@hviana/sema 0.4.2 → 0.4.3

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

@@ -0,0 +1,300 @@
+// 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;
+        }
+        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

@@ -0,0 +1,16 @@
+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;