@hviana/sema 0.2.3 → 0.2.4

package/src/mind/mechanisms/cast.ts

@@ -12,8 +12,9 @@
 // juxtapose.
 
 import type { MindContext } from "../types.js";
+import type { Vec } from "../../vec.js";
 import { read } from "../primitives.js";
-import { argmaxBy, hubBound } from "../traverse.js";
+import { argmaxBy, corpusN, hubBound } from "../traverse.js";
 import {
   analogyStrength,
   follow,
@@ -22,11 +23,17 @@ import {
   reverseContext,
 } from "../match.js";
 import { joinWithBridge } from "../resonance.js";
+import { restatesQuery } from "../reasoning.js";
 import { CONCEPT, STEP } from "../graph-search.js";
 import { concat2, indexOf } from "../../bytes.js";
-import { dominates } from "../../geometry.js";
-import { decodeText, unexplainedLabel } from "../rationale.js";
+import { consensusFloor, dominates } from "../../geometry.js";
+import {
+  decodeText,
+  unexplainedLabel,
+  unexplainedSpans,
+} from "../rationale.js";
 import { rItem, rNode } from "../trace.js";
+import { dismissedKnownContent } from "../bridge.js";
 
 // ── CAST gates ────────────────────────────────────────────────────────────
 //
@@ -80,6 +87,85 @@ export interface CastResult {
   unexplained: string;
 }
 
+/** The seat that establishes a node's role in an analogical comparison:
+ *  the REVERSE context (what leads to it) when a predecessor genuinely
+ *  ESTABLISHES id — introduces or describes it by name — else the FORWARD
+ *  continuation (what it leads to), else `fallback`.
+ *
+ *  An earlier version gated this purely on `prevCount(id) > 0`: any
+ *  predecessor at all was treated as proof of a genuine named ENTITY
+ *  (seat it by what established it), while no predecessor meant a bare
+ *  learnt CONTEXT (seat it by what it leads to, since voicing it verbatim
+ *  would answer a question with a question).  That test measured the wrong
+ *  thing — a broad sample of this store's own question-shaped nodes showed
+ *  the large majority (≈71%) have at least one predecessor, most of them a
+ *  handful of generic, high-fan-out sentences that recur as an INCIDENTAL
+ *  neighbour to dozens of otherwise-unrelated destinations (a SmolSent-
+ *  style sentence-adjacency artifact, never naming or describing what
+ *  follows).  Traced live: "What is the capital of France?" — whose own
+ *  forward edge unambiguously resolves to "The capital of France is
+ *  Paris." — has exactly one such incidental predecessor ("Create an
+ *  example of a types of questions a GPT model can answer.?"), wrongly
+ *  read as disqualifying proof of "genuine entity."
+ *
+ *  A plain forward-first swap (matching {@link project}'s universal
+ *  priority) over-corrected: test/29's C2/C3 pin that a genuine entity
+ *  analog (e.g. "Leonardo da Vinci", established by "The Mona Lisa was
+ *  painted by Leonardo da Vinci.") must be seated by that establishing
+ *  sentence, NOT by its own biography fact — voicing the bio leaks exactly
+ *  what a comparison must keep out, and loses the embedded "Mona Lisa"
+ *  term C3 relies on for a further hop.
+ *
+ *  The distinguishing signal is content-addressed, not a count: a genuine
+ *  establishing predecessor's bytes CONTAIN id's own bytes — it names or
+ *  describes id ("...painted by Leonardo da Vinci." contains "Leonardo da
+ *  Vinci").  An incidental adjacency predecessor never does — it merely
+ *  preceded id in some unrelated document without ever mentioning it.  No
+ *  new tuned constant: containment is the same primitive `restatesQuery`
+ *  and `dominates`-style checks already use throughout this codebase.
+ *
+ *  `allowForward` (default true) gates the FORWARD branch specifically —
+ *  see the call sites below: the DOMINANT is what the query is actually
+ *  ASKING, so completing it forward is the whole point; an ANALOG is only
+ *  being CITED for comparison; the query never asked about IT, so chasing
+ *  its own further continuation drifts onto whatever coincidentally
+ *  follows it in the corpus.  Traced live: the analog "What is the capital
+ *  of Japan?\nTokyo is the capital of Japan." is ALREADY a complete,
+ *  self-answering unit (prevCount 0, so no establishing predecessor
+ *  either) — its sole forward edge is "And what is the capital of the
+ *  Moon?", an unrelated quiz question sharing nothing but corpus
+ *  adjacency.  With forward disallowed, an analog like this falls through
+ *  to `fallback` — its own bytes, exactly the complete fact that made it a
+ *  genuine analog in the first place.  See
+ *  test/41-seatofnode-direction.test.mjs and
+ *  test/43-cast-analog-seat.test.mjs. */
+export async function seatOfNode(
+  ctx: MindContext,
+  id: number,
+  guide: Vec | null | undefined,
+  fallback: Uint8Array,
+  allowForward = true,
+): Promise<Uint8Array> {
+  const rev = ctx.store.prevFirst(id, hubBound(ctx));
+  if (rev.length > 0) {
+    const own = read(ctx, id);
+    const establishing = rev.some((p) => indexOf(read(ctx, p), own, 0) >= 0);
+    if (establishing) {
+      const back = reverseContext(ctx, id, guide, rev);
+      if (back !== null) return back;
+    }
+  }
+  // The "last resort, non-establishing reverse" fallback below is itself a
+  // LESS CERTAIN projection (the same tier as forward) — an analog
+  // (allowForward: false) must stop at `fallback` (its own bytes) here
+  // rather than fall back to a predecessor that already failed the
+  // establishing check just above.
+  if (!allowForward) return fallback;
+  const fwd = await follow(ctx, id, guide);
+  if (fwd !== null) return fwd;
+  return reverseContext(ctx, id, guide, rev) ?? fallback;
+}
+
 /** CAST's own entry gates, checked once here and reused by
 /** The main CAST entry point.  Given a query and its pre-computed pre.rec.sites,
  *  determine whether the query weaves together multiple independent learnt
@@ -297,7 +383,10 @@ export async function counterfactualTransfer(
     const tail = proj.ctx.subarray(seat.cs);
     let answer = await joinWithBridge(ctx, filler, tail);
     const fwd = await follow(ctx, proj.anchor, qv);
-    if (fwd !== null && indexOf(answer, fwd, 0) < 0) {
+    if (
+      fwd !== null && indexOf(answer, fwd, 0) < 0 &&
+      !restatesQuery(query, fwd)
+    ) {
       answer = concat2(answer, fwd);
     }
     ctx.trace?.step(
@@ -375,36 +464,14 @@ export async function counterfactualTransfer(
   // seed-dependent failure where the climb ranks an exemplar above a
   // person node and the person node is excluded from points by run-
   // overlap trimming.
-  // The seat that establishes a candidate's role: the REVERSE projection
-  // (the context it follows), voiced by the query gist — falling back to the
-  // candidate's own bytes when it follows nothing.  DELIBERATE STRENGTHENING
-  // over the pre-refactor code: reverseContext also returns null when the
-  // picked context READS EMPTY (a dangling id degrades to zero bytes), so a
-  // corrupted-store read now falls back here instead of voicing a hollow
-  // seat into the comparison — the same "empty bytes are no grounding"
-  // invariant project() has always enforced.
-  // A node that only ever CONTINUES — an edge SOURCE with no predecessor —
-  // is a learnt CONTEXT (a stored question-shaped frame), not an entity.
-  // Voicing it verbatim answers a question with a question (the observed
-  // cast echo: two stored questions concatenated as an "answer").  The
-  // context's role is established by what it LEADS TO, so its seat is its
-  // own continuation — the fact — with the reverse projection and the raw
-  // bytes as the ordinary fallbacks for genuine entities.
-  const seatOfNode = async (
-    id: number,
-    fallback: Uint8Array,
-  ): Promise<Uint8Array> => {
-    if (ctx.store.prevCount(id) === 0 && ctx.store.hasNext(id)) {
-      const fwd = await follow(ctx, id, qv);
-      if (fwd !== null) return fwd;
-    }
-    return reverseContext(ctx, id, qv) ?? fallback;
-  };
-  const seatOf = (p: Point): Promise<Uint8Array> => seatOfNode(p.anchor, p.ctx);
+  // The seat that establishes a candidate's role — see {@link seatOfNode}.
+  const seatOf = (p: Point, allowForward = true): Promise<Uint8Array> =>
+    seatOfNode(ctx, p.anchor, qv, p.ctx, allowForward);
   interface AnalogCandidate {
     anchor: number;
     /** The point this candidate came from, or null when it is a nextOf
-     *  descendant — then seatOfNode supplies the seat. */
+     *  descendant — then its own bytes ARE the seat (already one meaningful
+     *  hop from `src`; see the comparison gate below). */
     point: Point | null;
     /** For a nextOf descendant: the aligned point whose continuation edge
      *  named it.  Its runs ARE the query evidence this analog rests on
@@ -447,8 +514,46 @@ export async function counterfactualTransfer(
   }
   let bestAnalog: AnalogCandidate | null = null;
   let bestSim = 0;
+  let bestHalo = false;
+  // Whether the query itself NAMES a candidate.  A directly aligned point
+  // is named by construction — its runs ARE query bytes.  A hop-reached
+  // candidate is named when its own bytes contain the query text of an
+  // aligned run of the point whose continuation edge reached it (that
+  // alignment IS the query evidence the hop rests on — the same reading
+  // cmpAccounted already prices): "William Shakespeare", reached off
+  // "Macbeth was written by William Shakespeare.", contains the src's
+  // 12-byte aligned run " Shakespeare" — test/29 C2/C3.  The run must span
+  // at least TWO perception windows (2·W, the same two-quantum floor
+  // CAST's own entry gate holds the whole query to): a single shared
+  // W-window is exactly the frame tier's own evidence quantum — the level
+  // "half the corpus" shares — and stopword scraps (" the ", "he b",
+  // 4–5 bytes) never reach two windows, while a genuinely named entity
+  // does.  NOT the weave's usable()/frame filter: weave depth counts every
+  // ranked exemplar, so a query's own named entity recurring across
+  // exemplars ("Shakespeare" in Hamlet+Macbeth+…) is wrongly classified as
+  // frame — measured live, it silently disqualified C3's genuine analog.
+  const namedByQuery = (c: AnalogCandidate): boolean => {
+    if (c.point !== null) return true;
+    const bytes = read(ctx, c.anchor);
+    return c.src.runs.some((r) =>
+      r.qe - r.qs >= 2 * quantum &&
+      indexOf(bytes, query.subarray(r.qs, r.qe), 0) >= 0
+    );
+  };
+  // Whether any committed root's consensus vote clears the SAME trust bar
+  // recallByResonance applies before grounding through a climb root:
+  // consensusFloor(N) = ln(N) + 1/2.  The climb's FIRST root is
+  // deliberately floor-free (attention.ts: "the dominant one always
+  // grounds") — fine for ORIENTING mechanisms, not for voicing learnt
+  // content the query never asked about.  Computed once here; both the
+  // hub fallback below and the comparison gate consume it.
+  const rootTrusted = roots.some((r) => r.vote >= consensusFloor(corpusN(ctx)));
   for (const c of analogs) {
-    const sim = await analogyStrength(ctx, dominant.anchor, c.anchor);
+    const { score: sim, halo } = await analogyStrength(
+      ctx,
+      dominant.anchor,
+      c.anchor,
+    );
     ctx.trace?.step(
       "tryAnalog",
       [
@@ -456,11 +561,12 @@ export async function counterfactualTransfer(
         rNode(ctx, c.anchor, "candidate", sim),
       ],
       [],
-      `analogy strength ${sim.toFixed(4)}`,
+      `analogy strength ${sim.toFixed(4)}${halo ? " (halo tier)" : ""}`,
     );
     if (sim > bestSim) {
       bestSim = sim;
       bestAnalog = c;
+      bestHalo = halo;
     }
   }
   // When every candidate fails the similarity gates (halo company — now
@@ -489,6 +595,17 @@ export async function counterfactualTransfer(
     let hubMass = -1;
     const fanClamp = hubBound(ctx) + 1;
     for (const c of analogs) {
+      // A fallback comparison carries NO similarity evidence at all.  Its
+      // honesty rests on the grounding decider discounting it against
+      // richer candidates (the design note below) — an assumption that
+      // holds only when the climb itself settled on this query with real
+      // evidence.  Under a root the consensus floor does not trust, an
+      // unnamed, hop-reached hub is pure corpus adjacency: refusing it is
+      // what kept the live wrong echo silent.  A hub the query itself
+      // NAMED stays eligible either way (test/29 C2/C3's "William
+      // Shakespeare"); an unnamed one under a TRUSTED root stays eligible
+      // too (test/33 1b's deliberately weak second candidate).
+      if (!rootTrusted && !namedByQuery(c)) continue;
       // Evidence clamped at the hub bound: beyond √N + 1 the exact fan-out
       // no longer discriminates (every mega-hub ties at the clamp), and
       // counting it exactly would require the corpus-sized read.
@@ -538,10 +655,111 @@ export async function counterfactualTransfer(
   // climb's own forest, never tuned; substitution/redirection stay
   // unaffected — they orient around a displaced seat, not a whole-topic
   // analogy.
+  //
+  // roots.length <= 1 is a PROXY for "the query is about one thing" — it is
+  // only as good as the climb's own root-commitment, which depends on
+  // recognise() having found something to commit a root TO.  When the
+  // query's newest content genuinely isn't recognised (not boundary noise —
+  // real, uncommitted content; see the session's own investigation of the
+  // France→Spain live trace), the climb under-commits roots and this proxy
+  // is fooled: comparison looks licensed to treat the query as one topic
+  // when it is not.
+  //
+  // The direct check is the SAME accounted spans comparison is about to
+  // cite as its evidence: unexplainedSpans (rationale.ts, the same gap
+  // computation the trace's own `unexplained` diagnostic uses) names every
+  // stretch of the query NEITHER the dominant NOR the analog's evidence
+  // touches.  A short comparison query ("How is ice like steel?") legitimately
+  // accounts for only its two short entity spans — the surrounding "How is
+  // ... like ...?" framing is real but SHORT, split into several small gaps,
+  // none of them the bulk of the query.  The live bug's shape is different in
+  // kind, not degree: ONE contiguous, substantial gap — a whole second
+  // question the query added that comparison's two spans never touch at all.
+  //
+  // Two bars, both derived, neither tuned:
+  //   • the largest gap must not DOMINATE the whole query (the same
+  //     predicate CAST's own frame gate uses) — rules out a gap that is
+  //     most of the query outright;
+  //   • the largest gap must be SMALLER than the dominant's own established
+  //     context.  A gap can't be dismissed as mere connective framing once
+  //     it is at least as large as the topic being compared FROM — at that
+  //     scale it isn't glue between two named things, it's substantial
+  //     enough to be a second topic in its own right.  This is what
+  //     actually separates the live bug (a 47-byte gap against a 30-byte
+  //     dominant — the ignored content is bigger than the topic itself)
+  //     from ordinary short comparisons (a 9-byte gap against an 11-byte
+  //     dominant — the gap is smaller than what's being compared): the two
+  //     cases land on the same side of "half the query" often enough
+  //     (both can exceed or clear it) that the query-relative bar alone
+  //     does not reliably separate them — the topic-relative scale does.
+  const cmpAccounted: Array<[number, number]> = bestAnalog !== null
+    ? [...runSpans(dominant), ...runSpans(bestAnalog.point ?? bestAnalog.src)]
+    : [];
+  const cmpGaps = unexplainedSpans(query.length, cmpAccounted);
+  const cmpMaxGap = cmpGaps.reduce((n, [s, e]) => Math.max(n, e - s), 0);
+  // An analog that is not itself a directly ALIGNED point (point !== null —
+  // its own runs are query bytes, the query NAMED it) was only reached
+  // through a continuation hop or the structural-hub fallback.  Voicing
+  // learnt content the query never named is the same act recallByResonance
+  // refuses to perform through a climb root whose consensus vote is below
+  // consensusFloor(N) = ln(N) + 1/2 (recall.ts's minVote), so comparison
+  // holds the climb to that SAME bar before citing a hop-reached analog:
+  // some committed root must clear the floor.  The climb's FIRST root is
+  // deliberately floor-free (attention.ts: "the dominant one always
+  // grounds") — fine for ORIENTING mechanisms, not for transferring
+  // unnamed content through.  The live bug this gates (real trained store,
+  // 325k edge sources, floor 13.2): the query's stopword scraps pooled a
+  // 1.92 vote that committed an unrelated haiku exemplar as the sole root,
+  // and comparison voiced that exemplar's continuation through a
+  // hop-reached analog while every other mechanism honestly refused.  A
+  // directly aligned analog needs no floor — the query's own bytes are its
+  // evidence (test/29 C1's "Steel is hard" for "How is ice like steel?").
+  // See test/50-cast-analog-consensus-floor.
+  // A HALO-tier best analog needs neither: its similarity already cleared
+  // significanceBar-gated distributional company (analogyStrength's
+  // `halo`) — genuine evidence in its own right, the very case the halo
+  // gate exists for (test/33 1b's nickname-corroborated analog).  Only a
+  // FRAME-tier or fallback analog — whose "similarity" is an unbarred
+  // coverage fraction or nothing — needs the query's naming or the climb's
+  // trust.
+  const analogNamed = bestAnalog !== null && namedByQuery(bestAnalog);
+  // NOTE — two further gates were tried here and empirically REFUTED,
+  // recorded so they are not re-tried:
+  //   • dominant self-coverage (dominant's aligned runs must dominate its
+  //     own ctx): legitimate dominants sit at the same coverage as junk
+  //     ones ("The Mona Lisa was painted by…" 16/47 vs the live junk
+  //     haiku ~10/54) — no separation.
+  //   • denying the shared-frame similarity tier to hop-reached analogs:
+  //     semantically right in isolation, but it merely promoted the next
+  //     junk candidate — an ALIGNED scrap-matched point ("The affluence…",
+  //     frame 0.157) — into bestAnalog on the live store, and the aligned
+  //     configuration is byte-structurally IDENTICAL to test/29 C1's
+  //     legitimate one ("Steel is hard", frame 0.364): every derived
+  //     local separator measured (run length, site overlap, frame
+  //     query-containment, weave-usable classification) falls on the same
+  //     side for both.  Only corpus-scale consensus separates them, which
+  //     is exactly what `rootTrusted` prices.
+  // FRAME-tier evidence under an UNTRUSTED root is comparison's weakest
+  // licence (an unbarred coverage fraction, a climb the consensus floor
+  // does not trust).  There it is additionally held to the IGNORED-KNOWN
+  // principle (dismissedKnownContent, bridge.ts): the two analogs' aligned
+  // runs must account for every STORED window of the query.  This is the
+  // byte-structural separator the refuted-gates note below could not find
+  // locally: a legitimate small-corpus comparison ("How is ice like
+  // steel?") leaves only UNATTESTED spans ("How ", " like ") unexplained,
+  // while a scrap-matched junk pair leaves the query's own trained content
+  // ("…songs…times…", "…planet…sun.") dismissed as gaps.  Halo-tier and
+  // trusted-root comparisons are exempt — their evidence already stands.
+  const cmpDismisses = !(bestHalo || rootTrusted) &&
+    dismissedKnownContent(ctx, query, cmpAccounted);
   if (
     bestAnalog !== null &&
+    (bestHalo || analogNamed || rootTrusted) &&
+    !cmpDismisses &&
     dominant.ctx.length <= query.length &&
-    roots.length <= 1
+    roots.length <= 1 &&
+    !dominates(cmpMaxGap, query.length) &&
+    cmpMaxGap < dominant.ctx.length
   ) {
     ctx.trace?.step(
       "validateAnalogy",
@@ -553,9 +771,21 @@ export async function counterfactualTransfer(
       "the two structures keep distributional company beyond chance — genuine analogs",
     );
     const a = await seatOf(dominant);
+    // The analog is only being CITED for comparison — the query never asked
+    // about it — so its seat never chases a FORWARD continuation (see
+    // seatOfNode's `allowForward`): only reverse (if a predecessor genuinely
+    // establishes it) or its own bytes.  A DIRECTLY aligned point
+    // (bestAnalog.point !== null) still goes through seatOfNode for that
+    // reverse check (a bare entity NAME like "Leonardo da Vinci" needs it —
+    // test/29's C2/C3).  A nextOf DESCENDANT (point === null) was already
+    // reached by following ONE meaningful hop off another aligned point (the
+    // alignment loop above: "its nextOf is the hub... and the hub's own
+    // [...] context will be the seat") — its own bytes ARE that seat
+    // directly, with no predecessor to even check (it was found by a
+    // forward edge, not matched in the query).
     const b = bestAnalog.point !== null
-      ? await seatOf(bestAnalog.point)
-      : await seatOfNode(bestAnalog.anchor, read(ctx, bestAnalog.anchor));
+      ? await seatOf(bestAnalog.point, false)
+      : read(ctx, bestAnalog.anchor);
     const answer = await joinWithBridge(ctx, a, b);
     record(
       answer,
@@ -568,10 +798,32 @@ export async function counterfactualTransfer(
       // when it was an aligned point, else the source point whose
       // continuation edge reached it (that alignment IS the query evidence
       // the hop rests on).
+      cmpAccounted,
+    );
+  } else if (
+    bestAnalog !== null &&
+    dominant.ctx.length <= query.length &&
+    roots.length <= 1
+  ) {
+    ctx.trace?.step(
+      "validateAnalogy",
       [
-        ...runSpans(dominant),
-        ...runSpans(bestAnalog.point ?? bestAnalog.src),
+        rNode(ctx, dominant.anchor, "analog", bestSim),
+        rNode(ctx, bestAnalog.anchor, "analog", bestSim),
       ],
+      [],
+      !(bestHalo || analogNamed || rootTrusted)
+        ? `the best analog carries no halo-tier company evidence, was never ` +
+          `named by the query, and no committed root's consensus vote ` +
+          `clears the floor, so comparison refuses to voice it`
+        : cmpDismisses
+        ? `a frame-tier analog under an untrusted root dismisses stored ` +
+          `query content its alignment never accounted for — comparison ` +
+          `refuses to ignore what the store knows`
+        : `comparison's own accounted evidence leaves a ${cmpMaxGap}-byte gap in ` +
+          `a ${query.length}-byte query against a ${dominant.ctx.length}-byte ` +
+          `dominant — too large to be mere framing — so it refuses rather ` +
+          `than paper over it with an analog the query never asked about`,
     );
   }
   t?.done(

package/src/mind/mechanisms/cover.ts

@@ -11,12 +11,11 @@
 
 import type { MindContext } from "../types.js";
 import type { ComputedResult, Site } from "../graph-search.js";
-import { bytesEqual, indexOf } from "../../bytes.js";
 import { read, resolve } from "../primitives.js";
 import { guidedFirst } from "../traverse.js";
 import { conceptHop } from "../match.js";
 import { bridge } from "../resonance.js";
-import { liftAnswer } from "../types.js";
+import { liftAnswer, segRestatesQuery } from "../types.js";
 import { decodeText, unexplainedLabel } from "../rationale.js";
 import type { RationaleItem } from "../rationale.js";
 import { rItem, rNode, traceDerivation } from "../trace.js";
@@ -226,44 +225,31 @@ export const coverMechanism: PipelineMechanism = {
 
     if (segs === null) return [];
 
+    const W = ctx.space.maxGroup;
     // A chosen span's SUBSTITUTED bytes (an edge followed from a recognised
     // site, not the site's own literal text read back) that equal a byte
     // span the query ALREADY CONTAINS elsewhere restates part of the
-    // question — never an answer (the same principle recall.ts's tiers
-    // apply to a whole-query projection: "a projection that is a proper
-    // byte-subspan of the query restates part of the question").  A
+    // question — never an answer (see {@link segRestatesQuery}).  A
     // recognised site that is itself an entire PRIOR TURN of a multi-turn
     // query is exactly this shape: it carries a genuine learnt
     // continuation, but that continuation is something the asker already
-    // said moments later in the SAME query, not a new answer — the cover
-    // search has no notion of "turn" to gate this itself, so the check
-    // belongs here, over the derivation it already chose.
-    const W = ctx.space.maxGroup;
-    for (const s of segs) {
-      if (!s.rec) continue;
-      const literal = s.j - s.i === s.bytes.length &&
-        bytesEqual(s.bytes, query.subarray(s.i, s.j));
-      if (literal) continue;
-      // A span shorter than one river window is exactly the "ice"/"c"
-      // case: a chain hop's short terminal coincidentally recurring as a
-      // substring of an unrelated longer word.  Below the quantum, byte
-      // overlap is chance, not evidence — the same floor identityBar and
-      // reachThreshold hold every other structural-overlap claim to.
-      if (
-        s.bytes.length >= W && s.bytes.length < query.length &&
-        indexOf(query, s.bytes, 0) >= 0
-      ) {
-        ctx.trace?.step(
-          "restatedSpan",
-          [rItem(s.bytes, "substituted", s.node, [s.i, s.j])],
-          [],
-          "the chosen span's substitution already occurs elsewhere in the query — restates it, not an answer",
-        );
-        return [];
-      }
+    // said moments later in the SAME query.  liftAnswer TRIMS such spans
+    // out of both the framing decision and the final concatenation — the
+    // OTHER spans a derivation chose are independent evidence and must not
+    // be discarded along with the stale one.
+    const restated = segs.filter((s) =>
+      segRestatesQuery(s, query, query.length, W)
+    );
+    if (restated.length > 0) {
+      ctx.trace?.step(
+        "restatedSpan",
+        restated.map((s) => rItem(s.bytes, "substituted", s.node, [s.i, s.j])),
+        [],
+        "the chosen span's substitution already occurs elsewhere in the query — trimmed from the answer, not an answer itself",
+      );
     }
 
-    const composed = liftAnswer(segs, query.length);
+    const composed = liftAnswer(segs, query.length, query, W);
     if (composed === null) return [];
 
     ctx.trace?.step(
@@ -276,11 +262,12 @@ export const coverMechanism: PipelineMechanism = {
       tCover ? [tCover.index] : undefined,
     );
 
-    // accounted = RECOGNISED cover spans only — PASS-carried bytes
-    // are priced in cost already; the diagnostic label reflects the
-    // same distinction.
+    // accounted = RECOGNISED, non-restating cover spans only — a trimmed
+    // restated span contributes nothing to the composed answer, so it must
+    // not be priced as if it did (PASS-carried bytes are priced already;
+    // the diagnostic label reflects the same distinction).
     const accounted: Array<[number, number]> = segs
-      .filter((s) => s.rec)
+      .filter((s) => s.rec && !segRestatesQuery(s, query, query.length, W))
       .map((s) => [s.i, s.j]);
 
     return [{

package/src/mind/mechanisms/recall.ts

@@ -20,6 +20,7 @@ import { CONCEPT, STEP } from "../graph-search.js";
 import { unexplainedLabel } from "../rationale.js";
 import type { PipelineMechanism, Precomputed } from "../pipeline-mechanism.js";
 import { rItem, rNode } from "../trace.js";
+import { substitutionBridge } from "../bridge.js";
 
 /** A recall result. */
 export interface RecallResult {
@@ -293,6 +294,84 @@ export async function recallByResonance(
       }
     }
   }
+  // 3b. Corroborated-substitution bridge — refusal-path only (bridge.ts).
+  // Every gist-based tier has failed; before refusing, align the query
+  // byte-for-byte against the trained contexts its own stored windows
+  // anchor, accepting mismatches only as corpus-attested, concept-bar
+  // substitutions.  A bridged context grounds exactly like any hit —
+  // projected through its learnt edges — under the same restated-fragment
+  // guard tiers 0b/2 apply.  Costs nothing on any answering path.
+  {
+    // The resonance hits already ranked above are handed to the bridge as
+    // PROPOSED candidates alongside its own anchor climbs: on a corpus this
+    // size a W-byte window is far too common for the clamped climb to
+    // single out the right trained context, while the whole-query gist
+    // already ranked it nearest (observed live: "what is the capital of
+    // france" resonating straight to "What is the capital of France?" yet
+    // refusing on the reach bar).  Approximate scores propose; the bridge's
+    // byte-exact alignment and attestation gates decide.
+    //
+    // The proposal breadth here is widened PAST `k` — first by requesting
+    // hubBound(ctx) candidates instead of `k` (recall's own tiers above
+    // stay at `k`; this re-resonates only on the refusal path, exactly
+    // where the bridge itself already runs), AND by asking the index to
+    // search EXHAUSTIVELY.  Both matter: the IVF only ever probes
+    // ⌈√clusters⌉ of them (store.ts's efFor) REGARDLESS of k — widening k
+    // alone just returns more hits from the SAME already-probed clusters,
+    // never a hit whose vector lives in an unprobed one.  Measured live:
+    // "What is the chemical symbol for water?" needs "What is the
+    // chemical formula for water?", scoring only 0.58 against the
+    // query's gist (a MIDDLE-of-string word swap perturbs the river-fold
+    // tree hash far more than a same-length TAIL swap like the "carbon"/
+    // "oxygen" neighbours that outrank it at 0.87+) — absent from the
+    // resonance list even at k=5000, present and byte-exact-verified the
+    // moment it's force-fed to the bridge directly.  `exhaustive` is the
+    // natural, tuning-free ceiling (probe every cluster) for a call that
+    // is ALREADY refusal-path-only and must not miss a candidate hiding
+    // behind an unlucky structural distance.
+    const wide = k < hubBound(ctx)
+      ? await ctx.store.resonate(queryGist, hubBound(ctx), true)
+      : whole;
+    const bridged = await substitutionBridge(
+      ctx,
+      query,
+      wide.map((h) => h.id),
+    );
+    if (bridged !== null) {
+      const g = await project(ctx, bridged.id, queryGist);
+      // A projection contained in a substituted candidate-side span is the
+      // substitution RESTATED as if it were knowledge — the exact failure
+      // observed live: "Darwin was born in England." bridged to the
+      // Einstein fact through " England." → " Germany." and would have
+      // voiced "Germany", an answer the substitution itself manufactured.
+      // The same principle as the restated-fragment guards above, extended
+      // to the bridge's own substitutions.
+      const cBytes = ctx.store.bytes(bridged.id);
+      const manufactured = g !== null &&
+        bridged.subs.some((s) =>
+          indexOf(cBytes.subarray(s.cs, s.ce), g!, 0) >= 0
+        );
+      if (
+        g !== null && g.length > 0 && !restates(g) && !manufactured &&
+        !(g.length < query.length && indexOf(query, g, 0) >= 0)
+      ) {
+        return ground(
+          g,
+          `substitution bridge — a trained context accounts for the query ` +
+            `up to ${bridged.subs.length} corroborated substitution(s)`,
+          // Accounted NOTHING — the same epistemic humility as the echo
+          // tier below: a substitution-bridged grounding is a last resort
+          // that must lose to ANY mechanism that actually explained the
+          // query (observed: pricing the aligned spans here outweighed
+          // extraction's correct answer in the grounding decider), while
+          // still beating silence when everything else refused.
+          [],
+          CONCEPT * bridged.subs.length + STEP,
+        );
+      }
+    }
+  }
+
   // The refusal/echo decision.  The echo returns a stored form's bytes AS
   // the answer — a near-identity claim about the query — and identity-grade
   // decisions are never made on an estimated score ("approximate scores may