncated`, `unavailable`, and
+ * `internallyInconsistent`: "they arrive declared rather than derived."
+ * `disclosure` and `remediationState` join them on the same reasoning
+ * rather than a new one.
+ */
+import { walkExpression } from '../compile/expression-legality.js';
+import { evaluateCoverage } from '../coverage/coverage.js';
+import { makePointerDenotesCollection, makeResolveOperand, referenceSetKeysOf, } from '../evaluate/evidence-resolution.js';
+import { resolveCheck } from '../evaluate/resolution.js';
+import { SEVERITY_LEVELS, } from '../schemas/eval-contract.js';
+import { buildPlanIndex } from '../seal/plan-index.js';
+import { resolveCapturedBindings, selectWithBindings } from './bindings.js';
+import { resolveContractVerdict, resolveProductionVerdict } from './ladder.js';
+import { FINDING_BUCKETS, resolveOutcome, uncitedDefectFindingGaps, uncitedFindingIds, } from './outcome.js';
+import { resolveHomeOperation, sealProbeSet, } from './qualification.js';
+import { reduceTrialSet, TRIAL_VOTE_STATES, } from './reduce-trials.js';
+import { mapFindings, matchProbeWitness } from './witness.js';
+/** A mutable copy of the two record-shaped arrays every witness/finding function this stage calls wants, since `ValidatedObservations`' own arrays are `readonly`. */
+const recordPickOf = (trial) => ({
+    observations: [...trial.observations],
+    findings: [...trial.findings],
+});
+/**
+ * Every interaction-rooted step identifier one oracle's `check` addresses.
+ * Reimplemented from `scripts/worked-example-target.ts`'s own private
+ * helper of the same name: nothing in `src/` exports it, and `core/` cannot
+ * import `scripts/`.
+ */
+function addressedSteps(expression) {
+    const found = new Set();
+    const take = (operand) => {
+        if (!('pointer' in operand))
+            return;
+        const { pointer } = operand;
+        if (pointer.startsWith('@'))
+            return;
+        const [, root, stepId] = pointer.split('/');
+        if (root === 'interactions' && stepId !== undefined)
+            found.add(stepId);
+    };
+    walkExpression(expression, 0, '', { onOperand: take, onSetOperand: take });
+    return found;
+}
+/**
+ * AD-40 pairs a probe with exactly one designated oracle: the one
+ * discharging the behaviour its seeded defect breaks. Anchored on
+ * `probe.behaviorId` rather than `probe.defects[0]?.behaviorId` (the worked
+ * example's own anchor): AD-9 puts the behaviour on the probe itself, on
+ * every branch and every class, while `defects` is empty on a canary and
+ * absent-shaped on a clean control, so an anchor scoped to `defects[0]`
+ * cannot generalise across probe classes the way this stage's "never throw
+ * on a domain input" rule requires. Returns `null`, never throws, when the
+ * probe's own behaviour resolves to no oracle or to more than one: a real,
+ * reportable authoring gap rather than a crash.
+ */
+function designatedOracleIdOf(probe, contract) {
+    const behavior = contract.behaviors.find((entry) => entry.id === probe.behaviorId);
+    if (behavior === undefined)
+        return null;
+    if (behavior.oracles.length !== 1)
+        return null;
+    return behavior.oracles[0] ?? null;
+}
+/** A probe on the seeding branch whose signature is present, or `null` for a clean control, a canary, or a signature-less defect probe -- none of which AD-40's witness match applies to. */
+function signedProbeOf(probe) {
+    if (probe.expectedClean)
+        return null;
+    if (probe.defectSignature === null)
+        return null;
+    return probe;
+}
+/**
+ * The highest severity among every behaviour this oracle discharges, the
+ * same source the worked example reads (`resolution.resolvedFrom === null`
+ * case). The maximum, not the first, on `coverage.ts`'s `coverageSeverity`
+ * precedent: a lower reduction would let one trivial behaviour understate a
+ * shared oracle's real severity. `'low'` when no behaviour names the oracle
+ * at all -- a genuine authoring gap this pure function has no better signal
+ * for, and never a reason to throw.
+ */
+function severityOfBehaviourFor(contract, oracleId) {
+    let best = null;
+    for (const behavior of contract.behaviors) {
+        if (!behavior.oracles.includes(oracleId))
+            continue;
+        if (best === null ||
+            SEVERITY_LEVELS.indexOf(behavior.severity) > SEVERITY_LEVELS.indexOf(best)) {
+            best = behavior.severity;
+        }
+    }
+    return best ?? 'low';
+}
+/**
+ * `judgeConduct` derives once per run, not per oracle-criterion, since no
+ * schema field maps a rubric criterion to an oracle. `'absent'` is
+ * `outcome.ts:75`'s own "ordinary value" case; `'malformed'` when any trial
+ * carries a `judge-result-unscored` condition; `'conforming'` otherwise.
+ */
+function judgeConductOf(contract, trials) {
+    if (contract.rubrics.length === 0)
+        return 'absent';
+    const anyMalformed = trials.some((trial) => trial.conditions.some((condition) => condition.kind === 'judge-result-unscored'));
+    return anyMalformed ? 'malformed' : 'conforming';
+}
+/**
+ * The ninth new Invalid condition: an observation whose `operationId`
+ * matches an operation declared in more than one `permittedInterfaces`
+ * entry. Not an `IngestCondition` -- ingest never computes it, since it has
+ * no `eval-contract` input and `Observation` carries no interface
+ * qualifier -- so `score.ts` renders each basis line itself.
+ */
+function operationIdentifierCollisionsOf(contract, trials) {
+    const interfacesByOperationId = new Map();
+    for (const iface of contract.permittedInterfaces) {
+        // Deduplicated per interface first: `PermittedInterface.operations`
+        // carries no uniqueness constraint, so one interface declaring the
+        // same `operationId` on two different operations must still count as
+        // one interface, not two, or this row would falsely fire a collision
+        // naming the same interface twice.
+        const operationIdsInThisInterface = new Set(iface.operations.map((operation) => operation.operationId));
+        for (const operationId of operationIdsInThisInterface) {
+            const entry = interfacesByOperationId.get(operationId);
+            if (entry === undefined) {
+                interfacesByOperationId.set(operationId, [iface.logicalId]);
+            }
+            else {
+                entry.push(iface.logicalId);
+            }
+        }
+    }
+    const collisions = [];
+    trials.forEach((trial, trialIndex) => {
+        for (const observation of trial.observations) {
+            const interfaces = interfacesByOperationId.get(observation.operationId) ?? [];
+            if (interfaces.length <= 1)
+                continue;
+            collisions.push(`trial ${trialIndex + 1} observation ${observation.observationId}: operationId "${observation.operationId}" matches operations in ${interfaces.length} permittedInterfaces entries (${interfaces.join(', ')})`);
+        }
+    });
+    return collisions;
+}
+/**
+ * The tenth new Invalid condition: a caller assembling a trial set from
+ * records that disagree on `mode`, `evaluatorRecommendation`, or `runId`.
+ * Every trial is compared against the first: a trial set is not a genuine
+ * set once one trial's own value is picked as authoritative, regardless of
+ * which one, so basis lines name every disagreeing pair. `runId` joined the
+ * other two once `ValidatedObservations` carried it: batching trials from
+ * two different runs into one trial set is the single most important
+ * cross-trial mixup this check exists to catch, and it read the same
+ * fallback posture as the other two without being compared like them.
+ */
+function trialSetDisagreementsOf(trials) {
+    const first = trials[0];
+    if (first === undefined)
+        return [];
+    const disagreements = [];
+    trials.forEach((trial, index) => {
+        if (index === 0)
+            return;
+        if (trial.mode !== first.mode) {
+            disagreements.push(`mode: trial 1 = "${first.mode}", trial ${index + 1} = "${trial.mode}"`);
+        }
+        if (trial.evaluatorRecommendation !== first.evaluatorRecommendation) {
+            disagreements.push(`evaluatorRecommendation: trial 1 = "${first.evaluatorRecommendation}", trial ${index + 1} = "${trial.evaluatorRecommendation}"`);
+        }
+        if (trial.runId !== first.runId) {
+            disagreements.push(`runId: trial 1 = "${first.runId}", trial ${index + 1} = "${trial.runId}"`);
+        }
+    });
+    return disagreements;
+}
+/**
+ * Cross-trial identifier reuse is NOT a collision, and this is where that was
+ * settled after being carried as an open question.
+ *
+ * `ingest` computes `duplicate-record-identifier` per sealed run record, since
+ * it sees one trial at a time, and nothing checks whether two DIFFERENT trials
+ * of one set reuse an observation, finding, or oracle-disposition identifier.
+ * That asymmetry looks like a gap beside `trial-set-field-disagreement`, which
+ * does compare `mode` and `evaluatorRecommendation` across trials.
+ *
+ * It is not one. A trial set is n independent evaluator runs of one contract,
+ * each producing its own record, and a harness that names its first observation
+ * `obs-1` names it that in every run. Reporting the second run for it would
+ * make a repeated run Invalid by construction, which is the opposite of what a
+ * trial set is for. The fields that ARE compared across trials are compared
+ * because they describe the SET rather than a run within it: two trials
+ * disagreeing about which mode they ran under cannot both be true of one set,
+ * while two trials both carrying `obs-1` are two true statements about two
+ * runs.
+ *
+ * The pooled basis this stage builds carries each entry with the trial it came
+ * from, so a repeated identifier is never resolved against the wrong row.
+ */
+/** Every `IngestCondition` of one `kind` across every trial, in trial order. */
+function conditionsAcrossTrials(trials, kind) {
+    return trials.flatMap((trial) => trial.conditions.filter((condition) => condition.kind === kind));
+}
+/**
+ * Reused from `reduce-trials.ts` rather than a second literal: `resolveOutcome`
+ * is the one AD-6 state assigner, and re-listing the three AD-6 invalidating
+ * states here would be a second place that vocabulary could drift from the
+ * reducer's own `TRIAL_VOTE_STATES.invalidating` grouping.
+ */
+const INVALIDATING_OUTCOME_STATES = new Set(TRIAL_VOTE_STATES.invalidating);
+/**
+ * The stage. Signature order matches `ScoreStage`'s own: the five declared
+ * artifact inputs, then `waiver` and `evaluationFault`, the two documented
+ * caller-supplied parameters -- neither has a source among those five, and
+ * each arrives named and explicit rather than a hardcoded literal.
+ */
+export const score = (contract, trials, probe, preflightVerdict, policy, waiver, evaluationFault) => {
+    // Probe sealing: once per run, never per trial, since qualification reads
+    // the probe and the contract's operation inventory alone. `probeQualified`
+    // reads whichever bucket the probe actually lands in -- never a throw on
+    // rejection, since a rejected probe is a legitimate domain outcome the
+    // existing `unqualified-probe-in-sealed-set` condition already reports.
+    const homeOperationOf = (candidate) => candidate.expectedClean || candidate.defectSignature === null
+        ? null
+        : resolveHomeOperation(candidate.defectSignature, contract.permittedInterfaces);
+    const sealedProbes = sealProbeSet([probe], homeOperationOf);
+    const qualifiedEntry = sealedProbes.admitted[0] ?? sealedProbes.rejected[0];
+    const probeQualified = qualifiedEntry === undefined ? false : qualifiedEntry.result.qualified;
+    const signedProbe = signedProbeOf(probe);
+    const designatedOracleId = designatedOracleIdOf(probe, contract);
+    const probeSigned = !probe.expectedClean && probe.defectSignature !== null;
+    // Plan indexing and the resolvers built from it: contract-only, so built
+    // once and reused across every trial. `resolveCapturedBindings` walks
+    // `bindingOrder`'s own tiers internally and already degrades a cyclic
+    // plan to "resolves as unlisted, filters every candidate away" without
+    // throwing, so this stage does not re-run that check itself.
+    // `duplicateIds: 'unresolved'` rather than `buildPlanIndex`'s own default
+    // `'throw'`: two `permittedInterfaces` entries sharing an operationId is
+    // exactly the domain input `operationIdentifierCollisionsOf` below
+    // reports as `operation-identifier-collision`, so this stage cannot let
+    // the index builder crash on the same shape its own new Invalid row
+    // exists to describe.
+    const index = buildPlanIndex(contract.interactionPlan, contract.permittedInterfaces, {
+        duplicateIds: 'unresolved',
+    });
+    const pointerDenotesCollection = makePointerDenotesCollection(contract, index);
+    const referenceSets = Object.fromEntries(Object.entries(contract.referenceSets ?? {}).map(([id, declaration]) => [
+        id,
+        declaration.members,
+    ]));
+    // The declared keys the members map above drops. `set-membership`'s set
+    // position reads the single declared key off each member.
+    const referenceSetKeys = referenceSetKeysOf(contract);
+    const judgeConduct = judgeConductOf(contract, trials);
+    const allOutcomes = [];
+    // The full `EvidenceArtifact.outcomes` shape, built alongside `allOutcomes`
+    // above in the same loop rather than derived from it after the fact: the
+    // two carry different fields from the same per-oracle locals and neither
+    // is reconstructible from the other.
+    const outcomes = [];
+    const votes = [];
+    for (const trial of trials) {
+        const captured = resolveCapturedBindings(contract.interactionPlan, index, trial.observations);
+        const selectionOf = new Map();
+        for (const step of contract.interactionPlan) {
+            selectionOf.set(step.stepId, selectWithBindings(step, trial.observations, index, captured));
+        }
+        // One observation per step, mirroring `resolveTemporalAnchor`'s own
+        // rule: one match binds, several under a declared `any` binds the
+        // lowest sequence, anything else binds nothing. A step this trial
+        // never selected, or whose matched id names no observation, is
+        // skipped rather than failed: a check addressing that step simply
+        // resolves against an absent operand, which `resolveCheck` already
+        // handles.
+        const observationById = new Map(trial.observations.map((observation) => [
+            observation.observationId,
+            observation,
+        ]));
+        const stepObservations = {};
+        for (const step of contract.interactionPlan) {
+            const selection = selectionOf.get(step.stepId);
+            if (selection === undefined)
+                continue;
+            const [first] = selection.matchedObservationIds;
+            if (first === undefined)
+                continue;
+            if (selection.result === 'several' && step.cardinality !== 'any')
+                continue;
+            const observation = observationById.get(first);
+            if (observation === undefined)
+                continue;
+            stepObservations[step.stepId] = observation;
+        }
+        const resolveOperand = makeResolveOperand(stepObservations, referenceSets);
+        const recordPick = recordPickOf(trial);
+        const witness = signedProbe === null
+            ? null
+            : matchProbeWitness(signedProbe, contract.permittedInterfaces, recordPick);
+        const findingMap = mapFindings([probe], contract.permittedInterfaces, recordPick);
+        const bucketOf = new Map();
+        for (const bucket of FINDING_BUCKETS) {
+            for (const entry of findingMap[bucket])
+                bucketOf.set(entry.findingId, bucket);
+        }
+        let designatedState;
+        let firstInvalidatingState;
+        let firstState;
+        for (const oracle of contract.oracles) {
+            const check = oracle.check;
+            const steps = check === null ? new Set() : addressedSteps(check);
+            const addressed = contract.interactionPlan
+                .filter((step) => steps.has(step.stepId))
+                .map((step) => ({
+                step,
+                selection: selectionOf.get(step.stepId) ?? {
+                    result: 'none',
+                    matchedObservationIds: [],
+                },
+            }));
+            const selections = addressed.map((entry) => entry.selection);
+            const selectorAmbiguity = addressed.some((entry) => entry.selection.result === 'several' &&
+                entry.step.cardinality !== 'any');
+            const checkResolution = check === null
+                ? null
+                : resolveCheck(check, resolveOperand, pointerDenotesCollection, referenceSetKeys, policy.regexMatchStepBudget, `EvalContract.oracles[id=${oracle.id}].check`);
+            // Every disposition this trial records for this oracle, not just
+            // the first: `SealedRunRecord.oracleDispositions` carries no
+            // uniqueness constraint on `oracleId`, and ingest's
+            // `duplicate-record-identifier` condition only advisory-flags a
+            // repeat, never rejects it, so two dispositions naming one oracle
+            // is a legal, ambiguous input. More than one match is treated as
+            // ambiguous -- `disposition: null` -- rather than silently
+            // picking the array's first entry, the same guard-rather-than-pick
+            // idiom `citedFinding` below already uses.
+            const matchingDispositions = trial.dispositions.filter((entry) => entry.oracleId === oracle.id);
+            const [onlyDisposition] = matchingDispositions;
+            const disposition = matchingDispositions.length === 1 && onlyDisposition !== undefined
+                ? onlyDisposition
+                : null;
+            // Every defect finding this trial cites against this oracle, not
+            // just the first: two distinct findings citing the same oracle is
+            // schema-legal and not caught by ingest's `duplicate-record-identifier`
+            // (that one is keyed by `findingId`, not `oracleId`). Picking either
+            // one arbitrarily would make the outcome depend on array order, so
+            // more than one match is treated as ambiguous -- `citedFinding: null`
+            // -- rather than silently choosing one, the same guard-rather-than-pick
+            // idiom `outcomesByProbeId` (`strength.ts`) already uses.
+            const defectFindingsForOracle = trial.findings.filter((finding) => finding.findingType === 'defect' && finding.oracleId === oracle.id);
+            const [onlyDefectFinding] = defectFindingsForOracle;
+            const citedFinding = defectFindingsForOracle.length === 1 && onlyDefectFinding !== undefined
+                ? {
+                    findingId: onlyDefectFinding.findingId,
+                    // `mapFindings` buckets every defect finding in this trial, so a
+                    // miss here cannot happen for a finding this same trial
+                    // produced; the fallback is fail-closed, never a throw.
+                    bucket: bucketOf.get(onlyDefectFinding.findingId) ?? 'dangling',
+                }
+                : null;
+            const inputs = {
+                required: true,
+                disposition,
+                citedFinding,
+                witness: designatedOracleId !== null && oracle.id === designatedOracleId
+                    ? witness
+                    : null,
+                selections,
+                selectorAmbiguity,
+                checkResolution: checkResolution === null ? null : checkResolution.resolution,
+                polarity: oracle.polarity,
+                probeClass: probe.probeClass,
+                expectedClean: probe.expectedClean,
+                probeSigned,
+                probeQualified,
+                waiver,
+                judgeConduct,
+                evaluationFault,
+            };
+            const resolution = resolveOutcome(inputs);
+            // The same ambiguity guard as `disposition` and `citedFinding`
+            // above: `findingId` carries no uniqueness constraint either (also
+            // only advisory-flagged by `duplicate-record-identifier`, subject
+            // `'finding'`), so two findings sharing `resolution.resolvedFrom`
+            // is legal. Reading either one's `.severity` would make the
+            // outcome depend on array order; more than one match falls
+            // through to the same behaviour-severity floor a missing match
+            // already uses, rather than picking one arbitrarily.
+            const findingsResolvedFrom = resolution.resolvedFrom === null
+                ? []
+                : trial.findings.filter((finding) => finding.findingId === resolution.resolvedFrom);
+            const [onlyFindingResolvedFrom] = findingsResolvedFrom;
+            const severity = findingsResolvedFrom.length === 1 &&
+                onlyFindingResolvedFrom !== undefined
+                ? onlyFindingResolvedFrom.severity
+                : severityOfBehaviourFor(contract, oracle.id);
+            allOutcomes.push({
+                oracleId: oracle.id,
+                required: true,
+                severity,
+                checkResolved: inputs.checkResolution !== null,
+                resolution,
+            });
+            outcomes.push({
+                oracleId: oracle.id,
+                // Constant across every entry: one `score()` call scores exactly
+                // one probe.
+                probeId: probe.probeId,
+                state: resolution.state,
+                severity,
+                // `ORACLE_DISPOSITIONS`' third member, `'not-attempted'`, on a
+                // `null` local `disposition`: no disposition was recorded for
+                // this oracle, or the ambiguity guard above fired. Both mean
+                // "nothing was recorded", which is the honest reading of the
+                // one closed-three member that says so.
+                disposition: disposition === null ? 'not-attempted' : disposition.disposition,
+                resolvedFrom: resolution.resolvedFrom,
+                corroboration: resolution.corroboration,
+                selectedObservationIds: [...resolution.selectedObservationIds],
+                checkResolution,
+            });
+            if (firstState === undefined)
+                firstState = resolution.state;
+            if (firstInvalidatingState === undefined &&
+                INVALIDATING_OUTCOME_STATES.has(resolution.state)) {
+                firstInvalidatingState = resolution.state;
+            }
+            // First match only: `Oracle.id` carries no schema- or compile-time
+            // uniqueness constraint, so two oracles could share
+            // `id === designatedOracleId`. Guarding against overwrite here is
+            // the same last-write-wins fix `outcomesByProbeId` (`strength.ts`)
+            // already applies, for the identical reason.
+            if (designatedOracleId !== null &&
+                oracle.id === designatedOracleId &&
+                designatedState === undefined) {
+                designatedState = resolution.state;
+            }
+        }
+        // One vote per trial, ordinarily: the trial set's own cardinality
+        // (`Trials.completed`) is `votes.length`, so a trial still
+        // contributes a vote when the probe has no designated oracle (a
+        // clean control, a canary, or a malformed defect chain) -- the
+        // fallback order is the first invalidating state this trial's
+        // oracles produced, else the first oracle's state. Every candidate
+        // is a state `resolveOutcome` itself assigned this trial, never a
+        // literal this module invents: `resolveOutcome` stays the one
+        // assigner of an AD-6 state. `contract.oracles` declaring none at
+        // all is the one shape with no state to vote with, and that trial
+        // contributes no vote rather than a fabricated one.
+        const voteState = designatedState ?? firstInvalidatingState ?? firstState;
+        if (voteState !== undefined) {
+            votes.push({ state: voteState });
+        }
+    }
+    const reduced = reduceTrialSet(votes, policy.catchThreshold);
+    const trialsField = {
+        declaredMinimum: policy.minimumTrialCount,
+        completed: votes.length,
+        invalidatedAttempts: [...reduced.invalidatedAttempts],
+    };
+    const unwitnessedQuotations = trials.flatMap((trial) => trial.unwitnessedQuotations);
+    const isolationViolation = trials.flatMap((trial) => trial.isolationViolation);
+    const findings = trials.flatMap((trial) => trial.findings.map((finding) => ({
+        findingId: finding.findingId,
+        confidence: finding.confidence,
+    })));
+    const uncitedDefectFindings = trials.flatMap((trial) => uncitedDefectFindingGaps(recordPickOf(trial)));
+    // Every finding across every trial citing no oracle (`emit`'s own
+    // `uncitedFindings` field): broader than `uncitedDefectFindings` above
+    // (every finding type, not `defect` only) and thinner (an identifier
+    // only), per `outcome.ts`'s own doc comment on the two functions
+    // coexisting rather than one replacing the other.
+    const uncitedFindings = trials.flatMap((trial) => uncitedFindingIds(recordPickOf(trial)));
+    const coverageGaps = evaluateCoverage(contract);
+    const operationIdentifierCollisions = operationIdentifierCollisionsOf(contract, trials);
+    const trialSetDisagreements = trialSetDisagreementsOf(trials);
+    const evidenceIntegrity = {
+        // Declared, not derived: no declared input or caller-supplied
+        // parameter carries `EvidenceDisclosure`, `ingest`'s product does not
+        // restate it, and the other three siblings are declared `false` for
+        // the identical reason (see this module's own header comment).
+        disclosure: { truncationBound: null, reportedIncomplete: false },
+        overTruncated: false,
+        unavailable: false,
+        internallyInconsistent: false,
+        isolationViolation,
+        duplicateRecordIdentifiers: conditionsAcrossTrials(trials, 'duplicate-record-identifier'),
+        danglingCitations: conditionsAcrossTrials(trials, 'dangling-citation'),
+        danglingDispositionCitations: conditionsAcrossTrials(trials, 'dangling-disposition-citation'),
+        forbiddenInputsNotWithheld: conditionsAcrossTrials(trials, 'forbidden-input-not-withheld'),
+        crossArtifactDisagreements: conditionsAcrossTrials(trials, 'cross-artifact-disagreement'),
+        evaluatorConfigurationAbsent: conditionsAcrossTrials(trials, 'evaluator-configuration-absent'),
+        evaluatorConfigurationDigestMismatches: conditionsAcrossTrials(trials, 'evaluator-configuration-digest-mismatch'),
+        judgeResultsUnscored: conditionsAcrossTrials(trials, 'judge-result-unscored'),
+        operationIdentifierCollisions,
+        trialSetDisagreements,
+    };
+    // Every trial in the set is asserted to agree with the
+    // first on `mode` and `evaluatorRecommendation`; a disagreement is
+    // `trial-set-field-disagreement` above, never a throw. The first
+    // trial's own values build the one assessment TypeScript's
+    // discriminated union still requires -- non-silence comes from the
+    // Invalid row's basis line, not from withholding a value. A caller
+    // supplying zero trials has no first value to read; both fields fall
+    // back to a neutral default, and `below-minimum-trial-count` already
+    // reports zero completed trials against the declared minimum.
+    const firstTrial = trials[0];
+    const mode = firstTrial?.mode ?? 'production';
+    const evaluatorRecommendation = firstTrial?.evaluatorRecommendation ?? 'PASS';
+    // Same posture as `mode`/`evaluatorRecommendation` above: read off the
+    // first trial, never derived or defaulted from anything richer. A caller
+    // supplying zero trials has no first value to read either; the empty
+    // string is the honest "no run identifier was presented" reading, on the
+    // same terms `below-minimum-trial-count` already reports the shortfall.
+    const runId = firstTrial?.runId ?? '';
+    const commonBody = {
+        outcomeState: {
+            outcomes: allOutcomes,
+            unwitnessedQuotations,
+            trials: trialsField,
+            reExecutionCap: policy.reExecutionCap,
+        },
+        evidenceIntegrity,
+        evaluatorRecommendation,
+        coverageGaps,
+        uncitedDefectFindings,
+        findings,
+        confidenceThreshold: policy.confidenceThreshold,
+        // Declared, not derived, on the same posture as `disclosure` below:
+        // `AD-12`'s remediation chain needs a caller-presented ancestor
+        // sequence, and score's five declared inputs carry only the current
+        // contract, never a chain. `validateLineageChain([contract], {...})`
+        // looks buildable from what score already has, but is not: it treats
+        // the one-element array as the WHOLE chain, so any contract with
+        // `revisionCount > 0` -- an ordinarily revised one -- fails the root,
+        // parent-resolution, and length checks unconditionally, firing the
+        // existing `lineage-chain-inconsistent` FAIL row on every such run
+        // regardless of whether anything is actually wrong. The vacuously
+        // true value below is what "no chain was presented to validate"
+        // honestly means.
+        remediationState: {
+            lengthConsistent: true,
+            noRepeatedDigest: true,
+            noGap: true,
+        },
+        preflightPassed: preflightVerdict.passed,
+        severityFloor: policy.severityFloor,
+    };
+    if (mode === 'production') {
+        const assessment = {
+            mode: 'production',
+            ...commonBody,
+        };
+        return {
+            assessment,
+            ladder: resolveProductionVerdict(assessment),
+            runId,
+            contract,
+            policy,
+            probe,
+            sealedProbes,
+            trialSetResult: reduced,
+            outcomes,
+            uncitedFindings,
+        };
+    }
+    const assessment = {
+        mode: 'contract-scoring',
+        ...commonBody,
+        systemRecommendationRecorded: evaluatorRecommendation,
+        // No declared input or caller-supplied parameter carries authored
+        // prose for this field; `null` is its own legal, honest value rather
+        // than an invented note.
+        systemRecommendationNote: null,
+    };
+    return {
+        assessment,
+        ladder: resolveContractVerdict(assessment),
+        runId,
+        contract,
+        policy,
+        probe,
+        sealedProbes,
+        trialSetResult: reduced,
+        outcomes,
+        uncitedFindings,
+    };
+};