universal-physics-tensor 0.45.2 → 0.46.0

package/dist/atlas/benchmark/stats.js

@@ -0,0 +1,159 @@
+/**
+ * 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 const Z95 = 1.959963984540054;
+/**
+ * Wilson score interval for a binomial proportion.
+ *
+ * @throws Error when `n` is not a positive integer or `successes` is outside `[0, n]`.
+ * @internal
+ */
+export function wilsonInterval(successes, n, z = Z95) {
+    if (!Number.isInteger(n) || n <= 0)
+        throw new Error(`wilsonInterval: n must be a positive integer, got ${n}`);
+    if (!Number.isInteger(successes) || successes < 0 || successes > n) {
+        throw new Error(`wilsonInterval: successes must be an integer in [0, n], got ${successes}`);
+    }
+    const p = successes / n;
+    const z2 = z * z;
+    const denom = 1 + z2 / n;
+    const centre = (p + z2 / (2 * n)) / denom;
+    const half = (z / denom) * Math.sqrt((p * (1 - p)) / n + z2 / (4 * n * n));
+    // At x = 0 the lower bound, and at x = n the upper bound, are EXACTLY 0 and 1
+    // analytically; the formula reaches them only to within float rounding.
+    return {
+        lower: successes === 0 ? 0 : Math.max(0, centre - half),
+        upper: successes === n ? 1 : Math.min(1, centre + half),
+    };
+}
+/** `ln C(n, k)`, summed; exact enough for the n here. */
+function logChoose(n, k) {
+    let s = 0;
+    for (let i = 1; i <= k; i++)
+        s += Math.log(n - k + i) - Math.log(i);
+    return s;
+}
+/**
+ * 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 function mcnemar(table) {
+    const n = table.b + table.c;
+    if (n === 0)
+        return { chiSquared: 0, exactP: 1 };
+    const chiSquared = (Math.abs(table.b - table.c) - 1) ** 2 / n;
+    const k = Math.min(table.b, table.c);
+    let tail = 0;
+    for (let i = 0; i <= k; i++)
+        tail += Math.exp(logChoose(n, i) - n * Math.LN2);
+    return { chiSquared, exactP: Math.min(1, 2 * tail) };
+}
+/**
+ * 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 function pairedDifferenceInterval(table, z = Z95) {
+    const { a, b, c, d } = table;
+    const n = a + b + c + d;
+    if (n === 0)
+        throw new Error('pairedDifferenceInterval: empty table');
+    const p1 = (a + b) / n;
+    const p2 = (a + c) / n;
+    const w1 = wilsonInterval(a + b, n, z);
+    const w2 = wilsonInterval(a + c, n, z);
+    const denom = Math.sqrt((a + b) * (c + d) * (a + c) * (b + d));
+    let phi = 0;
+    if (denom > 0) {
+        const cross = a * d - b * c;
+        // Newcombe's continuity adjustment of φ, applied only when ad > bc.
+        phi = (cross > 0 ? Math.max(cross - n / 2, 0) : cross) / denom;
+    }
+    const diff = p1 - p2;
+    const dl1 = p1 - w1.lower;
+    const du1 = w1.upper - p1;
+    const dl2 = p2 - w2.lower;
+    const du2 = w2.upper - p2;
+    const lower = diff - Math.sqrt(Math.max(0, dl1 * dl1 - 2 * phi * dl1 * du2 + du2 * du2));
+    const upper = diff + Math.sqrt(Math.max(0, du1 * du1 - 2 * phi * du1 * dl2 + dl2 * dl2));
+    return { diff, lower: Math.max(-1, lower), upper: Math.min(1, upper) };
+}
+/**
+ * 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 function cohensKappa(matrix) {
+    const k = matrix.length;
+    if (k === 0 || matrix.some((row) => row.length !== k))
+        throw new Error('cohensKappa: matrix must be square and non-empty');
+    let n = 0;
+    let agree = 0;
+    const rows = new Array(k).fill(0);
+    const cols = new Array(k).fill(0);
+    for (let i = 0; i < k; i++) {
+        for (let j = 0; j < k; j++) {
+            const v = matrix[i][j];
+            n += v;
+            rows[i] += v;
+            cols[j] += v;
+            if (i === j)
+                agree += v;
+        }
+    }
+    if (n === 0)
+        throw new Error('cohensKappa: no ratings');
+    const po = agree / n;
+    let pe = 0;
+    for (let i = 0; i < k; i++)
+        pe += (rows[i] / n) * (cols[i] / n);
+    if (pe === 1)
+        return Number.NaN;
+    return (po - pe) / (1 - pe);
+}
+/**
+ * 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 function powerReport(nPerClass, assumedAccuracy = 0.8) {
+    const successes = Math.round(assumedAccuracy * nPerClass);
+    const interval = wilsonInterval(successes, nPerClass);
+    const halfWidth = (interval.upper - interval.lower) / 2;
+    const pct = (x) => `${(100 * x).toFixed(1)}%`;
+    return {
+        nPerClass,
+        assumedAccuracy,
+        interval,
+        halfWidth,
+        statement: `With ${nPerClass} items per class at ${pct(assumedAccuracy)} accuracy, the 95% Wilson interval ` +
+            `is [${pct(interval.lower)}, ${pct(interval.upper)}] (±${pct(halfWidth)}). Two methods whose ` +
+            `accuracies differ by less than about ${pct(2 * halfWidth)} cannot be separated on this set ` +
+            'by unpaired intervals; the paired McNemar design is what narrows it.',
+    };
+}
+//# sourceMappingURL=stats.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"stats.js","sourceRoot":"","sources":["../../../src/atlas/benchmark/stats.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,mDAAmD;AACnD,MAAM,CAAC,MAAM,GAAG,GAAG,iBAAiB,CAAC;AAQrC;;;;;GAKG;AACH,MAAM,UAAU,cAAc,CAAC,SAAiB,EAAE,CAAS,EAAE,CAAC,GAAG,GAAG;IAClE,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,qDAAqD,CAAC,EAAE,CAAC,CAAC;IAC9G,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,SAAS,CAAC,IAAI,SAAS,GAAG,CAAC,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;QACnE,MAAM,IAAI,KAAK,CAAC,+DAA+D,SAAS,EAAE,CAAC,CAAC;IAC9F,CAAC;IACD,MAAM,CAAC,GAAG,SAAS,GAAG,CAAC,CAAC;IACxB,MAAM,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC;IACjB,MAAM,KAAK,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC;IACzB,MAAM,MAAM,GAAG,CAAC,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;IAC1C,MAAM,IAAI,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;IAC3E,8EAA8E;IAC9E,wEAAwE;IACxE,OAAO;QACL,KAAK,EAAE,SAAS,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACvD,KAAK,EAAE,SAAS,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;KACxD,CAAC;AACJ,CAAC;AAkBD,yDAAyD;AACzD,SAAS,SAAS,CAAC,CAAS,EAAE,CAAS;IACrC,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE;QAAE,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;IACpE,OAAO,CAAC,CAAC;AACX,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,OAAO,CAAC,KAAkB;IACxC,MAAM,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC;IAC5B,IAAI,CAAC,KAAK,CAAC;QAAE,OAAO,EAAE,UAAU,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;IACjD,MAAM,UAAU,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IAC9D,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC;IACrC,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE;QAAE,IAAI,IAAI,IAAI,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC;IAC9E,OAAO,EAAE,UAAU,EAAE,MAAM,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,EAAE,CAAC;AACvD,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,UAAU,wBAAwB,CAAC,KAAkB,EAAE,CAAC,GAAG,GAAG;IAClE,MAAM,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,GAAG,KAAK,CAAC;IAC7B,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IACxB,IAAI,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC,CAAC;IACtE,MAAM,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC;IACvB,MAAM,EAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC;IACvB,MAAM,EAAE,GAAG,cAAc,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;IACvC,MAAM,EAAE,GAAG,cAAc,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;IACvC,MAAM,KAAK,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;IAC/D,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;QACd,MAAM,KAAK,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAC5B,oEAAoE;QACpE,GAAG,GAAG,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,GAAG,KAAK,CAAC;IACjE,CAAC;IACD,MAAM,IAAI,GAAG,EAAE,GAAG,EAAE,CAAC;IACrB,MAAM,GAAG,GAAG,EAAE,GAAG,EAAE,CAAC,KAAK,CAAC;IAC1B,MAAM,GAAG,GAAG,EAAE,CAAC,KAAK,GAAG,EAAE,CAAC;IAC1B,MAAM,GAAG,GAAG,EAAE,GAAG,EAAE,CAAC,KAAK,CAAC;IAC1B,MAAM,GAAG,GAAG,EAAE,CAAC,KAAK,GAAG,EAAE,CAAC;IAC1B,MAAM,KAAK,GAAG,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,GAAG,GAAG,GAAG,GAAG,CAAC,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,CAAC,CAAC,CAAC;IACzF,MAAM,KAAK,GAAG,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,GAAG,GAAG,GAAG,GAAG,CAAC,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,GAAG,CAAC,CAAC,CAAC;IACzF,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,KAAK,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,CAAC;AACzE,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,WAAW,CAAC,MAAsC;IAChE,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC;IACxB,IAAI,CAAC,KAAK,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,MAAM,KAAK,CAAC,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,kDAAkD,CAAC,CAAC;IAC3H,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,MAAM,IAAI,GAAG,IAAI,KAAK,CAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAC1C,MAAM,IAAI,GAAG,IAAI,KAAK,CAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAC1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC3B,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAE,CAAC,CAAC,CAAE,CAAC;YACzB,CAAC,IAAI,CAAC,CAAC;YACP,IAAI,CAAC,CAAC,CAAE,IAAI,CAAC,CAAC;YACd,IAAI,CAAC,CAAC,CAAE,IAAI,CAAC,CAAC;YACd,IAAI,CAAC,KAAK,CAAC;gBAAE,KAAK,IAAI,CAAC,CAAC;QAC1B,CAAC;IACH,CAAC;IACD,IAAI,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,yBAAyB,CAAC,CAAC;IACxD,MAAM,EAAE,GAAG,KAAK,GAAG,CAAC,CAAC;IACrB,IAAI,EAAE,GAAG,CAAC,CAAC;IACX,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC;IAClE,IAAI,EAAE,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC,GAAG,CAAC;IAChC,OAAO,CAAC,EAAE,GAAG,EAAE,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC;AAC9B,CAAC;AAaD;;;;;;GAMG;AACH,MAAM,UAAU,WAAW,CAAC,SAAiB,EAAE,eAAe,GAAG,GAAG;IAClE,MAAM,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,eAAe,GAAG,SAAS,CAAC,CAAC;IAC1D,MAAM,QAAQ,GAAG,cAAc,CAAC,SAAS,EAAE,SAAS,CAAC,CAAC;IACtD,MAAM,SAAS,GAAG,CAAC,QAAQ,CAAC,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IACxD,MAAM,GAAG,GAAG,CAAC,CAAS,EAAU,EAAE,CAAC,GAAG,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC;IAC9D,OAAO;QACL,SAAS;QACT,eAAe;QACf,QAAQ;QACR,SAAS;QACT,SAAS,EACP,QAAQ,SAAS,uBAAuB,GAAG,CAAC,eAAe,CAAC,qCAAqC;YACjG,OAAO,GAAG,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,GAAG,CAAC,QAAQ,CAAC,KAAK,CAAC,OAAO,GAAG,CAAC,SAAS,CAAC,uBAAuB;YAC9F,wCAAwC,GAAG,CAAC,CAAC,GAAG,SAAS,CAAC,mCAAmC;YAC7F,sEAAsE;KACzE,CAAC;AACJ,CAAC"}

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

@@ -0,0 +1,120 @@
+/**
+ * Atlas Phase 6, S6.1 — scoring a benchmark condition against the answer key.
+ *
+ * Design note: `docs/planning/Atlas-Phase-5-Design.md` §9.
+ *
+ * This module receives the answer key as an ARGUMENT. It never reads the scorer
+ * half of the fixture tree itself: the import guard forbids any `src/` path into
+ * it. The orchestration script (`scripts/run-atlas-study.mjs`) loads the labels
+ * and passes them in.
+ *
+ * **An empty item set produces NO metrics, not zero metrics.** `scoreCondition`
+ * throws on it: a table of zeros and NaNs from an empty set is indistinguishable,
+ * at a glance, from a condition that failed everything.
+ *
+ * @module atlas/benchmark/study
+ * @internal
+ */
+import type { FailureKind } from './types.js';
+import type { Interval, McNemarResult } from './stats.js';
+/** One condition's answer for one item. @internal */
+export interface ConditionAnswer {
+    readonly itemId: string;
+    readonly outcome: 'accept' | 'reject' | 'abstain';
+    readonly detectedFailure?: FailureKind;
+}
+/** The answer key for one item — scorer-side, never in the public half. @internal */
+export interface ItemLabel {
+    readonly itemId: string;
+    readonly kind: 'valid' | 'invalid';
+    /** For an invalid item: the failure kind a correct rejection names. */
+    readonly failureKind?: FailureKind;
+}
+/** One reason an answer key does not fit its item set. @internal */
+export interface LabelProblem {
+    readonly id: string;
+    readonly problem: string;
+}
+/**
+ * Check an answer key against the item ids it must cover: exactly one label per
+ * item, no label for an unknown item, a known failure kind on every invalid
+ * label and none on a valid one. Takes the key as an ARGUMENT, like everything
+ * here, because no `src/` file may read the scorer half.
+ *
+ * @param itemIds - the ids of the frozen items.
+ * @param labels - the answer key.
+ * @returns every problem found, labels first in key order, then missing labels
+ * in item order; empty means the key fits.
+ * @internal
+ */
+export declare function validateLabels(itemIds: readonly string[], labels: readonly ItemLabel[]): LabelProblem[];
+/** A condition's metrics on the frozen set. @internal */
+export interface ConditionMetrics {
+    readonly condition: string;
+    readonly nValid: number;
+    readonly nInvalid: number;
+    /** Invalid items rejected, with its Wilson interval. */
+    readonly invalidRejected: number;
+    readonly invalidRejectionInterval: Interval;
+    /** Of the rejected invalid items, how many named the right failure kind. */
+    readonly kindNamedCorrectly: number;
+    /** Invalid items ACCEPTED: the error the benchmark exists to count. */
+    readonly wrongAccepts: number;
+    /** Valid items accepted. */
+    readonly validAccepted: number;
+    /** Valid items rejected. */
+    readonly falseRejects: number;
+    /** All abstentions, reported and never folded into accuracy. */
+    readonly abstentions: number;
+    /** Items the condition did not answer at all: counted, never dropped. */
+    readonly unanswered: number;
+}
+/**
+ * Score one condition against the answer key.
+ *
+ * @throws Error when the key is empty (no measurement exists), or when an answer
+ * names an item the key does not contain.
+ * @internal
+ */
+export declare function scoreCondition(condition: string, answers: readonly ConditionAnswer[], labels: readonly ItemLabel[]): ConditionMetrics;
+/** The paired comparison of two conditions' rejection of invalid items. @internal */
+export interface PairedRejection {
+    readonly conditionA: string;
+    readonly conditionB: string;
+    readonly table: {
+        readonly a: number;
+        readonly b: number;
+        readonly c: number;
+        readonly d: number;
+    };
+    readonly difference: Interval & {
+        readonly diff: number;
+    };
+    readonly mcnemar: McNemarResult;
+    /** The pre-registered criterion: the interval for A − B excludes zero, above it. */
+    readonly aBetterExcludingZero: boolean;
+}
+/**
+ * Compare two conditions on the SAME invalid items. An item counts as "right"
+ * for a condition when it was rejected; an unanswered item is "wrong", not dropped.
+ *
+ * @throws Error when there are no invalid items.
+ * @internal
+ */
+export declare function pairedRejection(conditionA: string, answersA: readonly ConditionAnswer[], conditionB: string, answersB: readonly ConditionAnswer[], labels: readonly ItemLabel[]): PairedRejection;
+/** One row of the ablation: a configuration's metrics and its step over the previous one. @internal */
+export interface AblationRow {
+    readonly configuration: string;
+    readonly metrics: ConditionMetrics;
+    /** Paired rejection of THIS configuration vs the previous one; absent for the first. */
+    readonly stepOverPrevious?: PairedRejection;
+}
+/**
+ * Score a sequence of named condition runs as an ablation: each row is scored,
+ * and each row after the first is compared, paired on the same invalid items,
+ * against the row before it — so the table shows what each added layer bought.
+ *
+ * @internal
+ */
+export declare function scoreAblation(runs: ReadonlyArray<readonly [string, readonly ConditionAnswer[]]>, labels: readonly ItemLabel[]): AblationRow[];
+//# sourceMappingURL=study.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"study.d.ts","sourceRoot":"","sources":["../../../src/atlas/benchmark/study.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAGH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,YAAY,CAAC;AAE9C,OAAO,KAAK,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAE1D,qDAAqD;AACrD,MAAM,WAAW,eAAe;IAC9B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,OAAO,EAAE,QAAQ,GAAG,QAAQ,GAAG,SAAS,CAAC;IAClD,QAAQ,CAAC,eAAe,CAAC,EAAE,WAAW,CAAC;CACxC;AAED,qFAAqF;AACrF,MAAM,WAAW,SAAS;IACxB,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,IAAI,EAAE,OAAO,GAAG,SAAS,CAAC;IACnC,uEAAuE;IACvE,QAAQ,CAAC,WAAW,CAAC,EAAE,WAAW,CAAC;CACpC;AAED,oEAAoE;AACpE,MAAM,WAAW,YAAY;IAC3B,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAID;;;;;;;;;;;GAWG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,SAAS,MAAM,EAAE,EAAE,MAAM,EAAE,SAAS,SAAS,EAAE,GAAG,YAAY,EAAE,CAoBvG;AAED,yDAAyD;AACzD,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,wDAAwD;IACxD,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,wBAAwB,EAAE,QAAQ,CAAC;IAC5C,4EAA4E;IAC5E,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,uEAAuE;IACvE,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,4BAA4B;IAC5B,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,4BAA4B;IAC5B,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,gEAAgE;IAChE,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;IAC7B,yEAAyE;IACzE,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;CAC7B;AAED;;;;;;GAMG;AACH,wBAAgB,cAAc,CAC5B,SAAS,EAAE,MAAM,EACjB,OAAO,EAAE,SAAS,eAAe,EAAE,EACnC,MAAM,EAAE,SAAS,SAAS,EAAE,GAC3B,gBAAgB,CAmDlB;AAED,qFAAqF;AACrF,MAAM,WAAW,eAAe;IAC9B,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,KAAK,EAAE;QAAE,QAAQ,CAAC,CAAC,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,CAAC,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,CAAC,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IACnG,QAAQ,CAAC,UAAU,EAAE,QAAQ,GAAG;QAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC;IAC1D,QAAQ,CAAC,OAAO,EAAE,aAAa,CAAC;IAChC,oFAAoF;IACpF,QAAQ,CAAC,oBAAoB,EAAE,OAAO,CAAC;CACxC;AAED;;;;;;GAMG;AACH,wBAAgB,eAAe,CAC7B,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,SAAS,eAAe,EAAE,EACpC,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,SAAS,eAAe,EAAE,EACpC,MAAM,EAAE,SAAS,SAAS,EAAE,GAC3B,eAAe,CA6BjB;AAED,uGAAuG;AACvG,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,OAAO,EAAE,gBAAgB,CAAC;IACnC,wFAAwF;IACxF,QAAQ,CAAC,gBAAgB,CAAC,EAAE,eAAe,CAAC;CAC7C;AAED;;;;;;GAMG;AACH,wBAAgB,aAAa,CAC3B,IAAI,EAAE,aAAa,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,eAAe,EAAE,CAAC,CAAC,EAClE,MAAM,EAAE,SAAS,SAAS,EAAE,GAC3B,WAAW,EAAE,CAgBf"}

package/dist/atlas/benchmark/study.js

@@ -0,0 +1,192 @@
+/**
+ * Atlas Phase 6, S6.1 — scoring a benchmark condition against the answer key.
+ *
+ * Design note: `docs/planning/Atlas-Phase-5-Design.md` §9.
+ *
+ * This module receives the answer key as an ARGUMENT. It never reads the scorer
+ * half of the fixture tree itself: the import guard forbids any `src/` path into
+ * it. The orchestration script (`scripts/run-atlas-study.mjs`) loads the labels
+ * and passes them in.
+ *
+ * **An empty item set produces NO metrics, not zero metrics.** `scoreCondition`
+ * throws on it: a table of zeros and NaNs from an empty set is indistinguishable,
+ * at a glance, from a condition that failed everything.
+ *
+ * @module atlas/benchmark/study
+ * @internal
+ */
+import { FAILURE_KINDS } from './types.js';
+import { mcnemar, pairedDifferenceInterval, wilsonInterval } from './stats.js';
+const KNOWN_FAILURES = new Set(FAILURE_KINDS);
+/**
+ * Check an answer key against the item ids it must cover: exactly one label per
+ * item, no label for an unknown item, a known failure kind on every invalid
+ * label and none on a valid one. Takes the key as an ARGUMENT, like everything
+ * here, because no `src/` file may read the scorer half.
+ *
+ * @param itemIds - the ids of the frozen items.
+ * @param labels - the answer key.
+ * @returns every problem found, labels first in key order, then missing labels
+ * in item order; empty means the key fits.
+ * @internal
+ */
+export function validateLabels(itemIds, labels) {
+    const problems = [];
+    const ids = new Set(itemIds);
+    const seen = new Set();
+    for (const l of labels) {
+        const p = (problem) => {
+            problems.push({ id: l.itemId, problem });
+        };
+        if (seen.has(l.itemId))
+            p('duplicate label');
+        seen.add(l.itemId);
+        if (!ids.has(l.itemId))
+            p('label for an item not in the set');
+        if (l.kind === 'invalid') {
+            if (l.failureKind === undefined)
+                p('an invalid label must name its failureKind');
+            else if (!KNOWN_FAILURES.has(l.failureKind))
+                p(`unknown failureKind '${l.failureKind}'`);
+        }
+        else if (l.failureKind !== undefined) {
+            p('a valid label must not carry a failureKind');
+        }
+    }
+    for (const id of itemIds)
+        if (!seen.has(id))
+            problems.push({ id, problem: 'no label' });
+    return problems;
+}
+/**
+ * Score one condition against the answer key.
+ *
+ * @throws Error when the key is empty (no measurement exists), or when an answer
+ * names an item the key does not contain.
+ * @internal
+ */
+export function scoreCondition(condition, answers, labels) {
+    if (labels.length === 0) {
+        throw new Error(`scoreCondition(${condition}): the answer key is empty — there is nothing to measure`);
+    }
+    const byId = new Map(answers.map((a) => [a.itemId, a]));
+    const labelIds = new Set(labels.map((l) => l.itemId));
+    for (const a of answers) {
+        if (!labelIds.has(a.itemId))
+            throw new Error(`scoreCondition(${condition}): answer for unknown item ${a.itemId}`);
+    }
+    let nValid = 0;
+    let nInvalid = 0;
+    let invalidRejected = 0;
+    let kindNamedCorrectly = 0;
+    let wrongAccepts = 0;
+    let validAccepted = 0;
+    let falseRejects = 0;
+    let abstentions = 0;
+    let unanswered = 0;
+    for (const label of labels) {
+        const answer = byId.get(label.itemId);
+        if (answer === undefined)
+            unanswered++;
+        else if (answer.outcome === 'abstain')
+            abstentions++;
+        if (label.kind === 'valid') {
+            nValid++;
+            if (answer?.outcome === 'accept')
+                validAccepted++;
+            if (answer?.outcome === 'reject')
+                falseRejects++;
+        }
+        else {
+            nInvalid++;
+            if (answer?.outcome === 'accept')
+                wrongAccepts++;
+            if (answer?.outcome === 'reject') {
+                invalidRejected++;
+                if (answer.detectedFailure !== undefined && answer.detectedFailure === label.failureKind) {
+                    kindNamedCorrectly++;
+                }
+            }
+        }
+    }
+    return {
+        condition,
+        nValid,
+        nInvalid,
+        invalidRejected,
+        invalidRejectionInterval: nInvalid === 0 ? { lower: Number.NaN, upper: Number.NaN } : wilsonInterval(invalidRejected, nInvalid),
+        kindNamedCorrectly,
+        wrongAccepts,
+        validAccepted,
+        falseRejects,
+        abstentions,
+        unanswered,
+    };
+}
+/**
+ * Compare two conditions on the SAME invalid items. An item counts as "right"
+ * for a condition when it was rejected; an unanswered item is "wrong", not dropped.
+ *
+ * @throws Error when there are no invalid items.
+ * @internal
+ */
+export function pairedRejection(conditionA, answersA, conditionB, answersB, labels) {
+    const invalid = labels.filter((l) => l.kind === 'invalid');
+    if (invalid.length === 0)
+        throw new Error('pairedRejection: no invalid items to compare on');
+    const rejectedBy = (answers) => new Set(answers.filter((x) => x.outcome === 'reject').map((x) => x.itemId));
+    const ra = rejectedBy(answersA);
+    const rb = rejectedBy(answersB);
+    let a = 0;
+    let b = 0;
+    let c = 0;
+    let d = 0;
+    for (const l of invalid) {
+        const x = ra.has(l.itemId);
+        const y = rb.has(l.itemId);
+        if (x && y)
+            a++;
+        else if (x)
+            b++;
+        else if (y)
+            c++;
+        else
+            d++;
+    }
+    const table = { a, b, c, d };
+    const difference = pairedDifferenceInterval(table);
+    return {
+        conditionA,
+        conditionB,
+        table,
+        difference,
+        mcnemar: mcnemar(table),
+        aBetterExcludingZero: difference.lower > 0,
+    };
+}
+/**
+ * Score a sequence of named condition runs as an ablation: each row is scored,
+ * and each row after the first is compared, paired on the same invalid items,
+ * against the row before it — so the table shows what each added layer bought.
+ *
+ * @internal
+ */
+export function scoreAblation(runs, labels) {
+    const rows = [];
+    for (let i = 0; i < runs.length; i++) {
+        const [name, answers] = runs[i];
+        const metrics = scoreCondition(name, answers, labels);
+        if (i === 0)
+            rows.push({ configuration: name, metrics });
+        else {
+            const [prevName, prevAnswers] = runs[i - 1];
+            rows.push({
+                configuration: name,
+                metrics,
+                stepOverPrevious: pairedRejection(name, answers, prevName, prevAnswers, labels),
+            });
+        }
+    }
+    return rows;
+}
+//# sourceMappingURL=study.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"study.js","sourceRoot":"","sources":["../../../src/atlas/benchmark/study.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAE3C,OAAO,EAAE,OAAO,EAAE,wBAAwB,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAwB/E,MAAM,cAAc,GAAG,IAAI,GAAG,CAAS,aAAa,CAAC,CAAC;AAEtD;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,cAAc,CAAC,OAA0B,EAAE,MAA4B;IACrF,MAAM,QAAQ,GAAmB,EAAE,CAAC;IACpC,MAAM,GAAG,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,CAAC;IAC7B,MAAM,IAAI,GAAG,IAAI,GAAG,EAAU,CAAC;IAC/B,KAAK,MAAM,CAAC,IAAI,MAAM,EAAE,CAAC;QACvB,MAAM,CAAC,GAAG,CAAC,OAAe,EAAQ,EAAE;YAClC,QAAQ,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,MAAM,EAAE,OAAO,EAAE,CAAC,CAAC;QAC3C,CAAC,CAAC;QACF,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC;YAAE,CAAC,CAAC,iBAAiB,CAAC,CAAC;QAC7C,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QACnB,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC;YAAE,CAAC,CAAC,kCAAkC,CAAC,CAAC;QAC9D,IAAI,CAAC,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;YACzB,IAAI,CAAC,CAAC,WAAW,KAAK,SAAS;gBAAE,CAAC,CAAC,4CAA4C,CAAC,CAAC;iBAC5E,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC;gBAAE,CAAC,CAAC,wBAAwB,CAAC,CAAC,WAAW,GAAG,CAAC,CAAC;QAC3F,CAAC;aAAM,IAAI,CAAC,CAAC,WAAW,KAAK,SAAS,EAAE,CAAC;YACvC,CAAC,CAAC,4CAA4C,CAAC,CAAC;QAClD,CAAC;IACH,CAAC;IACD,KAAK,MAAM,EAAE,IAAI,OAAO;QAAE,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;YAAE,QAAQ,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,OAAO,EAAE,UAAU,EAAE,CAAC,CAAC;IACxF,OAAO,QAAQ,CAAC;AAClB,CAAC;AAwBD;;;;;;GAMG;AACH,MAAM,UAAU,cAAc,CAC5B,SAAiB,EACjB,OAAmC,EACnC,MAA4B;IAE5B,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACxB,MAAM,IAAI,KAAK,CAAC,kBAAkB,SAAS,0DAA0D,CAAC,CAAC;IACzG,CAAC;IACD,MAAM,IAAI,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IACxD,MAAM,QAAQ,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;IACtD,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACxB,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,kBAAkB,SAAS,8BAA8B,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC;IACpH,CAAC;IACD,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,IAAI,QAAQ,GAAG,CAAC,CAAC;IACjB,IAAI,eAAe,GAAG,CAAC,CAAC;IACxB,IAAI,kBAAkB,GAAG,CAAC,CAAC;IAC3B,IAAI,YAAY,GAAG,CAAC,CAAC;IACrB,IAAI,aAAa,GAAG,CAAC,CAAC;IACtB,IAAI,YAAY,GAAG,CAAC,CAAC;IACrB,IAAI,WAAW,GAAG,CAAC,CAAC;IACpB,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QAC3B,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;QACtC,IAAI,MAAM,KAAK,SAAS;YAAE,UAAU,EAAE,CAAC;aAClC,IAAI,MAAM,CAAC,OAAO,KAAK,SAAS;YAAE,WAAW,EAAE,CAAC;QACrD,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;YAC3B,MAAM,EAAE,CAAC;YACT,IAAI,MAAM,EAAE,OAAO,KAAK,QAAQ;gBAAE,aAAa,EAAE,CAAC;YAClD,IAAI,MAAM,EAAE,OAAO,KAAK,QAAQ;gBAAE,YAAY,EAAE,CAAC;QACnD,CAAC;aAAM,CAAC;YACN,QAAQ,EAAE,CAAC;YACX,IAAI,MAAM,EAAE,OAAO,KAAK,QAAQ;gBAAE,YAAY,EAAE,CAAC;YACjD,IAAI,MAAM,EAAE,OAAO,KAAK,QAAQ,EAAE,CAAC;gBACjC,eAAe,EAAE,CAAC;gBAClB,IAAI,MAAM,CAAC,eAAe,KAAK,SAAS,IAAI,MAAM,CAAC,eAAe,KAAK,KAAK,CAAC,WAAW,EAAE,CAAC;oBACzF,kBAAkB,EAAE,CAAC;gBACvB,CAAC;YACH,CAAC;QACH,CAAC;IACH,CAAC;IACD,OAAO;QACL,SAAS;QACT,MAAM;QACN,QAAQ;QACR,eAAe;QACf,wBAAwB,EACtB,QAAQ,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,cAAc,CAAC,eAAe,EAAE,QAAQ,CAAC;QACvG,kBAAkB;QAClB,YAAY;QACZ,aAAa;QACb,YAAY;QACZ,WAAW;QACX,UAAU;KACX,CAAC;AACJ,CAAC;AAaD;;;;;;GAMG;AACH,MAAM,UAAU,eAAe,CAC7B,UAAkB,EAClB,QAAoC,EACpC,UAAkB,EAClB,QAAoC,EACpC,MAA4B;IAE5B,MAAM,OAAO,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC;IAC3D,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,iDAAiD,CAAC,CAAC;IAC7F,MAAM,UAAU,GAAG,CAAC,OAAmC,EAAe,EAAE,CACtE,IAAI,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;IAC9E,MAAM,EAAE,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC;IAChC,MAAM,EAAE,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC;IAChC,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACxB,MAAM,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QAC3B,MAAM,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QAC3B,IAAI,CAAC,IAAI,CAAC;YAAE,CAAC,EAAE,CAAC;aACX,IAAI,CAAC;YAAE,CAAC,EAAE,CAAC;aACX,IAAI,CAAC;YAAE,CAAC,EAAE,CAAC;;YACX,CAAC,EAAE,CAAC;IACX,CAAC;IACD,MAAM,KAAK,GAAG,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC;IAC7B,MAAM,UAAU,GAAG,wBAAwB,CAAC,KAAK,CAAC,CAAC;IACnD,OAAO;QACL,UAAU;QACV,UAAU;QACV,KAAK;QACL,UAAU;QACV,OAAO,EAAE,OAAO,CAAC,KAAK,CAAC;QACvB,oBAAoB,EAAE,UAAU,CAAC,KAAK,GAAG,CAAC;KAC3C,CAAC;AACJ,CAAC;AAUD;;;;;;GAMG;AACH,MAAM,UAAU,aAAa,CAC3B,IAAkE,EAClE,MAA4B;IAE5B,MAAM,IAAI,GAAkB,EAAE,CAAC;IAC/B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACrC,MAAM,CAAC,IAAI,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC,CAAC,CAAE,CAAC;QACjC,MAAM,OAAO,GAAG,cAAc,CAAC,IAAI,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC;QACtD,IAAI,CAAC,KAAK,CAAC;YAAE,IAAI,CAAC,IAAI,CAAC,EAAE,aAAa,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;aACpD,CAAC;YACJ,MAAM,CAAC,QAAQ,EAAE,WAAW,CAAC,GAAG,IAAI,CAAC,CAAC,GAAG,CAAC,CAAE,CAAC;YAC7C,IAAI,CAAC,IAAI,CAAC;gBACR,aAAa,EAAE,IAAI;gBACnB,OAAO;gBACP,gBAAgB,EAAE,eAAe,CAAC,IAAI,EAAE,OAAO,EAAE,QAAQ,EAAE,WAAW,EAAE,MAAM,CAAC;aAChF,CAAC,CAAC;QACL,CAAC;IACH,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC"}

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

@@ -0,0 +1,79 @@
+/**
+ * Atlas Phase 5, S5.1 — the invalid-bridge benchmark's item schema.
+ *
+ * Design note: `docs/planning/Atlas-Phase-5-Design.md` §2.
+ *
+ * ⚠ **No agent authors a frozen item** (design note §0). This file defines the
+ * SHAPE an independent author fills in; it contains no items.
+ *
+ * @module atlas/benchmark/types
+ * @internal
+ */
+import type { ExprNode } from '../../dimensional/ast-types.js';
+import type { Conventions, RegimeInequality, RelationType } from '../types.js';
+/** The eight failure kinds, sampled evenly (ROADMAP Phase 5). @internal */
+export declare const FAILURE_KINDS: readonly ['omitted-premise', 'domain-violation', 'convention-mismatch', 'notation-collision', 'dimensional-coincidence', 'non-uniform-limit', 'false-inverse', 'analogy-promoted'];
+/** One failure kind. @internal */
+export type FailureKind = (typeof FAILURE_KINDS)[number];
+/**
+ * Who wrote an item. Only `'independent'` may be frozen; `'contested-draft'`
+ * is what an agent may produce, into `contested/`, and nothing else.
+ *
+ * @internal
+ */
+export type Authorship = 'independent' | 'contested-draft';
+/** Which split an item scores in. @internal */
+export type BenchmarkSplit = 'in-distribution' | 'held-out';
+/**
+ * The held-out family. Corrected from the plan's "first-order relaxation",
+ * which Phase 0 already encodes as `model-first-order` (design note §4).
+ *
+ * @internal
+ */
+export declare const HELD_OUT_FAMILY = "fluid-statics";
+/**
+ * Markers that identify the held-out family in model ids, dynamics and bridge
+ * ids. `tests/atlas/benchmark.test.ts` scans the atlas for them, and proves the
+ * scan can fire with the ORIGINAL family's markers as a positive control.
+ *
+ * @internal
+ */
+export declare const HELD_OUT_MARKERS: readonly string[];
+/**
+ * One benchmark item, as its PUBLIC half carries it — never the answer. Whether
+ * the item is valid, and which failure kind it has, live only in the scorer
+ * half (`ItemLabel` in `study.ts`), so a condition that reads the public file
+ * cannot read the key.
+ *
+ * @internal
+ */
+export interface BenchmarkItem {
+    readonly id: string;
+    readonly premises: readonly string[];
+    readonly conclusion: string;
+    readonly claimedRelation: RelationType;
+    readonly family: string;
+    readonly split: BenchmarkSplit;
+    /** The claimed relation as an AST, so the leakage check can run. */
+    readonly expr: ExprNode;
+    /** The id of the item this one renames, when it is a renamed-variable variant. */
+    readonly renamedVariant?: string;
+    readonly authorship: Authorship;
+    /** Where the item came from: an erratum, a documented misconception, a textbook. */
+    readonly source: string;
+    /** The claim's stated side conditions, as prose. `[]` states "none". */
+    readonly sideConditions?: readonly string[];
+    /** Conventions declared by the premises and by the conclusion. */
+    readonly conventions?: {
+        readonly premise?: Conventions;
+        readonly conclusion?: Conventions;
+    };
+    /** When the claim is a CHAIN: the two relations composed to reach `claimedRelation`. */
+    readonly composedFrom?: readonly [RelationType, RelationType];
+    /** The regime the claim is used in: its inequalities and the π-group values at the use site. */
+    readonly regime?: {
+        readonly inequalities: readonly RegimeInequality[];
+        readonly values: Readonly<Record<string, number>>;
+    };
+}
+//# sourceMappingURL=types.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../../src/atlas/benchmark/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,gCAAgC,CAAC;AAC/D,OAAO,KAAK,EAAE,WAAW,EAAE,gBAAgB,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAE/E,2EAA2E;AAC3E,eAAO,MAAM,aAAa,YACxB,iBAAiB,EACjB,kBAAkB,EAClB,qBAAqB,EACrB,oBAAoB,EACpB,yBAAyB,EACzB,mBAAmB,EACnB,eAAe,EACf,kBAAkB,CACV,CAAC;AAEX,kCAAkC;AAClC,MAAM,MAAM,WAAW,GAAG,CAAC,OAAO,aAAa,CAAC,CAAC,MAAM,CAAC,CAAC;AAEzD;;;;;GAKG;AACH,MAAM,MAAM,UAAU,GAAG,aAAa,GAAG,iBAAiB,CAAC;AAE3D,+CAA+C;AAC/C,MAAM,MAAM,cAAc,GAAG,iBAAiB,GAAG,UAAU,CAAC;AAE5D;;;;;GAKG;AACH,eAAO,MAAM,eAAe,kBAAkB,CAAC;AAE/C;;;;;;GAMG;AACH,eAAO,MAAM,gBAAgB,EAAE,SAAS,MAAM,EAQ7C,CAAC;AAEF;;;;;;;GAOG;AACH,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,QAAQ,EAAE,SAAS,MAAM,EAAE,CAAC;IACrC,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,eAAe,EAAE,YAAY,CAAC;IACvC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,cAAc,CAAC;IAC/B,oEAAoE;IACpE,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC;IACxB,kFAAkF;IAClF,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,UAAU,EAAE,UAAU,CAAC;IAChC,oFAAoF;IACpF,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IAMxB,wEAAwE;IACxE,QAAQ,CAAC,cAAc,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IAC5C,kEAAkE;IAClE,QAAQ,CAAC,WAAW,CAAC,EAAE;QACrB,QAAQ,CAAC,OAAO,CAAC,EAAE,WAAW,CAAC;QAC/B,QAAQ,CAAC,UAAU,CAAC,EAAE,WAAW,CAAC;KACnC,CAAC;IACF,wFAAwF;IACxF,QAAQ,CAAC,YAAY,CAAC,EAAE,SAAS,CAAC,YAAY,EAAE,YAAY,CAAC,CAAC;IAC9D,gGAAgG;IAChG,QAAQ,CAAC,MAAM,CAAC,EAAE;QAChB,QAAQ,CAAC,YAAY,EAAE,SAAS,gBAAgB,EAAE,CAAC;QACnD,QAAQ,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAAC;KACnD,CAAC;CACH"}

package/dist/atlas/benchmark/types.js

@@ -0,0 +1,46 @@
+/**
+ * Atlas Phase 5, S5.1 — the invalid-bridge benchmark's item schema.
+ *
+ * Design note: `docs/planning/Atlas-Phase-5-Design.md` §2.
+ *
+ * ⚠ **No agent authors a frozen item** (design note §0). This file defines the
+ * SHAPE an independent author fills in; it contains no items.
+ *
+ * @module atlas/benchmark/types
+ * @internal
+ */
+/** The eight failure kinds, sampled evenly (ROADMAP Phase 5). @internal */
+export const FAILURE_KINDS = [
+    'omitted-premise',
+    'domain-violation',
+    'convention-mismatch',
+    'notation-collision',
+    'dimensional-coincidence',
+    'non-uniform-limit',
+    'false-inverse',
+    'analogy-promoted',
+];
+/**
+ * The held-out family. Corrected from the plan's "first-order relaxation",
+ * which Phase 0 already encodes as `model-first-order` (design note §4).
+ *
+ * @internal
+ */
+export const HELD_OUT_FAMILY = 'fluid-statics';
+/**
+ * Markers that identify the held-out family in model ids, dynamics and bridge
+ * ids. `tests/atlas/benchmark.test.ts` scans the atlas for them, and proves the
+ * scan can fire with the ORIGINAL family's markers as a positive control.
+ *
+ * @internal
+ */
+export const HELD_OUT_MARKERS = [
+    'hydrostat',
+    'buoyan',
+    'archimed',
+    'pascal',
+    'barometric',
+    'fluid-static',
+    'fluid statics',
+];
+//# sourceMappingURL=types.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../../../src/atlas/benchmark/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAKH,2EAA2E;AAC3E,MAAM,CAAC,MAAM,aAAa,GAAG;IAC3B,iBAAiB;IACjB,kBAAkB;IAClB,qBAAqB;IACrB,oBAAoB;IACpB,yBAAyB;IACzB,mBAAmB;IACnB,eAAe;IACf,kBAAkB;CACV,CAAC;AAgBX;;;;;GAKG;AACH,MAAM,CAAC,MAAM,eAAe,GAAG,eAAe,CAAC;AAE/C;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAsB;IACjD,WAAW;IACX,QAAQ;IACR,UAAU;IACV,QAAQ;IACR,YAAY;IACZ,cAAc;IACd,eAAe;CAChB,CAAC"}