@hviana/sema 0.2.3 → 0.2.5

package/dist/src/mind/junction.js

@@ -16,6 +16,20 @@ import { windowIds } from "./canonical.js";
 import { hubBound } from "./traverse.js";
 import { haloSiblings } from "./match.js";
 import { indexOf } from "../bytes.js";
+/** Resolve `left`/`right` to their exact node ids (when known) and load each
+ *  resolved side's halo siblings once — deterministic (haloSiblings already
+ *  ranks nearest-first) and shared by every ladder rung that consults
+ *  siblings, so no ladder rung repeats a halo ANN query the previous one
+ *  already paid for. */
+export async function loadJunctionSynonymSides(ctx, left, right) {
+    const leftId = resolve(ctx, left);
+    const rightId = resolve(ctx, right);
+    const leftSiblings = leftId !== null ? await haloSiblings(ctx, leftId) : [];
+    const rightSiblings = rightId !== null
+        ? await haloSiblings(ctx, rightId)
+        : [];
+    return { leftId, rightId, leftSiblings, rightSiblings };
+}
 /** Seed node ids to ascend from for one side of a junction: the side's own
  *  node when it is a stored form, plus — when the node has no structural
  *  parents — its canonical window ids.  A non-W-aligned node may have no
@@ -225,38 +239,79 @@ export function junctionContainers(ctx, left, right, maxContainer, unordered = f
  *  cost is bounded at √N·W pops total regardless of how many siblings are
  *  tried.  A sibling whose bytes exceed `maxInterior` is skipped (it
  *  cannot be junction-sized). */
-export async function junctionSynonyms(ctx, left, right, maxInterior, unordered = false) {
-    const out = [];
-    const lId = resolve(ctx, left);
-    const rId = resolve(ctx, right);
-    if (lId === null && rId === null)
-        return out;
-    const budget = { n: hubBound(ctx) * ctx.space.maxGroup };
+export async function junctionSynonyms(ctx, left, right, maxInterior, unordered = false, sides) {
+    const s = sides ?? await loadJunctionSynonymSides(ctx, left, right);
+    if (s.leftId === null && s.rightId === null)
+        return [];
+    // ── Tier 2.5a: single-synonym — one side replaced by a halo sibling ──────
+    // ONE shared expansion budget across BOTH directions of this tier.
+    const singleBudget = { n: hubBound(ctx) * ctx.space.maxGroup };
+    const singleOut = new Map();
+    const keepBest = (map, j, tier, confidence) => {
+        const prev = map.get(j.id);
+        if (prev === undefined || confidence > prev.confidence) {
+            map.set(j.id, { ...j, tier, confidence });
+        }
+    };
     // Left-side synonyms: containers of sibling+right.  `right`'s seeds are
     // FIXED across every sibling this loop tries.
-    if (lId !== null) {
+    if (s.leftId !== null) {
         const rightSeeds = junctionSeeds(ctx, right);
-        for (const sib of await haloSiblings(ctx, lId)) {
+        for (const sib of s.leftSiblings) {
             const sibBytes = read(ctx, sib.id, maxInterior + 1);
             if (sibBytes.length === 0 || sibBytes.length > maxInterior)
                 continue;
-            const containers = junctionContainersFrom(ctx, sibBytes, right, sibBytes.length + right.length + maxInterior, junctionSeeds(ctx, sibBytes), rightSeeds, budget, unordered);
-            for (const c of containers)
-                out.push(c);
+            const containers = junctionContainersFrom(ctx, sibBytes, right, sibBytes.length + right.length + maxInterior, junctionSeeds(ctx, sibBytes), rightSeeds, singleBudget, unordered);
+            for (const c of containers) {
+                keepBest(singleOut, c, "single-synonym", sib.score);
+            }
         }
     }
     // Right-side synonyms: containers of left+sibling.  `left`'s seeds are
     // likewise fixed across this loop.
-    if (rId !== null) {
+    if (s.rightId !== null) {
         const leftSeeds = junctionSeeds(ctx, left);
-        for (const sib of await haloSiblings(ctx, rId)) {
+        for (const sib of s.rightSiblings) {
             const sibBytes = read(ctx, sib.id, maxInterior + 1);
             if (sibBytes.length === 0 || sibBytes.length > maxInterior)
                 continue;
-            const containers = junctionContainersFrom(ctx, left, sibBytes, left.length + sibBytes.length + maxInterior, leftSeeds, junctionSeeds(ctx, sibBytes), budget, unordered);
-            for (const c of containers)
-                out.push(c);
+            const containers = junctionContainersFrom(ctx, left, sibBytes, left.length + sibBytes.length + maxInterior, leftSeeds, junctionSeeds(ctx, sibBytes), singleBudget, unordered);
+            for (const c of containers) {
+                keepBest(singleOut, c, "single-synonym", sib.score);
+            }
         }
     }
-    return out;
+    if (singleOut.size > 0)
+        return [...singleOut.values()];
+    // ── Tier 2.5b: double-synonym — BOTH sides replaced, tried only when
+    // single-synonym found NOTHING.  Every (leftSibling, rightSibling) pair,
+    // sorted deterministically, bounded to haloQueryK pairs total, ONE fresh
+    // shared budget for the whole tier. ─────────────────────────────────────
+    if (s.leftSiblings.length === 0 || s.rightSiblings.length === 0)
+        return [];
+    const pairs = [];
+    for (const l of s.leftSiblings) {
+        for (const r of s.rightSiblings) {
+            pairs.push({ l, r, confidence: Math.min(l.score, r.score) });
+        }
+    }
+    pairs.sort((a, b) => b.confidence - a.confidence ||
+        a.l.id - b.l.id ||
+        a.r.id - b.r.id);
+    const doubleOut = new Map();
+    const budget = { n: hubBound(ctx) * ctx.space.maxGroup };
+    const tries = Math.min(pairs.length, ctx.cfg.haloQueryK);
+    for (let i = 0; i < tries; i++) {
+        const { l, r, confidence } = pairs[i];
+        const lBytes = read(ctx, l.id, maxInterior + 1);
+        const rBytes = read(ctx, r.id, maxInterior + 1);
+        if (lBytes.length === 0 || lBytes.length > maxInterior ||
+            rBytes.length === 0 || rBytes.length > maxInterior)
+            continue;
+        const containers = junctionContainersFrom(ctx, lBytes, rBytes, lBytes.length + rBytes.length + maxInterior, junctionSeeds(ctx, lBytes), junctionSeeds(ctx, rBytes), budget, unordered);
+        for (const c of containers) {
+            keepBest(doubleOut, c, "double-synonym", confidence);
+        }
+    }
+    return [...doubleOut.values()];
 }

package/dist/src/mind/match.d.ts

@@ -64,8 +64,21 @@ export declare function haloSiblings(ctx: MindContext, id: number, halo?: Vec |
  *  strength, not a pick.  Returns the direct halo cosine, or failing that the
  *  highest mutual-halo-sibling min-score (second-order analogy), or failing
  *  that the SHARED-FRAME strength (below) — the gate CAST's comparison
- *  schema validates genuine analogs with (bar: significanceBar). */
-export declare function analogyStrength(ctx: MindContext, a: number, b: number): Promise<number>;
+ *  schema validates genuine analogs with (bar: significanceBar).
+ *
+ *  The result names its TIER alongside the score: `halo: true` means the
+ *  score cleared a significanceBar-gated HALO tier (direct company cosine
+ *  or mutual-sibling) — genuine distributional evidence; `halo: false`
+ *  means only the structural shared-frame fallback matched, a coverage
+ *  fraction with no bar of its own.  CAST's comparison gate treats the two
+ *  differently (see cast.ts): halo evidence stands alone, frame evidence
+ *  needs the query to have named the analog or the climb root to be
+ *  trusted. */
+export interface AnalogyEvidence {
+    score: number;
+    halo: boolean;
+}
+export declare function analogyStrength(ctx: MindContext, a: number, b: number): Promise<AnalogyEvidence>;
 /** The STRUCTURAL analogy tier: two nodes are analogs when their byte
  *  streams share a LEARNT frame — a content-addressed flat form of at least
  *  one full river window (W bytes, the perception quantum) that occurs in

package/dist/src/mind/match.js

@@ -257,11 +257,6 @@ export async function haloSiblings(ctx, id, halo, bar = conceptThreshold(ctx.sto
         memo.set(id, out);
     return out;
 }
-/** The DISTRIBUTIONAL matcher between two nodes: mutual-nearest-neighbour
- *  strength, not a pick.  Returns the direct halo cosine, or failing that the
- *  highest mutual-halo-sibling min-score (second-order analogy), or failing
- *  that the SHARED-FRAME strength (below) — the gate CAST's comparison
- *  schema validates genuine analogs with (bar: significanceBar). */
 export async function analogyStrength(ctx, a, b) {
     const ha = ctx.store.halo(a);
     const hb = ctx.store.halo(b);
@@ -269,7 +264,7 @@ export async function analogyStrength(ctx, a, b) {
         const bar = significanceBar(ctx.store.D);
         const direct = cosine(ha, hb);
         if (direct >= bar)
-            return direct;
+            return { score: direct, halo: true };
         const sibsA = await haloSiblings(ctx, a, ha, bar);
         const sibsB = await haloSiblings(ctx, b, hb, bar);
         let best = 0;
@@ -282,9 +277,9 @@ export async function analogyStrength(ctx, a, b) {
             }
         }
         if (best > 0)
-            return best;
+            return { score: best, halo: true };
     }
-    return sharedFrameStrength(ctx, a, b);
+    return { score: sharedFrameStrength(ctx, a, b), halo: false };
 }
 /** The STRUCTURAL analogy tier: two nodes are analogs when their byte
  *  streams share a LEARNT frame — a content-addressed flat form of at least

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

@@ -1,4 +1,5 @@
 import type { MindContext } from "../types.js";
+import type { Vec } from "../../vec.js";
 /** A CAST answer plus its elementary evidence for think's grounding decider:
  *  `accounted` — the query spans the weave's aligned runs explain; `moves` —
  *  the ladder cost of the acts the taken branch performed (STEP per
@@ -13,6 +14,59 @@ export interface CastResult {
      *  Task 2 note in pipeline.ts's Candidate interface). */
     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 declare function seatOfNode(ctx: MindContext, id: number, guide: Vec | null | undefined, fallback: Uint8Array, allowForward?: boolean): Promise<Uint8Array>;
 /** 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

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

@@ -11,14 +11,16 @@
 // gate = the frame gate below + analogyStrength, projection = insert / project /
 // juxtapose.
 import { read } from "../primitives.js";
-import { argmaxBy, hubBound } from "../traverse.js";
+import { argmaxBy, corpusN, hubBound } from "../traverse.js";
 import { analogyStrength, follow, project, 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 { unexplainedLabel } from "../rationale.js";
+import { consensusFloor, dominates } from "../../geometry.js";
+import { unexplainedLabel, unexplainedSpans, } from "../rationale.js";
 import { rItem, rNode } from "../trace.js";
+import { dismissedKnownContent } from "../bridge.js";
 // ── CAST gates ────────────────────────────────────────────────────────────
 //
 // The frame gate has TWO components, both derived from the weave itself:
@@ -52,6 +54,81 @@ import { rItem, rNode } from "../trace.js";
  *  minimum: `depth > MIN_WEAVE` means at least three structures agree on a
  *  byte, so no byte is frame when only the minimum pair exists. */
 const MIN_WEAVE = 2;
+/** 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, id, guide, fallback, allowForward = true) {
+    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
@@ -222,7 +299,8 @@ export async function counterfactualTransfer(ctx, query, pre) {
         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("projectCounterfactual", [
@@ -278,30 +356,8 @@ export async function counterfactualTransfer(ctx, query, pre) {
     // 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, fallback) => {
-        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) => seatOfNode(p.anchor, p.ctx);
+    // The seat that establishes a candidate's role — see {@link seatOfNode}.
+    const seatOf = (p, allowForward = true) => seatOfNode(ctx, p.anchor, qv, p.ctx, allowForward);
     const analogs = [];
     for (const p of points) {
         if (p === dominant)
@@ -335,15 +391,49 @@ export async function counterfactualTransfer(ctx, query, pre) {
     }
     let bestAnalog = 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) => {
+        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", [
             rNode(ctx, dominant.anchor, "dominant"),
             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
@@ -372,6 +462,18 @@ export async function counterfactualTransfer(ctx, query, pre) {
         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.
@@ -412,17 +514,130 @@ export async function counterfactualTransfer(ctx, query, pre) {
     // 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 = 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", [
             rNode(ctx, dominant.anchor, "analog", bestSim),
             rNode(ctx, bestAnalog.anchor, "analog", bestSim),
         ], [], "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, "analogical comparison — each analog voiced by the context that establishes its role", new Set([dominant.anchor, bestAnalog.anchor]), 
         // A halo-mediated act (the analogy gate) plus two seat projections.
@@ -432,10 +647,26 @@ export async function counterfactualTransfer(ctx, query, pre) {
         // 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).
-        [
-            ...runSpans(dominant),
-            ...runSpans(bestAnalog.point ?? bestAnalog.src),
-        ]);
+        cmpAccounted);
+    }
+    else if (bestAnalog !== null &&
+        dominant.ctx.length <= query.length &&
+        roots.length <= 1) {
+        ctx.trace?.step("validateAnalogy", [
+            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 ${dominant.ctx.length}-byte ` +
+                    `dominant — too large to be mere framing — so it refuses rather ` +
+                    `than paper over it with an analog the query never asked about`);
     }
     t?.done(results.map((r) => rItem(r.bytes, "answer")), results.length > 0
         ? `${results.length} counterfactual schema(s) fired — the grounding decider weighs them`