universal-physics-tensor 0.45.1 → 0.46.0

package/dist/atlas/benchmark/loader.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"loader.d.ts","sourceRoot":"","sources":["../../../src/atlas/benchmark/loader.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAKH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAEhD,0DAA0D;AAC1D,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED;;;;;;;;;GASG;AACH,wBAAgB,aAAa,CAAC,KAAK,EAAE,SAAS,aAAa,EAAE,EAAE,MAAM,EAAE,OAAO,GAAG,WAAW,EAAE,CA0B7F;AAED,kFAAkF;AAClF,qBAAa,uBAAwB,SAAQ,KAAK;IACpC,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE;IAArD,YAAqB,QAAQ,EAAE,SAAS,WAAW,EAAE,EAMpD;CACF;AAQD;;;;;;GAMG;AACH,wBAAgB,eAAe,CAAC,YAAY,EAAE,MAAM,GAAG,aAAa,EAAE,CAKrE;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,YAAY,EAAE,MAAM,GAAG,aAAa,EAAE,CAKzE"}

package/dist/atlas/benchmark/loader.js

@@ -0,0 +1,104 @@
+/**
+ * Atlas Phase 5, S5.1 — load and validate benchmark items.
+ *
+ * Design note: `docs/planning/Atlas-Phase-5-Design.md` §0, §2.
+ *
+ * This module reads the PUBLIC half of a benchmark directory and nothing else.
+ * The answers live in a sibling half that no `src/` file may read;
+ * `tests/atlas/benchmark.test.ts` holds the import guard, with a positive
+ * control proving the guard can fire.
+ *
+ * **The loader enforces the one part of the independence rule that code CAN
+ * enforce: a frozen item whose `authorship` is not `'independent'` is refused.**
+ * Whether the author really was independent is a fact about people, recorded in
+ * the pre-registration note and checked by review.
+ *
+ * @module atlas/benchmark/loader
+ * @internal
+ */
+import { readFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { HELD_OUT_FAMILY } from './types.js';
+/**
+ * Validate an item set against the schema and the rules that make it a benchmark.
+ *
+ * @param items - the items, as parsed.
+ * @param frozen - true for the frozen set (authorship must be `'independent'`),
+ * false for the contested staging area (authorship must be `'contested-draft'`).
+ * @returns every problem found; empty means admissible. Never throws, so a
+ * caller sees ALL problems at once rather than the first.
+ * @internal
+ */
+export function validateItems(items, frozen) {
+    const problems = [];
+    const seen = new Set();
+    for (const item of items) {
+        const p = (problem) => {
+            problems.push({ id: item.id, problem });
+        };
+        if (seen.has(item.id))
+            p('duplicate id');
+        seen.add(item.id);
+        const wanted = frozen ? 'independent' : 'contested-draft';
+        if (item.authorship !== wanted) {
+            p(`authorship '${item.authorship}' where the ${frozen ? 'frozen' : 'contested'} set requires '${wanted}'`);
+        }
+        // The answer lives in the scorer half. A public item that carries it would
+        // hand the key to any condition that reads the public file.
+        if ('kind' in item || 'failureKind' in item) {
+            p('the public half carries the answer (kind/failureKind); it belongs in scorer/labels.json');
+        }
+        if ((item.family === HELD_OUT_FAMILY) !== (item.split === 'held-out')) {
+            p(`the held-out split is exactly the '${HELD_OUT_FAMILY}' family; got family '${item.family}' in split '${item.split}'`);
+        }
+        if (item.premises.length === 0)
+            p('no premises');
+        if (item.source.trim() === '')
+            p('no source');
+    }
+    return problems;
+}
+/** Thrown when a benchmark directory holds an inadmissible item set. @internal */
+export class BenchmarkAdmissionError extends Error {
+    problems;
+    constructor(problems) {
+        super(`benchmark items are not admissible: ${problems
+            .map((p) => `${p.id}: ${p.problem}`)
+            .join('; ')}`);
+        this.problems = problems;
+    }
+}
+function readItems(file) {
+    const parsed = JSON.parse(readFileSync(file, 'utf-8'));
+    if (!Array.isArray(parsed))
+        throw new Error(`${file}: expected a JSON array of items`);
+    return parsed;
+}
+/**
+ * Load the FROZEN items from `<benchmarkDir>/public/items.json`.
+ *
+ * @throws BenchmarkAdmissionError when any item is inadmissible — a frozen set
+ * with one contested draft in it is not a smaller benchmark, it is a broken one.
+ * @internal
+ */
+export function loadFrozenItems(benchmarkDir) {
+    const items = readItems(join(benchmarkDir, 'public', 'items.json'));
+    const problems = validateItems(items, true);
+    if (problems.length > 0)
+        throw new BenchmarkAdmissionError(problems);
+    return items;
+}
+/**
+ * Load the CONTESTED drafts from `<benchmarkDir>/contested/items.json`. These are
+ * never scored; they wait for an independent author to accept or rewrite them.
+ *
+ * @internal
+ */
+export function loadContestedDrafts(benchmarkDir) {
+    const items = readItems(join(benchmarkDir, 'contested', 'items.json'));
+    const problems = validateItems(items, false);
+    if (problems.length > 0)
+        throw new BenchmarkAdmissionError(problems);
+    return items;
+}
+//# sourceMappingURL=loader.js.map

package/dist/atlas/benchmark/loader.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"loader.js","sourceRoot":"","sources":["../../../src/atlas/benchmark/loader.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,EAAE,YAAY,EAAE,MAAM,SAAS,CAAC;AACvC,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AACjC,OAAO,EAAE,eAAe,EAAE,MAAM,YAAY,CAAC;AAS7C;;;;;;;;;GASG;AACH,MAAM,UAAU,aAAa,CAAC,KAA+B,EAAE,MAAe;IAC5E,MAAM,QAAQ,GAAkB,EAAE,CAAC;IACnC,MAAM,IAAI,GAAG,IAAI,GAAG,EAAU,CAAC;IAC/B,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,MAAM,CAAC,GAAG,CAAC,OAAe,EAAQ,EAAE;YAClC,QAAQ,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,OAAO,EAAE,CAAC,CAAC;QAC1C,CAAC,CAAC;QACF,IAAI,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;YAAE,CAAC,CAAC,cAAc,CAAC,CAAC;QACzC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QAElB,MAAM,MAAM,GAAG,MAAM,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,iBAAiB,CAAC;QAC1D,IAAI,IAAI,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;YAC/B,CAAC,CAAC,eAAe,IAAI,CAAC,UAAU,eAAe,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,WAAW,kBAAkB,MAAM,GAAG,CAAC,CAAC;QAC7G,CAAC;QACD,2EAA2E;QAC3E,4DAA4D;QAC5D,IAAI,MAAM,IAAI,IAAI,IAAI,aAAa,IAAI,IAAI,EAAE,CAAC;YAC5C,CAAC,CAAC,yFAAyF,CAAC,CAAC;QAC/F,CAAC;QACD,IAAI,CAAC,IAAI,CAAC,MAAM,KAAK,eAAe,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,KAAK,UAAU,CAAC,EAAE,CAAC;YACtE,CAAC,CAAC,sCAAsC,eAAe,yBAAyB,IAAI,CAAC,MAAM,eAAe,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC;QAC3H,CAAC;QACD,IAAI,IAAI,CAAC,QAAQ,CAAC,MAAM,KAAK,CAAC;YAAE,CAAC,CAAC,aAAa,CAAC,CAAC;QACjD,IAAI,IAAI,CAAC,MAAM,CAAC,IAAI,EAAE,KAAK,EAAE;YAAE,CAAC,CAAC,WAAW,CAAC,CAAC;IAChD,CAAC;IACD,OAAO,QAAQ,CAAC;AAClB,CAAC;AAED,kFAAkF;AAClF,MAAM,OAAO,uBAAwB,SAAQ,KAAK;IAC3B,QAAQ;IAA7B,YAAqB,QAAgC;QACnD,KAAK,CACH,uCAAuC,QAAQ;aAC5C,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC,OAAO,EAAE,CAAC;aACnC,IAAI,CAAC,IAAI,CAAC,EAAE,CAChB,CAAC;wBALiB,QAAQ;IAM7B,CAAC;CACF;AAED,SAAS,SAAS,CAAC,IAAY;IAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,EAAE,OAAO,CAAC,CAAY,CAAC;IAClE,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,GAAG,IAAI,kCAAkC,CAAC,CAAC;IACvF,OAAO,MAAyB,CAAC;AACnC,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,eAAe,CAAC,YAAoB;IAClD,MAAM,KAAK,GAAG,SAAS,CAAC,IAAI,CAAC,YAAY,EAAE,QAAQ,EAAE,YAAY,CAAC,CAAC,CAAC;IACpE,MAAM,QAAQ,GAAG,aAAa,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC;IAC5C,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC;QAAE,MAAM,IAAI,uBAAuB,CAAC,QAAQ,CAAC,CAAC;IACrE,OAAO,KAAK,CAAC;AACf,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,YAAoB;IACtD,MAAM,KAAK,GAAG,SAAS,CAAC,IAAI,CAAC,YAAY,EAAE,WAAW,EAAE,YAAY,CAAC,CAAC,CAAC;IACvE,MAAM,QAAQ,GAAG,aAAa,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;IAC7C,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC;QAAE,MAAM,IAAI,uBAAuB,CAAC,QAAQ,CAAC,CAAC;IACrE,OAAO,KAAK,CAAC;AACf,CAAC"}

package/dist/atlas/benchmark/run-atlas.d.ts

@@ -0,0 +1,72 @@
+/**
+ * Atlas Phase 5, S5.2 — the ATLAS condition of the invalid-bridge benchmark.
+ *
+ * Design note: `docs/planning/Atlas-Phase-5-Design.md` §6.
+ *
+ * For each item the runner applies three atlas instruments — the applicability
+ * checker (S4.1), the composition table (S3) and the regime check (S2) — and
+ * emits `accept`, `reject` or `abstain`.
+ *
+ * ## The decision rule, and why accept is the HARDEST outcome to reach
+ *
+ * - **reject** when an instrument DEMONSTRABLY fires: a blocking applicability
+ *   finding, a chain OVERCLAIMED as an exact equivalence, a structural analogy
+ *   promoted to an equivalence, or a regime inequality checked and violated.
+ * - **accept** only when EVERY ENABLED instrument RAN and CLEARED: side conditions were
+ *   stated, the applicability checker found nothing, the regime was supplied and
+ *   every inequality was checked and held, and any claimed chain agrees with the
+ *   table. At least one instrument must have RUN: an item that no enabled
+ *   instrument applies to abstains, because nothing cleared it. S4.1 already fixed the principle: an empty finding list means "no rule
+ *   fired", which is weaker than "valid". An item that leaves a check unrun has
+ *   not been cleared by it.
+ * - **abstain** otherwise. Abstention is a first-class outcome and is preferred
+ *   to a wrong accept (ROADMAP Phase 5, pre-stated criteria).
+ *
+ * The failure kind a rejection names is a MAPPING from the rule that fired, and
+ * it is only as good as that mapping: a dimensional inconsistency is reported as
+ * `notation-collision` because that is its commonest cause, not because the rule
+ * can see notation. The scorer grades "rejected" and "named the right kind"
+ * separately, so an imperfect mapping costs the second score only.
+ *
+ * @module atlas/benchmark/run-atlas
+ * @internal
+ */
+import type { BenchmarkItem, FailureKind } from './types.js';
+/** What the atlas condition concluded about one item. @internal */
+export interface AtlasVerdict {
+    readonly itemId: string;
+    readonly outcome: 'accept' | 'reject' | 'abstain';
+    /** Present on a reject when the rule that fired maps to a failure kind. */
+    readonly detectedFailure?: FailureKind;
+    /** Every reason, in rule order — including the unrun checks behind an abstention. */
+    readonly reasons: readonly string[];
+}
+/**
+ * Which instruments a run uses — the S6.2 ablation. The four named
+ * configurations in {@link ABLATION_CONFIGS} are CUMULATIVE.
+ *
+ * @internal
+ */
+export interface AtlasRunConfig {
+    /** The composition table over relation types. */
+    readonly types: boolean;
+    /** Side-condition rules: unguarded / zero divisors, squaring. */
+    readonly assumptions: boolean;
+    /** Dimensional consistency and convention agreement. */
+    readonly dimensionsAndConventions: boolean;
+    /** The regime check at the use site. */
+    readonly regimes: boolean;
+}
+/** Every instrument on: the atlas condition as scored. @internal */
+export declare const FULL_CONFIG: AtlasRunConfig;
+/** The four cumulative ablation configurations, in the plan's order. @internal */
+export declare const ABLATION_CONFIGS: ReadonlyArray<readonly [string, AtlasRunConfig]>;
+/**
+ * Run the atlas condition on one item. Pure and deterministic; never throws.
+ *
+ * @internal
+ */
+export declare function runAtlasOnItem(item: BenchmarkItem, config?: AtlasRunConfig): AtlasVerdict;
+/** Run the atlas condition over a set of items, in order. @internal */
+export declare function runAtlasCondition(items: readonly BenchmarkItem[], config?: AtlasRunConfig): AtlasVerdict[];
+//# sourceMappingURL=run-atlas.d.ts.map

package/dist/atlas/benchmark/run-atlas.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"run-atlas.d.ts","sourceRoot":"","sources":["../../../src/atlas/benchmark/run-atlas.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AAOH,OAAO,KAAK,EAAE,aAAa,EAAE,WAAW,EAAE,MAAM,YAAY,CAAC;AAE7D,mEAAmE;AACnE,MAAM,WAAW,YAAY;IAC3B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,OAAO,EAAE,QAAQ,GAAG,QAAQ,GAAG,SAAS,CAAC;IAClD,2EAA2E;IAC3E,QAAQ,CAAC,eAAe,CAAC,EAAE,WAAW,CAAC;IACvC,qFAAqF;IACrF,QAAQ,CAAC,OAAO,EAAE,SAAS,MAAM,EAAE,CAAC;CACrC;AAED;;;;;GAKG;AACH,MAAM,WAAW,cAAc;IAC7B,iDAAiD;IACjD,QAAQ,CAAC,KAAK,EAAE,OAAO,CAAC;IACxB,iEAAiE;IACjE,QAAQ,CAAC,WAAW,EAAE,OAAO,CAAC;IAC9B,wDAAwD;IACxD,QAAQ,CAAC,wBAAwB,EAAE,OAAO,CAAC;IAC3C,wCAAwC;IACxC,QAAQ,CAAC,OAAO,EAAE,OAAO,CAAC;CAC3B;AAED,oEAAoE;AACpE,eAAO,MAAM,WAAW,EAAE,cAKzB,CAAC;AAEF,kFAAkF;AAClF,eAAO,MAAM,gBAAgB,EAAE,aAAa,CAAC,SAAS,CAAC,MAAM,EAAE,cAAc,CAAC,CAK7E,CAAC;AA6CF;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,IAAI,EAAE,aAAa,EAAE,MAAM,GAAE,cAA4B,GAAG,YAAY,CAoHtG;AAED,uEAAuE;AACvE,wBAAgB,iBAAiB,CAC/B,KAAK,EAAE,SAAS,aAAa,EAAE,EAC/B,MAAM,GAAE,cAA4B,GACnC,YAAY,EAAE,CAEhB"}

package/dist/atlas/benchmark/run-atlas.js

@@ -0,0 +1,210 @@
+/**
+ * Atlas Phase 5, S5.2 — the ATLAS condition of the invalid-bridge benchmark.
+ *
+ * Design note: `docs/planning/Atlas-Phase-5-Design.md` §6.
+ *
+ * For each item the runner applies three atlas instruments — the applicability
+ * checker (S4.1), the composition table (S3) and the regime check (S2) — and
+ * emits `accept`, `reject` or `abstain`.
+ *
+ * ## The decision rule, and why accept is the HARDEST outcome to reach
+ *
+ * - **reject** when an instrument DEMONSTRABLY fires: a blocking applicability
+ *   finding, a chain OVERCLAIMED as an exact equivalence, a structural analogy
+ *   promoted to an equivalence, or a regime inequality checked and violated.
+ * - **accept** only when EVERY ENABLED instrument RAN and CLEARED: side conditions were
+ *   stated, the applicability checker found nothing, the regime was supplied and
+ *   every inequality was checked and held, and any claimed chain agrees with the
+ *   table. At least one instrument must have RUN: an item that no enabled
+ *   instrument applies to abstains, because nothing cleared it. S4.1 already fixed the principle: an empty finding list means "no rule
+ *   fired", which is weaker than "valid". An item that leaves a check unrun has
+ *   not been cleared by it.
+ * - **abstain** otherwise. Abstention is a first-class outcome and is preferred
+ *   to a wrong accept (ROADMAP Phase 5, pre-stated criteria).
+ *
+ * The failure kind a rejection names is a MAPPING from the rule that fired, and
+ * it is only as good as that mapping: a dimensional inconsistency is reported as
+ * `notation-collision` because that is its commonest cause, not because the rule
+ * can see notation. The scorer grades "rejected" and "named the right kind"
+ * separately, so an imperfect mapping costs the second score only.
+ *
+ * @module atlas/benchmark/run-atlas
+ * @internal
+ */
+import { checkApplicability } from '../applicability.js';
+import { composeRelation, NO_COMPOSITE_CLAIM } from '../composition-table.js';
+import { regimeHolds } from '../regime.js';
+/** Every instrument on: the atlas condition as scored. @internal */
+export const FULL_CONFIG = {
+    types: true,
+    assumptions: true,
+    dimensionsAndConventions: true,
+    regimes: true,
+};
+/** The four cumulative ablation configurations, in the plan's order. @internal */
+export const ABLATION_CONFIGS = [
+    ['types only', { types: true, assumptions: false, dimensionsAndConventions: false, regimes: false }],
+    ['+ assumptions', { types: true, assumptions: true, dimensionsAndConventions: false, regimes: false }],
+    ['+ dimensions & conventions', { types: true, assumptions: true, dimensionsAndConventions: true, regimes: false }],
+    ['+ regimes', FULL_CONFIG],
+];
+/** The instrument each applicability finding belongs to. */
+const FINDING_INSTRUMENT = {
+    'dimensional-inconsistency': 'dimensionsAndConventions',
+    'convention-mismatch': 'dimensionsAndConventions',
+    'convention-undeclared': 'dimensionsAndConventions',
+    'division-unguarded': 'assumptions',
+    'division-by-zero': 'assumptions',
+    'squaring-adds-solutions': 'assumptions',
+    'model-incompatibility': 'models',
+};
+/** Which failure kind a BLOCKING applicability finding points to. */
+const FINDING_TO_FAILURE = {
+    'dimensional-inconsistency': 'notation-collision',
+    'convention-mismatch': 'convention-mismatch',
+    'division-by-zero': 'domain-violation',
+};
+/** Relations that assert more than an analogy can carry. */
+const STRONG = new Set(['exact-equivalence', 'derivation']);
+/**
+ * The failure kind an OVERCLAIMED chain points to: a chain claimed as an exact
+ * equivalence that the table composes to something weaker.
+ *
+ * - through a restriction ⇒ the restriction's condition was dropped:
+ *   `omitted-premise`;
+ * - to a derivation or a coarse-graining ⇒ the claim asserts an inverse the
+ *   one-way step does not have: `false-inverse`.
+ */
+function overclaimFailure(chain) {
+    return chain.includes('restriction') ? 'omitted-premise' : 'false-inverse';
+}
+/**
+ * Whether `claimed` follows from `composed`. The table's own non-declined
+ * results are exact-equivalence, derivation, restriction and coarse-graining;
+ * of these only exact-equivalence ⇒ derivation is an implication this module
+ * asserts. Anything else unequal is "not shown", never "shown".
+ */
+const implies = (composed, claimed) => composed === claimed || (composed === 'exact-equivalence' && claimed === 'derivation');
+/**
+ * Run the atlas condition on one item. Pure and deterministic; never throws.
+ *
+ * @internal
+ */
+export function runAtlasOnItem(item, config = FULL_CONFIG) {
+    const reasons = [];
+    let rejectKind;
+    let rejected = false;
+    const reject = (reason, kind) => {
+        reasons.push(`REJECT: ${reason}`);
+        if (!rejected)
+            rejectKind = kind;
+        rejected = true;
+    };
+    let allRanAndCleared = true;
+    // An accept needs at least one instrument that actually RAN. Without this,
+    // an item no enabled instrument applies to is accepted with nothing checked.
+    let ranAny = false;
+    const unrun = (reason) => {
+        reasons.push(`UNRUN: ${reason}`);
+        allRanAndCleared = false;
+    };
+    // 1. Applicability — split by instrument so the ablation can switch each off.
+    const applicabilityOn = config.assumptions || config.dimensionsAndConventions;
+    if (!applicabilityOn) {
+        // Instrument disabled: neither run nor required.
+    }
+    else if (item.sideConditions === undefined) {
+        unrun('no side conditions stated, so the applicability checker has nothing to check against');
+    }
+    else {
+        ranAny = true;
+        const findings = checkApplicability({
+            ast: item.expr,
+            sideConditions: item.sideConditions,
+            ...(item.conventions?.premise === undefined ? {} : { premiseConventions: item.conventions.premise }),
+            ...(item.conventions?.conclusion === undefined
+                ? {}
+                : { conclusionConventions: item.conventions.conclusion }),
+        });
+        for (const f of findings) {
+            const instrument = FINDING_INSTRUMENT[f.kind];
+            if (instrument === 'assumptions' && !config.assumptions)
+                continue;
+            if (instrument === 'dimensionsAndConventions' && !config.dimensionsAndConventions)
+                continue;
+            if (f.severity === 'blocking')
+                reject(`${f.kind}: ${f.detail}`, FINDING_TO_FAILURE[f.kind]);
+            else {
+                reasons.push(`QUESTION: ${f.kind}: ${f.detail}`);
+                allRanAndCleared = false;
+            }
+        }
+    }
+    // 2. Composition. Only for items that claim a chain.
+    if (config.types && item.composedFrom !== undefined) {
+        ranAny = true;
+        const [first, second] = item.composedFrom;
+        const composed = composeRelation(first, second);
+        if (item.composedFrom.includes('structural-analogy') && STRONG.has(item.claimedRelation)) {
+            reject(`a chain through structural-analogy is claimed as ${item.claimedRelation}`, 'analogy-promoted');
+        }
+        else if (composed === NO_COMPOSITE_CLAIM) {
+            reasons.push(`QUESTION: the table declines to compose ${first} ∘ ${second}`);
+            allRanAndCleared = false;
+        }
+        else if (implies(composed, item.claimedRelation)) {
+            reasons.push(`CLEARED: ${first} ∘ ${second} composes to ${composed}, which gives ${item.claimedRelation}`);
+        }
+        else if (item.claimedRelation === 'exact-equivalence') {
+            reject(`${first} ∘ ${second} composes only to ${composed}; the exact equivalence is overclaimed`, overclaimFailure(item.composedFrom));
+        }
+        else {
+            reasons.push(`QUESTION: ${first} ∘ ${second} composes to ${composed}; the table does not show ${item.claimedRelation}`);
+            allRanAndCleared = false;
+        }
+    }
+    // 3. Regime.
+    if (!config.regimes) {
+        // Instrument disabled: neither run nor required.
+    }
+    else if (item.regime === undefined) {
+        unrun('no regime supplied, so the use site was never checked against a domain');
+    }
+    else {
+        ranAny = true;
+        const check = regimeHolds({ family: item.family, inequalities: item.regime.inequalities, groupDefinitions: {} }, item.regime.values);
+        if (check.ok === false) {
+            reject(`regime violated: ${check.violated.map((i) => `${i.group} ${i.op} ${i.bound}`).join(', ')}`, 'domain-violation');
+        }
+        else if (check.ok === 'unknown') {
+            unrun(`regime unchecked for ${check.unchecked.map((i) => i.group).join(', ')}`);
+        }
+        else if (item.regime.inequalities.length === 0) {
+            // regimeHolds returns true for an empty list; that is "nothing to check",
+            // not "checked and held" — the silent pass RegimeCheck warns about.
+            unrun('the regime states no inequality, so nothing was checked');
+        }
+    }
+    if (rejected) {
+        return {
+            itemId: item.id,
+            outcome: 'reject',
+            ...(rejectKind === undefined ? {} : { detectedFailure: rejectKind }),
+            reasons,
+        };
+    }
+    if (allRanAndCleared && !ranAny) {
+        reasons.push('UNRUN: no enabled instrument applies to this item, so nothing was checked');
+        return { itemId: item.id, outcome: 'abstain', reasons };
+    }
+    if (allRanAndCleared) {
+        reasons.push('ACCEPT: every instrument ran and cleared');
+        return { itemId: item.id, outcome: 'accept', reasons };
+    }
+    return { itemId: item.id, outcome: 'abstain', reasons };
+}
+/** Run the atlas condition over a set of items, in order. @internal */
+export function runAtlasCondition(items, config = FULL_CONFIG) {
+    return items.map((item) => runAtlasOnItem(item, config));
+}
+//# sourceMappingURL=run-atlas.js.map

package/dist/atlas/benchmark/run-atlas.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"run-atlas.js","sourceRoot":"","sources":["../../../src/atlas/benchmark/run-atlas.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AAEH,OAAO,EAAE,kBAAkB,EAAE,MAAM,qBAAqB,CAAC;AAEzD,OAAO,EAAE,eAAe,EAAE,kBAAkB,EAAE,MAAM,yBAAyB,CAAC;AAC9E,OAAO,EAAE,WAAW,EAAE,MAAM,cAAc,CAAC;AA+B3C,oEAAoE;AACpE,MAAM,CAAC,MAAM,WAAW,GAAmB;IACzC,KAAK,EAAE,IAAI;IACX,WAAW,EAAE,IAAI;IACjB,wBAAwB,EAAE,IAAI;IAC9B,OAAO,EAAE,IAAI;CACd,CAAC;AAEF,kFAAkF;AAClF,MAAM,CAAC,MAAM,gBAAgB,GAAqD;IAChF,CAAC,YAAY,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE,WAAW,EAAE,KAAK,EAAE,wBAAwB,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;IACpG,CAAC,eAAe,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,wBAAwB,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;IACtG,CAAC,4BAA4B,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,wBAAwB,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;IAClH,CAAC,WAAW,EAAE,WAAW,CAAC;CAC3B,CAAC;AAEF,4DAA4D;AAC5D,MAAM,kBAAkB,GAAsG;IAC5H,2BAA2B,EAAE,0BAA0B;IACvD,qBAAqB,EAAE,0BAA0B;IACjD,uBAAuB,EAAE,0BAA0B;IACnD,oBAAoB,EAAE,aAAa;IACnC,kBAAkB,EAAE,aAAa;IACjC,yBAAyB,EAAE,aAAa;IACxC,uBAAuB,EAAE,QAAQ;CAClC,CAAC;AAEF,qEAAqE;AACrE,MAAM,kBAAkB,GAAqE;IAC3F,2BAA2B,EAAE,oBAAoB;IACjD,qBAAqB,EAAE,qBAAqB;IAC5C,kBAAkB,EAAE,kBAAkB;CACvC,CAAC;AAEF,4DAA4D;AAC5D,MAAM,MAAM,GAA8B,IAAI,GAAG,CAAC,CAAC,mBAAmB,EAAE,YAAY,CAAC,CAAC,CAAC;AAEvF;;;;;;;;GAQG;AACH,SAAS,gBAAgB,CAAC,KAA8B;IACtD,OAAO,KAAK,CAAC,QAAQ,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC,CAAC,eAAe,CAAC;AAC7E,CAAC;AAED;;;;;GAKG;AACH,MAAM,OAAO,GAAG,CAAC,QAAsB,EAAE,OAAqB,EAAW,EAAE,CACzE,QAAQ,KAAK,OAAO,IAAI,CAAC,QAAQ,KAAK,mBAAmB,IAAI,OAAO,KAAK,YAAY,CAAC,CAAC;AAEzF;;;;GAIG;AACH,MAAM,UAAU,cAAc,CAAC,IAAmB,EAAE,MAAM,GAAmB,WAAW;IACtF,MAAM,OAAO,GAAa,EAAE,CAAC;IAC7B,IAAI,UAAmC,CAAC;IACxC,IAAI,QAAQ,GAAG,KAAK,CAAC;IACrB,MAAM,MAAM,GAAG,CAAC,MAAc,EAAE,IAAkB,EAAQ,EAAE;QAC1D,OAAO,CAAC,IAAI,CAAC,WAAW,MAAM,EAAE,CAAC,CAAC;QAClC,IAAI,CAAC,QAAQ;YAAE,UAAU,GAAG,IAAI,CAAC;QACjC,QAAQ,GAAG,IAAI,CAAC;IAClB,CAAC,CAAC;IACF,IAAI,gBAAgB,GAAG,IAAI,CAAC;IAC5B,2EAA2E;IAC3E,6EAA6E;IAC7E,IAAI,MAAM,GAAG,KAAK,CAAC;IACnB,MAAM,KAAK,GAAG,CAAC,MAAc,EAAQ,EAAE;QACrC,OAAO,CAAC,IAAI,CAAC,UAAU,MAAM,EAAE,CAAC,CAAC;QACjC,gBAAgB,GAAG,KAAK,CAAC;IAC3B,CAAC,CAAC;IAEF,8EAA8E;IAC9E,MAAM,eAAe,GAAG,MAAM,CAAC,WAAW,IAAI,MAAM,CAAC,wBAAwB,CAAC;IAC9E,IAAI,CAAC,eAAe,EAAE,CAAC;QACrB,iDAAiD;IACnD,CAAC;SAAM,IAAI,IAAI,CAAC,cAAc,KAAK,SAAS,EAAE,CAAC;QAC7C,KAAK,CAAC,sFAAsF,CAAC,CAAC;IAChG,CAAC;SAAM,CAAC;QACN,MAAM,GAAG,IAAI,CAAC;QACd,MAAM,QAAQ,GAAG,kBAAkB,CAAC;YAClC,GAAG,EAAE,IAAI,CAAC,IAAI;YACd,cAAc,EAAE,IAAI,CAAC,cAAc;YACnC,GAAG,CAAC,IAAI,CAAC,WAAW,EAAE,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,kBAAkB,EAAE,IAAI,CAAC,WAAW,CAAC,OAAO,EAAE,CAAC;YACpG,GAAG,CAAC,IAAI,CAAC,WAAW,EAAE,UAAU,KAAK,SAAS;gBAC5C,CAAC,CAAC,EAAE;gBACJ,CAAC,CAAC,EAAE,qBAAqB,EAAE,IAAI,CAAC,WAAW,CAAC,UAAU,EAAE,CAAC;SAC5D,CAAC,CAAC;QACH,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;YACzB,MAAM,UAAU,GAAG,kBAAkB,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;YAC9C,IAAI,UAAU,KAAK,aAAa,IAAI,CAAC,MAAM,CAAC,WAAW;gBAAE,SAAS;YAClE,IAAI,UAAU,KAAK,0BAA0B,IAAI,CAAC,MAAM,CAAC,wBAAwB;gBAAE,SAAS;YAC5F,IAAI,CAAC,CAAC,QAAQ,KAAK,UAAU;gBAAE,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,MAAM,EAAE,EAAE,kBAAkB,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;iBACvF,CAAC;gBACJ,OAAO,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC;gBACjD,gBAAgB,GAAG,KAAK,CAAC;YAC3B,CAAC;QACH,CAAC;IACH,CAAC;IAED,qDAAqD;IACrD,IAAI,MAAM,CAAC,KAAK,IAAI,IAAI,CAAC,YAAY,KAAK,SAAS,EAAE,CAAC;QACpD,MAAM,GAAG,IAAI,CAAC;QACd,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,GAAG,IAAI,CAAC,YAAY,CAAC;QAC1C,MAAM,QAAQ,GAAG,eAAe,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QAChD,IAAI,IAAI,CAAC,YAAY,CAAC,QAAQ,CAAC,oBAAoB,CAAC,IAAI,MAAM,CAAC,GAAG,CAAC,IAAI,CAAC,eAAe,CAAC,EAAE,CAAC;YACzF,MAAM,CACJ,oDAAoD,IAAI,CAAC,eAAe,EAAE,EAC1E,kBAAkB,CACnB,CAAC;QACJ,CAAC;aAAM,IAAI,QAAQ,KAAK,kBAAkB,EAAE,CAAC;YAC3C,OAAO,CAAC,IAAI,CAAC,2CAA2C,KAAK,MAAM,MAAM,EAAE,CAAC,CAAC;YAC7E,gBAAgB,GAAG,KAAK,CAAC;QAC3B,CAAC;aAAM,IAAI,OAAO,CAAC,QAAQ,EAAE,IAAI,CAAC,eAAe,CAAC,EAAE,CAAC;YACnD,OAAO,CAAC,IAAI,CAAC,YAAY,KAAK,MAAM,MAAM,gBAAgB,QAAQ,iBAAiB,IAAI,CAAC,eAAe,EAAE,CAAC,CAAC;QAC7G,CAAC;aAAM,IAAI,IAAI,CAAC,eAAe,KAAK,mBAAmB,EAAE,CAAC;YACxD,MAAM,CACJ,GAAG,KAAK,MAAM,MAAM,qBAAqB,QAAQ,wCAAwC,EACzF,gBAAgB,CAAC,IAAI,CAAC,YAAY,CAAC,CACpC,CAAC;QACJ,CAAC;aAAM,CAAC;YACN,OAAO,CAAC,IAAI,CACV,aAAa,KAAK,MAAM,MAAM,gBAAgB,QAAQ,6BAA6B,IAAI,CAAC,eAAe,EAAE,CAC1G,CAAC;YACF,gBAAgB,GAAG,KAAK,CAAC;QAC3B,CAAC;IACH,CAAC;IAED,aAAa;IACb,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;QACpB,iDAAiD;IACnD,CAAC;SAAM,IAAI,IAAI,CAAC,MAAM,KAAK,SAAS,EAAE,CAAC;QACrC,KAAK,CAAC,wEAAwE,CAAC,CAAC;IAClF,CAAC;SAAM,CAAC;QACN,MAAM,GAAG,IAAI,CAAC;QACd,MAAM,KAAK,GAAG,WAAW,CACvB,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,YAAY,EAAE,IAAI,CAAC,MAAM,CAAC,YAAY,EAAE,gBAAgB,EAAE,EAAE,EAAE,EACrF,IAAI,CAAC,MAAM,CAAC,MAAM,CACnB,CAAC;QACF,IAAI,KAAK,CAAC,EAAE,KAAK,KAAK,EAAE,CAAC;YACvB,MAAM,CACJ,oBAAoB,KAAK,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,KAAK,IAAI,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,EAC3F,kBAAkB,CACnB,CAAC;QACJ,CAAC;aAAM,IAAI,KAAK,CAAC,EAAE,KAAK,SAAS,EAAE,CAAC;YAClC,KAAK,CAAC,wBAAwB,KAAK,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QAClF,CAAC;aAAM,IAAI,IAAI,CAAC,MAAM,CAAC,YAAY,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACjD,0EAA0E;YAC1E,oEAAoE;YACpE,KAAK,CAAC,yDAAyD,CAAC,CAAC;QACnE,CAAC;IACH,CAAC;IAED,IAAI,QAAQ,EAAE,CAAC;QACb,OAAO;YACL,MAAM,EAAE,IAAI,CAAC,EAAE;YACf,OAAO,EAAE,QAAQ;YACjB,GAAG,CAAC,UAAU,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,eAAe,EAAE,UAAU,EAAE,CAAC;YACpE,OAAO;SACR,CAAC;IACJ,CAAC;IACD,IAAI,gBAAgB,IAAI,CAAC,MAAM,EAAE,CAAC;QAChC,OAAO,CAAC,IAAI,CAAC,2EAA2E,CAAC,CAAC;QAC1F,OAAO,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,OAAO,EAAE,SAAS,EAAE,OAAO,EAAE,CAAC;IAC1D,CAAC;IACD,IAAI,gBAAgB,EAAE,CAAC;QACrB,OAAO,CAAC,IAAI,CAAC,0CAA0C,CAAC,CAAC;QACzD,OAAO,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE,CAAC;IACzD,CAAC;IACD,OAAO,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,OAAO,EAAE,SAAS,EAAE,OAAO,EAAE,CAAC;AAC1D,CAAC;AAED,uEAAuE;AACvE,MAAM,UAAU,iBAAiB,CAC/B,KAA+B,EAC/B,MAAM,GAAmB,WAAW;IAEpC,OAAO,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,cAAc,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC,CAAC;AAC3D,CAAC"}

package/dist/atlas/benchmark/stats.d.ts

@@ -0,0 +1,90 @@
+/**
+ * Atlas Phase 5, S5.4 — the benchmark's statistics and power report.
+ *
+ * Design note: `docs/planning/Atlas-Phase-5-Design.md` §8.
+ *
+ * Every function here is tested against a value computed independently, not
+ * against its own output: the Wilson intervals against the plan's textbook
+ * values, McNemar against a hand-computed 2×2 table, and κ against a hand-computed
+ * confusion matrix. The paired-difference interval (Newcombe's method 10) has no
+ * textbook value in the repository; it is tested by properties that pin it to
+ * Wilson and to the unpaired square-and-add interval where those must agree.
+ *
+ * @module atlas/benchmark/stats
+ * @internal
+ */
+/** The two-sided 95% normal quantile. @internal */
+export declare const Z95 = 1.959963984540054;
+/** A closed interval. @internal */
+export interface Interval {
+    readonly lower: number;
+    readonly upper: number;
+}
+/**
+ * Wilson score interval for a binomial proportion.
+ *
+ * @throws Error when `n` is not a positive integer or `successes` is outside `[0, n]`.
+ * @internal
+ */
+export declare function wilsonInterval(successes: number, n: number, z?: number): Interval;
+/** A paired 2×2 table: a = both right, b = only A right, c = only B right, d = both wrong. @internal */
+export interface PairedTable {
+    readonly a: number;
+    readonly b: number;
+    readonly c: number;
+    readonly d: number;
+}
+/** McNemar's test on the discordant cells. @internal */
+export interface McNemarResult {
+    /** Continuity-corrected χ² = (|b − c| − 1)² / (b + c); 0 when b + c = 0. */
+    readonly chiSquared: number;
+    /** Exact two-sided p-value from Binomial(b + c, ½); 1 when b + c = 0. */
+    readonly exactP: number;
+}
+/**
+ * McNemar's paired test. The EXACT binomial p-value is the one to report at
+ * these sample sizes; the χ² is given because readers expect it.
+ *
+ * @internal
+ */
+export declare function mcnemar(table: PairedTable): McNemarResult;
+/**
+ * Newcombe's method 10: a 95% interval for the PAIRED difference in proportions
+ * p_A − p_B = (b − c)/n, built from the two marginal Wilson intervals and a
+ * correlation correction φ. Pre-registered criterion: the interval for the
+ * atlas-minus-baseline difference must exclude zero.
+ *
+ * Newcombe, Statistics in Medicine 17 (1998) 2635.
+ *
+ * @internal
+ */
+export declare function pairedDifferenceInterval(table: PairedTable, z?: number): Interval & {
+    readonly diff: number;
+};
+/**
+ * Cohen's κ for two raters over the same items, from a square confusion matrix
+ * (`matrix[i][j]` = items rater 1 put in category i and rater 2 in category j).
+ *
+ * @returns NaN when chance agreement is 1 (κ is undefined there), never 1.
+ * @internal
+ */
+export declare function cohensKappa(matrix: readonly (readonly number[])[]): number;
+/** What a frozen set of a given size can and cannot resolve. @internal */
+export interface PowerReport {
+    readonly nPerClass: number;
+    /** The assumed per-class accuracy the interval is computed at. */
+    readonly assumedAccuracy: number;
+    readonly interval: Interval;
+    /** Half the interval width: differences smaller than about twice this are not resolvable. */
+    readonly halfWidth: number;
+    readonly statement: string;
+}
+/**
+ * The interval a frozen set of `nPerClass` items per class can afford at an
+ * assumed accuracy. This is the "report the interval it can afford rather than
+ * the claim it cannot" line of ROADMAP Phase 5.
+ *
+ * @internal
+ */
+export declare function powerReport(nPerClass: number, assumedAccuracy?: number): PowerReport;
+//# sourceMappingURL=stats.d.ts.map

package/dist/atlas/benchmark/stats.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"stats.d.ts","sourceRoot":"","sources":["../../../src/atlas/benchmark/stats.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,mDAAmD;AACnD,eAAO,MAAM,GAAG,oBAAoB,CAAC;AAErC,mCAAmC;AACnC,MAAM,WAAW,QAAQ;IACvB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB;AAED;;;;;GAKG;AACH,wBAAgB,cAAc,CAAC,SAAS,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,SAAM,GAAG,QAAQ,CAgB9E;AAED,wGAAwG;AACxG,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,CAAC,EAAE,MAAM,CAAC;IACnB,QAAQ,CAAC,CAAC,EAAE,MAAM,CAAC;IACnB,QAAQ,CAAC,CAAC,EAAE,MAAM,CAAC;IACnB,QAAQ,CAAC,CAAC,EAAE,MAAM,CAAC;CACpB;AAED,wDAAwD;AACxD,MAAM,WAAW,aAAa;IAC5B,4EAA4E;IAC5E,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,yEAAyE;IACzE,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AASD;;;;;GAKG;AACH,wBAAgB,OAAO,CAAC,KAAK,EAAE,WAAW,GAAG,aAAa,CAQzD;AAED;;;;;;;;;GASG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,WAAW,EAAE,CAAC,SAAM,GAAG,QAAQ,GAAG;IAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAuB1G;AAED;;;;;;GAMG;AACH,wBAAgB,WAAW,CAAC,MAAM,EAAE,SAAS,CAAC,SAAS,MAAM,EAAE,CAAC,EAAE,GAAG,MAAM,CAsB1E;AAED,0EAA0E;AAC1E,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,kEAAkE;IAClE,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,QAAQ,EAAE,QAAQ,CAAC;IAC5B,6FAA6F;IAC7F,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;CAC5B;AAED;;;;;;GAMG;AACH,wBAAgB,WAAW,CAAC,SAAS,EAAE,MAAM,EAAE,eAAe,SAAM,GAAG,WAAW,CAgBjF"}