@hviana/sema 0.2.3 → 0.2.5
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/src/geometry.d.ts +26 -0
- package/dist/src/geometry.js +32 -3
- package/dist/src/mind/attention.d.ts +246 -7
- package/dist/src/mind/attention.js +771 -173
- package/dist/src/mind/bridge.d.ts +30 -0
- package/dist/src/mind/bridge.js +569 -0
- package/dist/src/mind/index.d.ts +2 -0
- package/dist/src/mind/junction.d.ts +30 -1
- package/dist/src/mind/junction.js +73 -18
- package/dist/src/mind/match.d.ts +15 -2
- package/dist/src/mind/match.js +3 -8
- package/dist/src/mind/mechanisms/cast.d.ts +54 -0
- package/dist/src/mind/mechanisms/cast.js +268 -37
- package/dist/src/mind/mechanisms/cover.js +16 -31
- package/dist/src/mind/mechanisms/recall.js +66 -0
- package/dist/src/mind/mind.d.ts +2 -0
- package/dist/src/mind/rationale.d.ts +7 -2
- package/dist/src/mind/rationale.js +6 -5
- package/dist/src/mind/reasoning.d.ts +11 -0
- package/dist/src/mind/reasoning.js +58 -2
- package/dist/src/mind/recognition.js +127 -7
- package/dist/src/mind/traverse.js +90 -25
- package/dist/src/mind/types.d.ts +57 -2
- package/dist/src/mind/types.js +38 -7
- package/dist/src/store.d.ts +12 -3
- package/dist/src/store.js +9 -3
- package/package.json +1 -1
- package/src/geometry.ts +52 -3
- package/src/mind/attention.ts +1199 -128
- package/src/mind/bridge.ts +596 -0
- package/src/mind/index.ts +16 -0
- package/src/mind/junction.ts +125 -16
- package/src/mind/match.ts +19 -5
- package/src/mind/mechanisms/cast.ts +290 -38
- package/src/mind/mechanisms/cover.ts +23 -36
- package/src/mind/mechanisms/recall.ts +79 -0
- package/src/mind/mind.ts +15 -0
- package/src/mind/rationale.ts +12 -4
- package/src/mind/reasoning.ts +71 -2
- package/src/mind/recognition.ts +132 -7
- package/src/mind/traverse.ts +91 -24
- package/src/mind/types.ts +95 -6
- package/src/store.ts +19 -5
- package/test/36-already-answered-fusion.test.mjs +128 -0
- package/test/37-cluster-dispersion-fusion.test.mjs +190 -0
- package/test/38-reason-restate-guard.test.mjs +94 -0
- package/test/39-cast-restate-guard.test.mjs +102 -0
- package/test/40-choosenext-scale-guard.test.mjs +75 -0
- package/test/41-seatofnode-direction.test.mjs +85 -0
- package/test/42-recognise-trace-idempotence.test.mjs +106 -0
- package/test/43-cast-analog-seat.test.mjs +244 -0
- package/test/44-recognise-edge-whitespace.test.mjs +63 -0
- package/test/45-liftanswer-restated-trim.test.mjs +60 -0
- package/test/46-recognise-multibyte-edge.test.mjs +85 -0
- package/test/47-cast-comparison-coverage.test.mjs +134 -0
- package/test/48-recognise-turn-connective.test.mjs +125 -0
- package/test/49-natural-units-synonym-bridge.test.mjs +175 -0
- package/test/50-cast-analog-consensus-floor.test.mjs +179 -0
- package/test/51-structural-resonance-ladder.test.mjs +552 -0
- package/test/52-climb-consensus-instrumentation.test.mjs +324 -0
package/src/mind/junction.ts
CHANGED
|
@@ -12,6 +12,7 @@
|
|
|
12
12
|
// adjacent ANSWER pieces) and cross-region attention (the joint CONTEXT of two
|
|
13
13
|
// non-adjacent QUERY regions) ascend by the same disciplined, bounded walk.
|
|
14
14
|
|
|
15
|
+
import type { Hit } from "../store.js";
|
|
15
16
|
import type { MindContext } from "./types.js";
|
|
16
17
|
import { read, resolve } from "./primitives.js";
|
|
17
18
|
import { windowIds } from "./canonical.js";
|
|
@@ -27,6 +28,52 @@ export interface Junction {
|
|
|
27
28
|
interior: Uint8Array;
|
|
28
29
|
}
|
|
29
30
|
|
|
31
|
+
/** Which relaxation produced a {@link SynonymJunction}: one side replaced by
|
|
32
|
+
* a distributional halo sibling (`single-synonym`), or both (`double-
|
|
33
|
+
* synonym`) — the two remaining rungs of the graded ladder below exact DAG
|
|
34
|
+
* containment (see the module doc atop {@link junctionSynonyms}). */
|
|
35
|
+
export type SynonymJunctionTier =
|
|
36
|
+
| "single-synonym"
|
|
37
|
+
| "double-synonym";
|
|
38
|
+
|
|
39
|
+
export interface SynonymJunction extends Junction {
|
|
40
|
+
tier: SynonymJunctionTier;
|
|
41
|
+
/** Sibling score for a single-synonym junction; min(left, right) sibling
|
|
42
|
+
* score for a double-synonym junction. */
|
|
43
|
+
confidence: number;
|
|
44
|
+
}
|
|
45
|
+
|
|
46
|
+
/** The exact node ids and halo siblings resolved for one junction call's two
|
|
47
|
+
* sides — computed ONCE and reused by every ladder rung that needs them
|
|
48
|
+
* (junctionSynonyms' two tiers, and the structural-resonance tier beyond
|
|
49
|
+
* it). A failed synonym junction means only "no common DAG container was
|
|
50
|
+
* proven" — it does NOT mean the loaded siblings stop being useful. */
|
|
51
|
+
export interface JunctionSynonymSides {
|
|
52
|
+
leftId: number | null;
|
|
53
|
+
rightId: number | null;
|
|
54
|
+
leftSiblings: Hit[];
|
|
55
|
+
rightSiblings: Hit[];
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
/** Resolve `left`/`right` to their exact node ids (when known) and load each
|
|
59
|
+
* resolved side's halo siblings once — deterministic (haloSiblings already
|
|
60
|
+
* ranks nearest-first) and shared by every ladder rung that consults
|
|
61
|
+
* siblings, so no ladder rung repeats a halo ANN query the previous one
|
|
62
|
+
* already paid for. */
|
|
63
|
+
export async function loadJunctionSynonymSides(
|
|
64
|
+
ctx: MindContext,
|
|
65
|
+
left: Uint8Array,
|
|
66
|
+
right: Uint8Array,
|
|
67
|
+
): Promise<JunctionSynonymSides> {
|
|
68
|
+
const leftId = resolve(ctx, left);
|
|
69
|
+
const rightId = resolve(ctx, right);
|
|
70
|
+
const leftSiblings = leftId !== null ? await haloSiblings(ctx, leftId) : [];
|
|
71
|
+
const rightSiblings = rightId !== null
|
|
72
|
+
? await haloSiblings(ctx, rightId)
|
|
73
|
+
: [];
|
|
74
|
+
return { leftId, rightId, leftSiblings, rightSiblings };
|
|
75
|
+
}
|
|
76
|
+
|
|
30
77
|
/** Seed node ids to ascend from for one side of a junction: the side's own
|
|
31
78
|
* node when it is a stored form, plus — when the node has no structural
|
|
32
79
|
* parents — its canonical window ids. A non-W-aligned node may have no
|
|
@@ -292,20 +339,32 @@ export async function junctionSynonyms(
|
|
|
292
339
|
right: Uint8Array,
|
|
293
340
|
maxInterior: number,
|
|
294
341
|
unordered = false,
|
|
295
|
-
|
|
296
|
-
|
|
297
|
-
|
|
298
|
-
|
|
299
|
-
const rId = resolve(ctx, right);
|
|
300
|
-
if (lId === null && rId === null) return out;
|
|
342
|
+
sides?: JunctionSynonymSides,
|
|
343
|
+
): Promise<SynonymJunction[]> {
|
|
344
|
+
const s = sides ?? await loadJunctionSynonymSides(ctx, left, right);
|
|
345
|
+
if (s.leftId === null && s.rightId === null) return [];
|
|
301
346
|
|
|
302
|
-
|
|
347
|
+
// ── Tier 2.5a: single-synonym — one side replaced by a halo sibling ──────
|
|
348
|
+
// ONE shared expansion budget across BOTH directions of this tier.
|
|
349
|
+
const singleBudget = { n: hubBound(ctx) * ctx.space.maxGroup };
|
|
350
|
+
const singleOut = new Map<number, SynonymJunction>();
|
|
351
|
+
const keepBest = (
|
|
352
|
+
map: Map<number, SynonymJunction>,
|
|
353
|
+
j: Junction,
|
|
354
|
+
tier: SynonymJunctionTier,
|
|
355
|
+
confidence: number,
|
|
356
|
+
) => {
|
|
357
|
+
const prev = map.get(j.id);
|
|
358
|
+
if (prev === undefined || confidence > prev.confidence) {
|
|
359
|
+
map.set(j.id, { ...j, tier, confidence });
|
|
360
|
+
}
|
|
361
|
+
};
|
|
303
362
|
|
|
304
363
|
// Left-side synonyms: containers of sibling+right. `right`'s seeds are
|
|
305
364
|
// FIXED across every sibling this loop tries.
|
|
306
|
-
if (
|
|
365
|
+
if (s.leftId !== null) {
|
|
307
366
|
const rightSeeds = junctionSeeds(ctx, right);
|
|
308
|
-
for (const sib of
|
|
367
|
+
for (const sib of s.leftSiblings) {
|
|
309
368
|
const sibBytes = read(ctx, sib.id, maxInterior + 1);
|
|
310
369
|
if (sibBytes.length === 0 || sibBytes.length > maxInterior) continue;
|
|
311
370
|
const containers = junctionContainersFrom(
|
|
@@ -315,18 +374,20 @@ export async function junctionSynonyms(
|
|
|
315
374
|
sibBytes.length + right.length + maxInterior,
|
|
316
375
|
junctionSeeds(ctx, sibBytes),
|
|
317
376
|
rightSeeds,
|
|
318
|
-
|
|
377
|
+
singleBudget,
|
|
319
378
|
unordered,
|
|
320
379
|
);
|
|
321
|
-
for (const c of containers)
|
|
380
|
+
for (const c of containers) {
|
|
381
|
+
keepBest(singleOut, c, "single-synonym", sib.score);
|
|
382
|
+
}
|
|
322
383
|
}
|
|
323
384
|
}
|
|
324
385
|
|
|
325
386
|
// Right-side synonyms: containers of left+sibling. `left`'s seeds are
|
|
326
387
|
// likewise fixed across this loop.
|
|
327
|
-
if (
|
|
388
|
+
if (s.rightId !== null) {
|
|
328
389
|
const leftSeeds = junctionSeeds(ctx, left);
|
|
329
|
-
for (const sib of
|
|
390
|
+
for (const sib of s.rightSiblings) {
|
|
330
391
|
const sibBytes = read(ctx, sib.id, maxInterior + 1);
|
|
331
392
|
if (sibBytes.length === 0 || sibBytes.length > maxInterior) continue;
|
|
332
393
|
const containers = junctionContainersFrom(
|
|
@@ -336,12 +397,60 @@ export async function junctionSynonyms(
|
|
|
336
397
|
left.length + sibBytes.length + maxInterior,
|
|
337
398
|
leftSeeds,
|
|
338
399
|
junctionSeeds(ctx, sibBytes),
|
|
339
|
-
|
|
400
|
+
singleBudget,
|
|
340
401
|
unordered,
|
|
341
402
|
);
|
|
342
|
-
for (const c of containers)
|
|
403
|
+
for (const c of containers) {
|
|
404
|
+
keepBest(singleOut, c, "single-synonym", sib.score);
|
|
405
|
+
}
|
|
343
406
|
}
|
|
344
407
|
}
|
|
345
408
|
|
|
346
|
-
return
|
|
409
|
+
if (singleOut.size > 0) return [...singleOut.values()];
|
|
410
|
+
|
|
411
|
+
// ── Tier 2.5b: double-synonym — BOTH sides replaced, tried only when
|
|
412
|
+
// single-synonym found NOTHING. Every (leftSibling, rightSibling) pair,
|
|
413
|
+
// sorted deterministically, bounded to haloQueryK pairs total, ONE fresh
|
|
414
|
+
// shared budget for the whole tier. ─────────────────────────────────────
|
|
415
|
+
if (s.leftSiblings.length === 0 || s.rightSiblings.length === 0) return [];
|
|
416
|
+
|
|
417
|
+
const pairs: Array<{ l: Hit; r: Hit; confidence: number }> = [];
|
|
418
|
+
for (const l of s.leftSiblings) {
|
|
419
|
+
for (const r of s.rightSiblings) {
|
|
420
|
+
pairs.push({ l, r, confidence: Math.min(l.score, r.score) });
|
|
421
|
+
}
|
|
422
|
+
}
|
|
423
|
+
pairs.sort((a, b) =>
|
|
424
|
+
b.confidence - a.confidence ||
|
|
425
|
+
a.l.id - b.l.id ||
|
|
426
|
+
a.r.id - b.r.id
|
|
427
|
+
);
|
|
428
|
+
|
|
429
|
+
const doubleOut = new Map<number, SynonymJunction>();
|
|
430
|
+
const budget = { n: hubBound(ctx) * ctx.space.maxGroup };
|
|
431
|
+
const tries = Math.min(pairs.length, ctx.cfg.haloQueryK);
|
|
432
|
+
for (let i = 0; i < tries; i++) {
|
|
433
|
+
const { l, r, confidence } = pairs[i];
|
|
434
|
+
const lBytes = read(ctx, l.id, maxInterior + 1);
|
|
435
|
+
const rBytes = read(ctx, r.id, maxInterior + 1);
|
|
436
|
+
if (
|
|
437
|
+
lBytes.length === 0 || lBytes.length > maxInterior ||
|
|
438
|
+
rBytes.length === 0 || rBytes.length > maxInterior
|
|
439
|
+
) continue;
|
|
440
|
+
const containers = junctionContainersFrom(
|
|
441
|
+
ctx,
|
|
442
|
+
lBytes,
|
|
443
|
+
rBytes,
|
|
444
|
+
lBytes.length + rBytes.length + maxInterior,
|
|
445
|
+
junctionSeeds(ctx, lBytes),
|
|
446
|
+
junctionSeeds(ctx, rBytes),
|
|
447
|
+
budget,
|
|
448
|
+
unordered,
|
|
449
|
+
);
|
|
450
|
+
for (const c of containers) {
|
|
451
|
+
keepBest(doubleOut, c, "double-synonym", confidence);
|
|
452
|
+
}
|
|
453
|
+
}
|
|
454
|
+
|
|
455
|
+
return [...doubleOut.values()];
|
|
347
456
|
}
|
package/src/mind/match.ts
CHANGED
|
@@ -325,18 +325,32 @@ export async function haloSiblings(
|
|
|
325
325
|
* strength, not a pick. Returns the direct halo cosine, or failing that the
|
|
326
326
|
* highest mutual-halo-sibling min-score (second-order analogy), or failing
|
|
327
327
|
* that the SHARED-FRAME strength (below) — the gate CAST's comparison
|
|
328
|
-
* schema validates genuine analogs with (bar: significanceBar).
|
|
328
|
+
* schema validates genuine analogs with (bar: significanceBar).
|
|
329
|
+
*
|
|
330
|
+
* The result names its TIER alongside the score: `halo: true` means the
|
|
331
|
+
* score cleared a significanceBar-gated HALO tier (direct company cosine
|
|
332
|
+
* or mutual-sibling) — genuine distributional evidence; `halo: false`
|
|
333
|
+
* means only the structural shared-frame fallback matched, a coverage
|
|
334
|
+
* fraction with no bar of its own. CAST's comparison gate treats the two
|
|
335
|
+
* differently (see cast.ts): halo evidence stands alone, frame evidence
|
|
336
|
+
* needs the query to have named the analog or the climb root to be
|
|
337
|
+
* trusted. */
|
|
338
|
+
export interface AnalogyEvidence {
|
|
339
|
+
score: number;
|
|
340
|
+
halo: boolean;
|
|
341
|
+
}
|
|
342
|
+
|
|
329
343
|
export async function analogyStrength(
|
|
330
344
|
ctx: MindContext,
|
|
331
345
|
a: number,
|
|
332
346
|
b: number,
|
|
333
|
-
): Promise<
|
|
347
|
+
): Promise<AnalogyEvidence> {
|
|
334
348
|
const ha = ctx.store.halo(a);
|
|
335
349
|
const hb = ctx.store.halo(b);
|
|
336
350
|
if (ha && hb) {
|
|
337
351
|
const bar = significanceBar(ctx.store.D);
|
|
338
352
|
const direct = cosine(ha, hb);
|
|
339
|
-
if (direct >= bar) return direct;
|
|
353
|
+
if (direct >= bar) return { score: direct, halo: true };
|
|
340
354
|
const sibsA = await haloSiblings(ctx, a, ha, bar);
|
|
341
355
|
const sibsB = await haloSiblings(ctx, b, hb, bar);
|
|
342
356
|
let best = 0;
|
|
@@ -347,9 +361,9 @@ export async function analogyStrength(
|
|
|
347
361
|
best = Math.max(best, Math.min(x.score, y.score));
|
|
348
362
|
}
|
|
349
363
|
}
|
|
350
|
-
if (best > 0) return best;
|
|
364
|
+
if (best > 0) return { score: best, halo: true };
|
|
351
365
|
}
|
|
352
|
-
return sharedFrameStrength(ctx, a, b);
|
|
366
|
+
return { score: sharedFrameStrength(ctx, a, b), halo: false };
|
|
353
367
|
}
|
|
354
368
|
|
|
355
369
|
/** The STRUCTURAL analogy tier: two nodes are analogs when their byte
|
|
@@ -12,8 +12,9 @@
|
|
|
12
12
|
// juxtapose.
|
|
13
13
|
|
|
14
14
|
import type { MindContext } from "../types.js";
|
|
15
|
+
import type { Vec } from "../../vec.js";
|
|
15
16
|
import { read } from "../primitives.js";
|
|
16
|
-
import { argmaxBy, hubBound } from "../traverse.js";
|
|
17
|
+
import { argmaxBy, corpusN, hubBound } from "../traverse.js";
|
|
17
18
|
import {
|
|
18
19
|
analogyStrength,
|
|
19
20
|
follow,
|
|
@@ -22,11 +23,17 @@ import {
|
|
|
22
23
|
reverseContext,
|
|
23
24
|
} from "../match.js";
|
|
24
25
|
import { joinWithBridge } from "../resonance.js";
|
|
26
|
+
import { restatesQuery } from "../reasoning.js";
|
|
25
27
|
import { CONCEPT, STEP } from "../graph-search.js";
|
|
26
28
|
import { concat2, indexOf } from "../../bytes.js";
|
|
27
|
-
import { dominates } from "../../geometry.js";
|
|
28
|
-
import {
|
|
29
|
+
import { consensusFloor, dominates } from "../../geometry.js";
|
|
30
|
+
import {
|
|
31
|
+
decodeText,
|
|
32
|
+
unexplainedLabel,
|
|
33
|
+
unexplainedSpans,
|
|
34
|
+
} from "../rationale.js";
|
|
29
35
|
import { rItem, rNode } from "../trace.js";
|
|
36
|
+
import { dismissedKnownContent } from "../bridge.js";
|
|
30
37
|
|
|
31
38
|
// ── CAST gates ────────────────────────────────────────────────────────────
|
|
32
39
|
//
|
|
@@ -80,6 +87,85 @@ export interface CastResult {
|
|
|
80
87
|
unexplained: string;
|
|
81
88
|
}
|
|
82
89
|
|
|
90
|
+
/** The seat that establishes a node's role in an analogical comparison:
|
|
91
|
+
* the REVERSE context (what leads to it) when a predecessor genuinely
|
|
92
|
+
* ESTABLISHES id — introduces or describes it by name — else the FORWARD
|
|
93
|
+
* continuation (what it leads to), else `fallback`.
|
|
94
|
+
*
|
|
95
|
+
* An earlier version gated this purely on `prevCount(id) > 0`: any
|
|
96
|
+
* predecessor at all was treated as proof of a genuine named ENTITY
|
|
97
|
+
* (seat it by what established it), while no predecessor meant a bare
|
|
98
|
+
* learnt CONTEXT (seat it by what it leads to, since voicing it verbatim
|
|
99
|
+
* would answer a question with a question). That test measured the wrong
|
|
100
|
+
* thing — a broad sample of this store's own question-shaped nodes showed
|
|
101
|
+
* the large majority (≈71%) have at least one predecessor, most of them a
|
|
102
|
+
* handful of generic, high-fan-out sentences that recur as an INCIDENTAL
|
|
103
|
+
* neighbour to dozens of otherwise-unrelated destinations (a SmolSent-
|
|
104
|
+
* style sentence-adjacency artifact, never naming or describing what
|
|
105
|
+
* follows). Traced live: "What is the capital of France?" — whose own
|
|
106
|
+
* forward edge unambiguously resolves to "The capital of France is
|
|
107
|
+
* Paris." — has exactly one such incidental predecessor ("Create an
|
|
108
|
+
* example of a types of questions a GPT model can answer.?"), wrongly
|
|
109
|
+
* read as disqualifying proof of "genuine entity."
|
|
110
|
+
*
|
|
111
|
+
* A plain forward-first swap (matching {@link project}'s universal
|
|
112
|
+
* priority) over-corrected: test/29's C2/C3 pin that a genuine entity
|
|
113
|
+
* analog (e.g. "Leonardo da Vinci", established by "The Mona Lisa was
|
|
114
|
+
* painted by Leonardo da Vinci.") must be seated by that establishing
|
|
115
|
+
* sentence, NOT by its own biography fact — voicing the bio leaks exactly
|
|
116
|
+
* what a comparison must keep out, and loses the embedded "Mona Lisa"
|
|
117
|
+
* term C3 relies on for a further hop.
|
|
118
|
+
*
|
|
119
|
+
* The distinguishing signal is content-addressed, not a count: a genuine
|
|
120
|
+
* establishing predecessor's bytes CONTAIN id's own bytes — it names or
|
|
121
|
+
* describes id ("...painted by Leonardo da Vinci." contains "Leonardo da
|
|
122
|
+
* Vinci"). An incidental adjacency predecessor never does — it merely
|
|
123
|
+
* preceded id in some unrelated document without ever mentioning it. No
|
|
124
|
+
* new tuned constant: containment is the same primitive `restatesQuery`
|
|
125
|
+
* and `dominates`-style checks already use throughout this codebase.
|
|
126
|
+
*
|
|
127
|
+
* `allowForward` (default true) gates the FORWARD branch specifically —
|
|
128
|
+
* see the call sites below: the DOMINANT is what the query is actually
|
|
129
|
+
* ASKING, so completing it forward is the whole point; an ANALOG is only
|
|
130
|
+
* being CITED for comparison; the query never asked about IT, so chasing
|
|
131
|
+
* its own further continuation drifts onto whatever coincidentally
|
|
132
|
+
* follows it in the corpus. Traced live: the analog "What is the capital
|
|
133
|
+
* of Japan?\nTokyo is the capital of Japan." is ALREADY a complete,
|
|
134
|
+
* self-answering unit (prevCount 0, so no establishing predecessor
|
|
135
|
+
* either) — its sole forward edge is "And what is the capital of the
|
|
136
|
+
* Moon?", an unrelated quiz question sharing nothing but corpus
|
|
137
|
+
* adjacency. With forward disallowed, an analog like this falls through
|
|
138
|
+
* to `fallback` — its own bytes, exactly the complete fact that made it a
|
|
139
|
+
* genuine analog in the first place. See
|
|
140
|
+
* test/41-seatofnode-direction.test.mjs and
|
|
141
|
+
* test/43-cast-analog-seat.test.mjs. */
|
|
142
|
+
export async function seatOfNode(
|
|
143
|
+
ctx: MindContext,
|
|
144
|
+
id: number,
|
|
145
|
+
guide: Vec | null | undefined,
|
|
146
|
+
fallback: Uint8Array,
|
|
147
|
+
allowForward = true,
|
|
148
|
+
): Promise<Uint8Array> {
|
|
149
|
+
const rev = ctx.store.prevFirst(id, hubBound(ctx));
|
|
150
|
+
if (rev.length > 0) {
|
|
151
|
+
const own = read(ctx, id);
|
|
152
|
+
const establishing = rev.some((p) => indexOf(read(ctx, p), own, 0) >= 0);
|
|
153
|
+
if (establishing) {
|
|
154
|
+
const back = reverseContext(ctx, id, guide, rev);
|
|
155
|
+
if (back !== null) return back;
|
|
156
|
+
}
|
|
157
|
+
}
|
|
158
|
+
// The "last resort, non-establishing reverse" fallback below is itself a
|
|
159
|
+
// LESS CERTAIN projection (the same tier as forward) — an analog
|
|
160
|
+
// (allowForward: false) must stop at `fallback` (its own bytes) here
|
|
161
|
+
// rather than fall back to a predecessor that already failed the
|
|
162
|
+
// establishing check just above.
|
|
163
|
+
if (!allowForward) return fallback;
|
|
164
|
+
const fwd = await follow(ctx, id, guide);
|
|
165
|
+
if (fwd !== null) return fwd;
|
|
166
|
+
return reverseContext(ctx, id, guide, rev) ?? fallback;
|
|
167
|
+
}
|
|
168
|
+
|
|
83
169
|
/** CAST's own entry gates, checked once here and reused by
|
|
84
170
|
/** The main CAST entry point. Given a query and its pre-computed pre.rec.sites,
|
|
85
171
|
* determine whether the query weaves together multiple independent learnt
|
|
@@ -297,7 +383,10 @@ export async function counterfactualTransfer(
|
|
|
297
383
|
const tail = proj.ctx.subarray(seat.cs);
|
|
298
384
|
let answer = await joinWithBridge(ctx, filler, tail);
|
|
299
385
|
const fwd = await follow(ctx, proj.anchor, qv);
|
|
300
|
-
if (
|
|
386
|
+
if (
|
|
387
|
+
fwd !== null && indexOf(answer, fwd, 0) < 0 &&
|
|
388
|
+
!restatesQuery(query, fwd)
|
|
389
|
+
) {
|
|
301
390
|
answer = concat2(answer, fwd);
|
|
302
391
|
}
|
|
303
392
|
ctx.trace?.step(
|
|
@@ -375,36 +464,14 @@ export async function counterfactualTransfer(
|
|
|
375
464
|
// seed-dependent failure where the climb ranks an exemplar above a
|
|
376
465
|
// person node and the person node is excluded from points by run-
|
|
377
466
|
// overlap trimming.
|
|
378
|
-
// The seat that establishes a candidate's role
|
|
379
|
-
|
|
380
|
-
|
|
381
|
-
// over the pre-refactor code: reverseContext also returns null when the
|
|
382
|
-
// picked context READS EMPTY (a dangling id degrades to zero bytes), so a
|
|
383
|
-
// corrupted-store read now falls back here instead of voicing a hollow
|
|
384
|
-
// seat into the comparison — the same "empty bytes are no grounding"
|
|
385
|
-
// invariant project() has always enforced.
|
|
386
|
-
// A node that only ever CONTINUES — an edge SOURCE with no predecessor —
|
|
387
|
-
// is a learnt CONTEXT (a stored question-shaped frame), not an entity.
|
|
388
|
-
// Voicing it verbatim answers a question with a question (the observed
|
|
389
|
-
// cast echo: two stored questions concatenated as an "answer"). The
|
|
390
|
-
// context's role is established by what it LEADS TO, so its seat is its
|
|
391
|
-
// own continuation — the fact — with the reverse projection and the raw
|
|
392
|
-
// bytes as the ordinary fallbacks for genuine entities.
|
|
393
|
-
const seatOfNode = async (
|
|
394
|
-
id: number,
|
|
395
|
-
fallback: Uint8Array,
|
|
396
|
-
): Promise<Uint8Array> => {
|
|
397
|
-
if (ctx.store.prevCount(id) === 0 && ctx.store.hasNext(id)) {
|
|
398
|
-
const fwd = await follow(ctx, id, qv);
|
|
399
|
-
if (fwd !== null) return fwd;
|
|
400
|
-
}
|
|
401
|
-
return reverseContext(ctx, id, qv) ?? fallback;
|
|
402
|
-
};
|
|
403
|
-
const seatOf = (p: Point): Promise<Uint8Array> => seatOfNode(p.anchor, p.ctx);
|
|
467
|
+
// The seat that establishes a candidate's role — see {@link seatOfNode}.
|
|
468
|
+
const seatOf = (p: Point, allowForward = true): Promise<Uint8Array> =>
|
|
469
|
+
seatOfNode(ctx, p.anchor, qv, p.ctx, allowForward);
|
|
404
470
|
interface AnalogCandidate {
|
|
405
471
|
anchor: number;
|
|
406
472
|
/** The point this candidate came from, or null when it is a nextOf
|
|
407
|
-
* descendant — then
|
|
473
|
+
* descendant — then its own bytes ARE the seat (already one meaningful
|
|
474
|
+
* hop from `src`; see the comparison gate below). */
|
|
408
475
|
point: Point | null;
|
|
409
476
|
/** For a nextOf descendant: the aligned point whose continuation edge
|
|
410
477
|
* named it. Its runs ARE the query evidence this analog rests on
|
|
@@ -447,8 +514,46 @@ export async function counterfactualTransfer(
|
|
|
447
514
|
}
|
|
448
515
|
let bestAnalog: AnalogCandidate | null = null;
|
|
449
516
|
let bestSim = 0;
|
|
517
|
+
let bestHalo = false;
|
|
518
|
+
// Whether the query itself NAMES a candidate. A directly aligned point
|
|
519
|
+
// is named by construction — its runs ARE query bytes. A hop-reached
|
|
520
|
+
// candidate is named when its own bytes contain the query text of an
|
|
521
|
+
// aligned run of the point whose continuation edge reached it (that
|
|
522
|
+
// alignment IS the query evidence the hop rests on — the same reading
|
|
523
|
+
// cmpAccounted already prices): "William Shakespeare", reached off
|
|
524
|
+
// "Macbeth was written by William Shakespeare.", contains the src's
|
|
525
|
+
// 12-byte aligned run " Shakespeare" — test/29 C2/C3. The run must span
|
|
526
|
+
// at least TWO perception windows (2·W, the same two-quantum floor
|
|
527
|
+
// CAST's own entry gate holds the whole query to): a single shared
|
|
528
|
+
// W-window is exactly the frame tier's own evidence quantum — the level
|
|
529
|
+
// "half the corpus" shares — and stopword scraps (" the ", "he b",
|
|
530
|
+
// 4–5 bytes) never reach two windows, while a genuinely named entity
|
|
531
|
+
// does. NOT the weave's usable()/frame filter: weave depth counts every
|
|
532
|
+
// ranked exemplar, so a query's own named entity recurring across
|
|
533
|
+
// exemplars ("Shakespeare" in Hamlet+Macbeth+…) is wrongly classified as
|
|
534
|
+
// frame — measured live, it silently disqualified C3's genuine analog.
|
|
535
|
+
const namedByQuery = (c: AnalogCandidate): boolean => {
|
|
536
|
+
if (c.point !== null) return true;
|
|
537
|
+
const bytes = read(ctx, c.anchor);
|
|
538
|
+
return c.src.runs.some((r) =>
|
|
539
|
+
r.qe - r.qs >= 2 * quantum &&
|
|
540
|
+
indexOf(bytes, query.subarray(r.qs, r.qe), 0) >= 0
|
|
541
|
+
);
|
|
542
|
+
};
|
|
543
|
+
// Whether any committed root's consensus vote clears the SAME trust bar
|
|
544
|
+
// recallByResonance applies before grounding through a climb root:
|
|
545
|
+
// consensusFloor(N) = ln(N) + 1/2. The climb's FIRST root is
|
|
546
|
+
// deliberately floor-free (attention.ts: "the dominant one always
|
|
547
|
+
// grounds") — fine for ORIENTING mechanisms, not for voicing learnt
|
|
548
|
+
// content the query never asked about. Computed once here; both the
|
|
549
|
+
// hub fallback below and the comparison gate consume it.
|
|
550
|
+
const rootTrusted = roots.some((r) => r.vote >= consensusFloor(corpusN(ctx)));
|
|
450
551
|
for (const c of analogs) {
|
|
451
|
-
const sim = await analogyStrength(
|
|
552
|
+
const { score: sim, halo } = await analogyStrength(
|
|
553
|
+
ctx,
|
|
554
|
+
dominant.anchor,
|
|
555
|
+
c.anchor,
|
|
556
|
+
);
|
|
452
557
|
ctx.trace?.step(
|
|
453
558
|
"tryAnalog",
|
|
454
559
|
[
|
|
@@ -456,11 +561,12 @@ export async function counterfactualTransfer(
|
|
|
456
561
|
rNode(ctx, c.anchor, "candidate", sim),
|
|
457
562
|
],
|
|
458
563
|
[],
|
|
459
|
-
`analogy strength ${sim.toFixed(4)}`,
|
|
564
|
+
`analogy strength ${sim.toFixed(4)}${halo ? " (halo tier)" : ""}`,
|
|
460
565
|
);
|
|
461
566
|
if (sim > bestSim) {
|
|
462
567
|
bestSim = sim;
|
|
463
568
|
bestAnalog = c;
|
|
569
|
+
bestHalo = halo;
|
|
464
570
|
}
|
|
465
571
|
}
|
|
466
572
|
// When every candidate fails the similarity gates (halo company — now
|
|
@@ -489,6 +595,17 @@ export async function counterfactualTransfer(
|
|
|
489
595
|
let hubMass = -1;
|
|
490
596
|
const fanClamp = hubBound(ctx) + 1;
|
|
491
597
|
for (const c of analogs) {
|
|
598
|
+
// A fallback comparison carries NO similarity evidence at all. Its
|
|
599
|
+
// honesty rests on the grounding decider discounting it against
|
|
600
|
+
// richer candidates (the design note below) — an assumption that
|
|
601
|
+
// holds only when the climb itself settled on this query with real
|
|
602
|
+
// evidence. Under a root the consensus floor does not trust, an
|
|
603
|
+
// unnamed, hop-reached hub is pure corpus adjacency: refusing it is
|
|
604
|
+
// what kept the live wrong echo silent. A hub the query itself
|
|
605
|
+
// NAMED stays eligible either way (test/29 C2/C3's "William
|
|
606
|
+
// Shakespeare"); an unnamed one under a TRUSTED root stays eligible
|
|
607
|
+
// too (test/33 1b's deliberately weak second candidate).
|
|
608
|
+
if (!rootTrusted && !namedByQuery(c)) continue;
|
|
492
609
|
// Evidence clamped at the hub bound: beyond √N + 1 the exact fan-out
|
|
493
610
|
// no longer discriminates (every mega-hub ties at the clamp), and
|
|
494
611
|
// counting it exactly would require the corpus-sized read.
|
|
@@ -538,10 +655,111 @@ export async function counterfactualTransfer(
|
|
|
538
655
|
// climb's own forest, never tuned; substitution/redirection stay
|
|
539
656
|
// unaffected — they orient around a displaced seat, not a whole-topic
|
|
540
657
|
// analogy.
|
|
658
|
+
//
|
|
659
|
+
// roots.length <= 1 is a PROXY for "the query is about one thing" — it is
|
|
660
|
+
// only as good as the climb's own root-commitment, which depends on
|
|
661
|
+
// recognise() having found something to commit a root TO. When the
|
|
662
|
+
// query's newest content genuinely isn't recognised (not boundary noise —
|
|
663
|
+
// real, uncommitted content; see the session's own investigation of the
|
|
664
|
+
// France→Spain live trace), the climb under-commits roots and this proxy
|
|
665
|
+
// is fooled: comparison looks licensed to treat the query as one topic
|
|
666
|
+
// when it is not.
|
|
667
|
+
//
|
|
668
|
+
// The direct check is the SAME accounted spans comparison is about to
|
|
669
|
+
// cite as its evidence: unexplainedSpans (rationale.ts, the same gap
|
|
670
|
+
// computation the trace's own `unexplained` diagnostic uses) names every
|
|
671
|
+
// stretch of the query NEITHER the dominant NOR the analog's evidence
|
|
672
|
+
// touches. A short comparison query ("How is ice like steel?") legitimately
|
|
673
|
+
// accounts for only its two short entity spans — the surrounding "How is
|
|
674
|
+
// ... like ...?" framing is real but SHORT, split into several small gaps,
|
|
675
|
+
// none of them the bulk of the query. The live bug's shape is different in
|
|
676
|
+
// kind, not degree: ONE contiguous, substantial gap — a whole second
|
|
677
|
+
// question the query added that comparison's two spans never touch at all.
|
|
678
|
+
//
|
|
679
|
+
// Two bars, both derived, neither tuned:
|
|
680
|
+
// • the largest gap must not DOMINATE the whole query (the same
|
|
681
|
+
// predicate CAST's own frame gate uses) — rules out a gap that is
|
|
682
|
+
// most of the query outright;
|
|
683
|
+
// • the largest gap must be SMALLER than the dominant's own established
|
|
684
|
+
// context. A gap can't be dismissed as mere connective framing once
|
|
685
|
+
// it is at least as large as the topic being compared FROM — at that
|
|
686
|
+
// scale it isn't glue between two named things, it's substantial
|
|
687
|
+
// enough to be a second topic in its own right. This is what
|
|
688
|
+
// actually separates the live bug (a 47-byte gap against a 30-byte
|
|
689
|
+
// dominant — the ignored content is bigger than the topic itself)
|
|
690
|
+
// from ordinary short comparisons (a 9-byte gap against an 11-byte
|
|
691
|
+
// dominant — the gap is smaller than what's being compared): the two
|
|
692
|
+
// cases land on the same side of "half the query" often enough
|
|
693
|
+
// (both can exceed or clear it) that the query-relative bar alone
|
|
694
|
+
// does not reliably separate them — the topic-relative scale does.
|
|
695
|
+
const cmpAccounted: Array<[number, number]> = bestAnalog !== null
|
|
696
|
+
? [...runSpans(dominant), ...runSpans(bestAnalog.point ?? bestAnalog.src)]
|
|
697
|
+
: [];
|
|
698
|
+
const cmpGaps = unexplainedSpans(query.length, cmpAccounted);
|
|
699
|
+
const cmpMaxGap = cmpGaps.reduce((n, [s, e]) => Math.max(n, e - s), 0);
|
|
700
|
+
// An analog that is not itself a directly ALIGNED point (point !== null —
|
|
701
|
+
// its own runs are query bytes, the query NAMED it) was only reached
|
|
702
|
+
// through a continuation hop or the structural-hub fallback. Voicing
|
|
703
|
+
// learnt content the query never named is the same act recallByResonance
|
|
704
|
+
// refuses to perform through a climb root whose consensus vote is below
|
|
705
|
+
// consensusFloor(N) = ln(N) + 1/2 (recall.ts's minVote), so comparison
|
|
706
|
+
// holds the climb to that SAME bar before citing a hop-reached analog:
|
|
707
|
+
// some committed root must clear the floor. The climb's FIRST root is
|
|
708
|
+
// deliberately floor-free (attention.ts: "the dominant one always
|
|
709
|
+
// grounds") — fine for ORIENTING mechanisms, not for transferring
|
|
710
|
+
// unnamed content through. The live bug this gates (real trained store,
|
|
711
|
+
// 325k edge sources, floor 13.2): the query's stopword scraps pooled a
|
|
712
|
+
// 1.92 vote that committed an unrelated haiku exemplar as the sole root,
|
|
713
|
+
// and comparison voiced that exemplar's continuation through a
|
|
714
|
+
// hop-reached analog while every other mechanism honestly refused. A
|
|
715
|
+
// directly aligned analog needs no floor — the query's own bytes are its
|
|
716
|
+
// evidence (test/29 C1's "Steel is hard" for "How is ice like steel?").
|
|
717
|
+
// See test/50-cast-analog-consensus-floor.
|
|
718
|
+
// A HALO-tier best analog needs neither: its similarity already cleared
|
|
719
|
+
// significanceBar-gated distributional company (analogyStrength's
|
|
720
|
+
// `halo`) — genuine evidence in its own right, the very case the halo
|
|
721
|
+
// gate exists for (test/33 1b's nickname-corroborated analog). Only a
|
|
722
|
+
// FRAME-tier or fallback analog — whose "similarity" is an unbarred
|
|
723
|
+
// coverage fraction or nothing — needs the query's naming or the climb's
|
|
724
|
+
// trust.
|
|
725
|
+
const analogNamed = bestAnalog !== null && namedByQuery(bestAnalog);
|
|
726
|
+
// NOTE — two further gates were tried here and empirically REFUTED,
|
|
727
|
+
// recorded so they are not re-tried:
|
|
728
|
+
// • dominant self-coverage (dominant's aligned runs must dominate its
|
|
729
|
+
// own ctx): legitimate dominants sit at the same coverage as junk
|
|
730
|
+
// ones ("The Mona Lisa was painted by…" 16/47 vs the live junk
|
|
731
|
+
// haiku ~10/54) — no separation.
|
|
732
|
+
// • denying the shared-frame similarity tier to hop-reached analogs:
|
|
733
|
+
// semantically right in isolation, but it merely promoted the next
|
|
734
|
+
// junk candidate — an ALIGNED scrap-matched point ("The affluence…",
|
|
735
|
+
// frame 0.157) — into bestAnalog on the live store, and the aligned
|
|
736
|
+
// configuration is byte-structurally IDENTICAL to test/29 C1's
|
|
737
|
+
// legitimate one ("Steel is hard", frame 0.364): every derived
|
|
738
|
+
// local separator measured (run length, site overlap, frame
|
|
739
|
+
// query-containment, weave-usable classification) falls on the same
|
|
740
|
+
// side for both. Only corpus-scale consensus separates them, which
|
|
741
|
+
// is exactly what `rootTrusted` prices.
|
|
742
|
+
// FRAME-tier evidence under an UNTRUSTED root is comparison's weakest
|
|
743
|
+
// licence (an unbarred coverage fraction, a climb the consensus floor
|
|
744
|
+
// does not trust). There it is additionally held to the IGNORED-KNOWN
|
|
745
|
+
// principle (dismissedKnownContent, bridge.ts): the two analogs' aligned
|
|
746
|
+
// runs must account for every STORED window of the query. This is the
|
|
747
|
+
// byte-structural separator the refuted-gates note below could not find
|
|
748
|
+
// locally: a legitimate small-corpus comparison ("How is ice like
|
|
749
|
+
// steel?") leaves only UNATTESTED spans ("How ", " like ") unexplained,
|
|
750
|
+
// while a scrap-matched junk pair leaves the query's own trained content
|
|
751
|
+
// ("…songs…times…", "…planet…sun.") dismissed as gaps. Halo-tier and
|
|
752
|
+
// trusted-root comparisons are exempt — their evidence already stands.
|
|
753
|
+
const cmpDismisses = !(bestHalo || rootTrusted) &&
|
|
754
|
+
dismissedKnownContent(ctx, query, cmpAccounted);
|
|
541
755
|
if (
|
|
542
756
|
bestAnalog !== null &&
|
|
757
|
+
(bestHalo || analogNamed || rootTrusted) &&
|
|
758
|
+
!cmpDismisses &&
|
|
543
759
|
dominant.ctx.length <= query.length &&
|
|
544
|
-
roots.length <= 1
|
|
760
|
+
roots.length <= 1 &&
|
|
761
|
+
!dominates(cmpMaxGap, query.length) &&
|
|
762
|
+
cmpMaxGap < dominant.ctx.length
|
|
545
763
|
) {
|
|
546
764
|
ctx.trace?.step(
|
|
547
765
|
"validateAnalogy",
|
|
@@ -553,9 +771,21 @@ export async function counterfactualTransfer(
|
|
|
553
771
|
"the two structures keep distributional company beyond chance — genuine analogs",
|
|
554
772
|
);
|
|
555
773
|
const a = await seatOf(dominant);
|
|
774
|
+
// The analog is only being CITED for comparison — the query never asked
|
|
775
|
+
// about it — so its seat never chases a FORWARD continuation (see
|
|
776
|
+
// seatOfNode's `allowForward`): only reverse (if a predecessor genuinely
|
|
777
|
+
// establishes it) or its own bytes. A DIRECTLY aligned point
|
|
778
|
+
// (bestAnalog.point !== null) still goes through seatOfNode for that
|
|
779
|
+
// reverse check (a bare entity NAME like "Leonardo da Vinci" needs it —
|
|
780
|
+
// test/29's C2/C3). A nextOf DESCENDANT (point === null) was already
|
|
781
|
+
// reached by following ONE meaningful hop off another aligned point (the
|
|
782
|
+
// alignment loop above: "its nextOf is the hub... and the hub's own
|
|
783
|
+
// [...] context will be the seat") — its own bytes ARE that seat
|
|
784
|
+
// directly, with no predecessor to even check (it was found by a
|
|
785
|
+
// forward edge, not matched in the query).
|
|
556
786
|
const b = bestAnalog.point !== null
|
|
557
|
-
? await seatOf(bestAnalog.point)
|
|
558
|
-
:
|
|
787
|
+
? await seatOf(bestAnalog.point, false)
|
|
788
|
+
: read(ctx, bestAnalog.anchor);
|
|
559
789
|
const answer = await joinWithBridge(ctx, a, b);
|
|
560
790
|
record(
|
|
561
791
|
answer,
|
|
@@ -568,10 +798,32 @@ export async function counterfactualTransfer(
|
|
|
568
798
|
// when it was an aligned point, else the source point whose
|
|
569
799
|
// continuation edge reached it (that alignment IS the query evidence
|
|
570
800
|
// the hop rests on).
|
|
801
|
+
cmpAccounted,
|
|
802
|
+
);
|
|
803
|
+
} else if (
|
|
804
|
+
bestAnalog !== null &&
|
|
805
|
+
dominant.ctx.length <= query.length &&
|
|
806
|
+
roots.length <= 1
|
|
807
|
+
) {
|
|
808
|
+
ctx.trace?.step(
|
|
809
|
+
"validateAnalogy",
|
|
571
810
|
[
|
|
572
|
-
|
|
573
|
-
|
|
811
|
+
rNode(ctx, dominant.anchor, "analog", bestSim),
|
|
812
|
+
rNode(ctx, bestAnalog.anchor, "analog", bestSim),
|
|
574
813
|
],
|
|
814
|
+
[],
|
|
815
|
+
!(bestHalo || analogNamed || rootTrusted)
|
|
816
|
+
? `the best analog carries no halo-tier company evidence, was never ` +
|
|
817
|
+
`named by the query, and no committed root's consensus vote ` +
|
|
818
|
+
`clears the floor, so comparison refuses to voice it`
|
|
819
|
+
: cmpDismisses
|
|
820
|
+
? `a frame-tier analog under an untrusted root dismisses stored ` +
|
|
821
|
+
`query content its alignment never accounted for — comparison ` +
|
|
822
|
+
`refuses to ignore what the store knows`
|
|
823
|
+
: `comparison's own accounted evidence leaves a ${cmpMaxGap}-byte gap in ` +
|
|
824
|
+
`a ${query.length}-byte query against a ${dominant.ctx.length}-byte ` +
|
|
825
|
+
`dominant — too large to be mere framing — so it refuses rather ` +
|
|
826
|
+
`than paper over it with an analog the query never asked about`,
|
|
575
827
|
);
|
|
576
828
|
}
|
|
577
829
|
t?.done(
|