@hviana/sema 0.5.2 → 0.5.4

package/dist/src/mind/mechanisms/extraction.js

@@ -1,327 +0,0 @@
-// extraction.ts — Extraction (Skill) — Section 4 of the mind.
-//
-// Given a query and its consensus-ranked anchors, find the first span-shaped
-// skill exemplar among the ranked anchors and read the analogous span of the
-// query.  A skill exemplar is a learnt fact whose context and answer together
-// form a span-in-context pattern: the answer is a subsequence of the context
-// (or one of its pieces is), and the context is the smallest spanning frame
-// that contains it.
-import { locate } from "../match.js";
-import { concatBytes, indexOf } from "../../bytes.js";
-import { decodeText, unexplainedLabel } from "../rationale.js";
-import { CONCEPT, STEP } from "../graph-search.js";
-import { rItem, rNode, traceFail } from "../trace.js";
-// ── Extraction ────────────────────────────────────────────────────────────
-/** Find the first span-shaped skill exemplar among the ranked anchors from
- *  climbAttentionAll and read the analogous span from the query.  Returns
- *  the extracted bytes PLUS the query spans the skill ACCOUNTED FOR — the
- *  located frames AND any read span BOUNDED by located frames on both
- *  sides, the elementary evidence think's grounding decider weighs.  A
- *  bounded read is explained: the skill located both its borders in the
- *  query and emitted exactly what sits between them.  An OPEN-ENDED read
- *  (the exemplar's answer reaches the context's end, so the query is read
- *  to its own end with no located right border) remains a guess about where
- *  the span stops — it stays unaccounted, priced by exclusion like the
- *  cover's bridged bytes.  (Accounting frames only — the earlier convention
- *  — let a CAST juxtaposition that merely echoed the query's exact site
- *  outweigh a correct bounded extraction: the same span counted as
- *  explained for one mechanism and not the other, and the asymmetry, not
- *  the answers' merits, decided the grounding.)  Null when no skill
- *  applies. */
-export async function extractBySkill(ctx, query, pre) {
-    const t = ctx.trace?.enter("extractBySkill", [
-        rItem(query, "query"),
-    ]);
-    const fail = traceFail(t);
-    // Use climbAttentionAll to get the FULL ranked list, not just the
-    // roots that cleared commitVotes' significance floor.  The floor
-    // gates further points of attention for fusion, but extraction only
-    // needs ONE anchor that IS a span-shaped skill exemplar — and on
-    // some seeds the top-voted anchor is not one (e.g. a concept-merge
-    // nickname outvotes the painting exemplars on shared substrings,
-    // while the exemplars' votes fall below the floor).  Iterating the
-    // ranked list instead of just the roots lets extraction reach the
-    // first painting-exemplar anchor regardless of its floor status.
-    //
-    const { ranked } = await pre.attention();
-    if (ranked.length === 0) {
-        return fail("no consensus anchor — no skill to apply");
-    }
-    // Try ranked anchors IN ORDER until one yields a USABLE extraction — not
-    // merely a span-shaped exemplar, but one whose extracted span clears the
-    // same one-river-fold quantum (W) cover.ts's restatedSpan gate already
-    // treats as the floor below which byte overlap is chance, not evidence.
-    // isSpanShaped (spanShapedOf) is a deliberately permissive sparse-
-    // subsequence check — see the section note below — so it accepts exemplars
-    // whose relation to the query is coincidental gap-matching, not genuine
-    // structure.  Stopping at the FIRST such exemplar let a coincidental match
-    // early in the ranked list win outright and read out a sub-quantum
-    // fragment (observed: a 3-byte "Hel" pulled from an unrelated exemplar,
-    // while a later ranked anchor would have read the query's own "Hello…"
-    // correctly).  Trying further anchors when one produces nothing usable is
-    // the same idiom this loop already uses for non-exemplars — extended to
-    // cover a bad extraction, not just a structural non-match.
-    //
-    // The retry is bounded at pre.k — the SAME evidence-breadth constant every
-    // other consumer of a ranked list already self-limits to (resonance, the
-    // weave, the climb itself; see Precomputed.k's own doc comment) — not the
-    // full ranked list.  locate()'s frame match has an EXACT-byte tier with no
-    // significance correction of its own (short W-byte frames are cheap to
-    // match by pure chance), so trying every ranked anchor turns that per-
-    // anchor chance into a near-certainty over enough attempts: on a pure-
-    // gibberish query, 170 anchors deep found an unrelated Zulu exemplar whose
-    // short frame happened to byte-match, producing "xyzzy pl" — a coincidence
-    // no different in kind from the "RaBitQ estimate overshot the reach bar
-    // and grounded pure gibberish" failure recall.ts's own significance
-    // correction exists to prevent (see recallByResonance's reach-threshold
-    // comment).  Bounding the search to the ranked list's own top-k restores
-    // the "genuinely relevant but not root-significant" exemplars this loop
-    // was built for, without the unbounded tail's chance collisions.
-    const W = ctx.space.maxGroup;
-    const searched = ranked.slice(0, pre.k);
-    let shapeMisses = 0;
-    let subQuantum = 0;
-    for (const cand of searched) {
-        const exemplar = await pre.spanShapedOf(cand.anchor);
-        if (!exemplar) {
-            shapeMisses++;
-            continue;
-        }
-        const built = buildFromExemplar(ctx, query, pre, exemplar);
-        if (built === null || built.bytes.length < W) {
-            subQuantum++;
-            continue;
-        }
-        // AN UNANCHORED READ IS NOT AN EXTRACTION.  This function's contract (see
-        // the doc above) is that `accounted` carries "the located frames AND any
-        // read span BOUNDED by located frames on both sides", while an open-ended
-        // read "remains a guess about where the span stops — it stays unaccounted".
-        // EMPTY accounted is the degenerate case of that: NO frame of the exemplar
-        // was located in the query at all, so nothing ties the bytes just read to
-        // this question — the skill applied its exemplar's geometry to a query it
-        // never matched.
-        //
-        // The live case (analyze_training.ts F, the battery's ONLY wrong non-silent
-        // answer): "Which city is France's seat of government?" answered "Which ci"
-        // — a fragment of the query itself — from the exemplar "What is dll", with
-        // accounted=[] and pieces=1.  isSpanShaped is deliberately permissive (a
-        // sparse-subsequence check), so it accepts exemplars whose relation to the
-        // query is coincidental gap-matching; requiring at least one LOCATED frame
-        // is the structural evidence that permissiveness leaves out.
-        //
-        // Scoped to extraction ON PURPOSE.  The same test at the pipeline's
-        // post-grounding density check was tried and REVERTED: `accounted` is passed
-        // empty BY CONVENTION on recall's own tiers (recall.ts's ground(…, [], …)),
-        // so a density veto there refused six legitimate reverse-recall groundings.
-        // Here the field is this mechanism's own output and carries its documented
-        // meaning, so the test is sound exactly where the convention does not reach.
-        if (built.accounted.length === 0) {
-            subQuantum++;
-            continue;
-        }
-        if (shapeMisses > 0 || subQuantum > 0) {
-            ctx.trace?.step("trySkillAnchors", [
-                rItem(query.subarray(0, 0), `skipped ${shapeMisses + subQuantum}`),
-                rNode(ctx, cand.anchor, "chosen"),
-            ], [], `skipped ${shapeMisses} non-exemplar and ${subQuantum} sub-quantum ` +
-                `anchor(s) before one yielded a usable extraction`);
-        }
-        t?.done([rItem(built.bytes, "extracted")], built.pieces === 1
-            ? `apply a learnt extraction skill — read the analogous span of the query` +
-                ` framed like "${decodeText(exemplar.answerBytes)}" sits in its exemplar`
-            : `apply a learnt MULTI-PIECE skill — read ${built.pieces} analogous` +
-                ` pieces of the query and synthesize them like "${decodeText(exemplar.answerBytes)}"`);
-        return {
-            bytes: built.bytes,
-            accounted: built.accounted,
-            unexplained: unexplainedLabel(query, built.accounted),
-        };
-    }
-    if (shapeMisses === searched.length) {
-        ctx.trace?.step("trySkillAnchors", [], [], `none of the top ${searched.length} ranked anchor(s) (of ${ranked.length} total) ` +
-            `is a span-shaped skill exemplar`);
-        return fail("no consensus root is a span-shaped skill exemplar");
-    }
-    return fail("no ranked anchor yielded an extraction at or above the quantum floor");
-}
-/** Build the extracted bytes for ONE already-accepted span-shaped exemplar —
- *  factored out of {@link extractBySkill} so its anchor loop can try
- *  successive ranked candidates instead of committing to the first
- *  structural match.  Null when the exemplar's answer does not decompose
- *  against its context, or no piece's frame locates in the query. */
-function buildFromExemplar(ctx, query, pre, exemplar) {
-    const { contextBytes, answerBytes } = exemplar;
-    const ansCtxRuns = answerRunsInContext(ctx, contextBytes, answerBytes);
-    if (ansCtxRuns === null || ansCtxRuns.length === 0) {
-        return null;
-    }
-    if (ansCtxRuns.length > 1) {
-        ctx.trace?.step("decomposeAnswer", [rItem(answerBytes, "multi-piece-answer")], ansCtxRuns.map((r) => rItem(contextBytes.subarray(r.start, r.end), "piece", undefined, [
-            r.start,
-            r.end,
-        ])), `answer splits into ${ansCtxRuns.length} piece(s) within the exemplar context`);
-    }
-    const pieces = [];
-    const accounted = [];
-    for (let ri = 0; ri < ansCtxRuns.length; ri++) {
-        const run = ansCtxRuns[ri];
-        const isLast = ri === ansCtxRuns.length - 1;
-        const framePreLen = Math.min(run.start, ctx.space.maxGroup);
-        const framePre = run.start > 0
-            ? contextBytes.subarray(run.start - framePreLen, run.start)
-            : null;
-        const frames = [];
-        let start = 0;
-        if (framePre) {
-            const prePos = locate(ctx, query, framePre, 0, pre.rec.sites);
-            if (prePos < 0)
-                continue;
-            start = prePos + framePre.length;
-            frames.push([prePos, start]); // the located frame IS matched evidence
-        }
-        let end;
-        if (isLast) {
-            if (run.end < contextBytes.length) {
-                const framePostLen = Math.min(contextBytes.length - run.end, ctx.space.maxGroup);
-                const framePost = contextBytes.subarray(run.end, run.end + framePostLen);
-                const postPos = locate(ctx, query.subarray(start), framePost, 0, pre.rec.sites);
-                if (postPos < 0)
-                    continue;
-                end = start + postPos;
-                frames.push([end, end + framePost.length]); // matched post-frame
-            }
-            else {
-                end = query.length;
-            }
-        }
-        else {
-            const nextRun = ansCtxRuns[ri + 1];
-            const nextPreLen = Math.min(nextRun.start, ctx.space.maxGroup);
-            const nextPre = contextBytes.subarray(nextRun.start - nextPreLen, nextRun.start);
-            const nextPos = locate(ctx, query.subarray(start), nextPre, 0, pre.rec.sites);
-            if (nextPos < 0) {
-                end = start + run.ansLen;
-            }
-            else {
-                end = start + nextPos;
-                frames.push([end, end + nextPre.length]); // matched next-frame
-            }
-        }
-        if (start >= end)
-            continue;
-        pieces.push(query.subarray(start, end));
-        accounted.push(...frames);
-        // Bounded on both sides ⇒ the read span itself is explained (see doc).
-        // frames carries the pre-border (when the answer is not at the context's
-        // start) and the located right border (post-frame or next piece's
-        // pre-frame); only when BOTH borders were located is the read bounded.
-        const preBounded = run.start === 0 || frames.some(([, e]) => e === start);
-        const postBounded = frames.some(([b]) => b === end);
-        if (preBounded && postBounded)
-            accounted.push([start, end]);
-    }
-    if (pieces.length === 0) {
-        return null;
-    }
-    const out = pieces.length === 1 ? pieces[0] : concatBytes(pieces);
-    return { bytes: out, accounted, pieces: pieces.length };
-}
-// ── The two span-shape readings: OPEN acceptance vs. STRONG decomposition ──
-//
-// isSpanShaped and answerRunsInContext read the SAME relation ("the answer is
-// drawn from the context") at two deliberately different strengths, and they
-// are NOT interchangeable:
-//
-//   • isSpanShaped — the OPEN reading: any in-order embedding (a sparse
-//     subsequence, arbitrary gaps).  O(|context|) byte scan.  Used to ACCEPT
-//     an exemplar candidate.
-//   • answerRunsInContext — the STRONG reading: a greedy longest-run
-//     DECOMPOSITION into contiguous pieces.  Greedy-longest is strictly
-//     stronger than subsequence (a long late match can consume context an
-//     earlier shorter choice needed), so an ACCEPTED exemplar can still fail
-//     to decompose — extractBySkill then fails with "answer is not a
-//     subsequence of the context" and think falls through to recall.  That
-//     fall-through is BEHAVIOUR, pinned by the extraction suites: do not
-//     "unify" the two into one machine — replacing the open reading with the
-//     strong one silently rejects exemplars extraction today accepts, and
-//     replacing the strong one with a backtracking embedding changes which
-//     pieces are read out of the query.
-/** Decompose an answer into substrings of its surrounding context, in order —
- *  the STRONG span-shape reading (see the section note above).  Returns null
- *  when no greedy longest-run decomposition exists.  Adjacent runs that
- *  connect contiguously are merged. */
-export function answerRunsInContext(_ctx, context, answer) {
-    const pos = indexOf(context, answer, 0);
-    if (pos >= 0) {
-        return [{ start: pos, end: pos + answer.length, ansLen: answer.length }];
-    }
-    const runs = [];
-    let ai = 0;
-    let ci = 0;
-    while (ai < answer.length) {
-        // Longest match of the remaining answer at any position of the remaining
-        // context: one direct extend per context position — O(|ctx|·match) per
-        // run, replacing the previous longest-first indexOf countdown whose
-        // repeated scans were cubic on long sparse-subsequence answers.
-        let bestLen = 0;
-        let bestPos = -1;
-        for (let p = ci; p < context.length; p++) {
-            let l = 0;
-            const maxL = Math.min(context.length - p, answer.length - ai);
-            if (maxL <= bestLen)
-                break; // no later position can beat the best
-            while (l < maxL && context[p + l] === answer[ai + l])
-                l++;
-            if (l > bestLen) {
-                bestLen = l;
-                bestPos = p;
-                if (ai + l === answer.length)
-                    break; // the whole remainder matched
-            }
-        }
-        if (bestLen === 0)
-            return null;
-        runs.push({ start: bestPos, end: bestPos + bestLen, ansLen: bestLen });
-        ai += bestLen;
-        ci = bestPos + bestLen;
-    }
-    const merged = [];
-    for (const r of runs) {
-        const last = merged[merged.length - 1];
-        if (last && r.start === last.end) {
-            last.end = r.end;
-            last.ansLen += r.ansLen;
-        }
-        else {
-            merged.push({ ...r });
-        }
-    }
-    return merged.length > 0 ? merged : null;
-}
-// ── Pipeline mechanism ──────────────────────────────────────────────────────
-export const extractionMechanism = {
-    name: "extraction",
-    provenance: "extract",
-    async floor(_ctx, _query, pre, worthRunning) {
-        // Extraction's floor is always exactly CONCEPT+STEP when it exists —
-        // same investment discipline as CAST's (see cast.ts): when the bound
-        // already cannot beat the incumbent, return it UNINVESTED (never
-        // first-touch the climb just to be pruned).
-        if (!worthRunning(CONCEPT + STEP))
-            return CONCEPT + STEP;
-        if ((await pre.attention()).ranked.length === 0)
-            return null;
-        return CONCEPT + STEP;
-    },
-    async run(ctx, query, pre) {
-        const ex = await extractBySkill(ctx, query, pre);
-        if (!ex)
-            return [];
-        return [{
-                bytes: ex.bytes,
-                accounted: ex.accounted,
-                moves: CONCEPT + STEP * ex.accounted.length,
-                unexplained: ex.unexplained,
-            }];
-    },
-};

package/dist/src/mind/mechanisms/recall.d.ts

@@ -1,16 +0,0 @@
-import type { MindContext } from "../types.js";
-import type { PipelineMechanism, Precomputed } from "../pipeline-mechanism.js";
-/** A recall result. */
-export interface RecallResult {
-    bytes: Uint8Array;
-    echoed: boolean;
-    accounted: Array<[number, number]>;
-    moves: number;
-    unexplained: string;
-    /** See {@link import("../pipeline-mechanism.js").MechanismResult.complete}
-     *  — set by the IDENTITY-bridge tier alone. */
-    complete?: boolean;
-}
-/** Recall the answer by resonating the whole query against the content index. */
-export declare function recallByResonance(ctx: MindContext, query: Uint8Array, pre: Precomputed): Promise<RecallResult | null>;
-export declare const recallMechanism: PipelineMechanism;