@hviana/sema 0.2.4 → 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.
@@ -8,14 +8,39 @@
8
8
  // attention for the rest of the pipeline to follow.
9
9
  import { isChunk } from "../sema.js";
10
10
  import { lightestDerivation, } from "../derive/src/index.js";
11
- import { consensusFloor, dominates, estimatorNoise } from "../geometry.js";
11
+ import { composeStructuralGist, consensusFloor, dominates, estimatorNoise, } from "../geometry.js";
12
12
  import { foldTree, gistOf, latin1Key, perceive, read } from "./primitives.js";
13
13
  import { recognise } from "./recognition.js";
14
14
  import { leafIdRun } from "./canonical.js";
15
- import { corpusN, edgeAncestors } from "./traverse.js";
16
- import { cachedRead, junctionContainersFrom, junctionSeeds, junctionSynonyms, walkCache, } from "./junction.js";
15
+ import { corpusN, edgeAncestors, hubBound } from "./traverse.js";
16
+ import { cachedRead, junctionContainersFrom, junctionSeeds, junctionSynonyms, loadJunctionSynonymSides, walkCache, } from "./junction.js";
17
17
  import { indexOf } from "../bytes.js";
18
18
  import { rItem, rNode, traceDerivation } from "./trace.js";
19
+ function newTraceDraft(perceivedCount) {
20
+ return {
21
+ perceivedCount,
22
+ regions: [],
23
+ crossRegionJunctionVotes: [],
24
+ supersededOrdinaryVotes: 0,
25
+ anchors: [],
26
+ };
27
+ }
28
+ /** Serialise the shared `reachMemo` into the plain, authoritative saturation
29
+ * profile (spec §5) — every distinct node any tier's `edgeAncestors` call
30
+ * climbed from during this response, in insertion (first-consulted) order. */
31
+ function serialiseReaches(reachMemo) {
32
+ const out = [];
33
+ for (const [node, r] of reachMemo) {
34
+ out.push({
35
+ node,
36
+ roots: [...r.roots],
37
+ contextsReached: r.contextsReached,
38
+ saturated: r.saturated,
39
+ ...(r.saturation ? { saturation: r.saturation } : {}),
40
+ });
41
+ }
42
+ return out;
43
+ }
19
44
  // ── Public entry points ───────────────────────────────────────────────────
20
45
  /** Climb the query's perceived byte regions up the structural DAG via
21
46
  * resonance, pool the evidence, and return only the ROOT points of
@@ -47,8 +72,21 @@ export async function climbAttentionAll(ctx, query, k, mode = "inverse") {
47
72
  }
48
73
  const hit = byRead.get(modeKey);
49
74
  if (hit !== undefined) {
75
+ // Cache-hit exit (spec §9): the abbreviated payload shape — only what
76
+ // is actually stored in the cached AttentionRead is reported. No
77
+ // candidate, reach, saturation, pooling or anchor detail is fabricated
78
+ // (that per-region detail was never retained by the memo).
79
+ const data = ctx.trace
80
+ ? {
81
+ version: 1,
82
+ cache: { hit: true, detailAvailable: false },
83
+ config: { annK: k, crossRegionProbeLimit: k, mode },
84
+ candidates: { perceived: 0, recognised: 0, total: 0 },
85
+ result: hit,
86
+ }
87
+ : undefined;
50
88
  ctx.trace?.step("climbConsensus", [rItem(query, "query")], hit.roots.map((r) => rNode(ctx, r.anchor, "anchor", r.vote)), `(cached) consensus already computed for this query — ` +
51
- `${hit.roots.length} point(s) of attention`);
89
+ `${hit.roots.length} point(s) of attention`, undefined, data);
52
90
  return hit;
53
91
  }
54
92
  const read = await computeAttention(ctx, query, k, mode);
@@ -60,6 +98,7 @@ export async function climbAttentionAll(ctx, query, k, mode = "inverse") {
60
98
  // ── Pipeline ──────────────────────────────────────────────────────────────
61
99
  export async function computeAttention(ctx, query, k, mode) {
62
100
  const regions = collectRegions(ctx, query);
101
+ const perceivedCount = regions.length;
63
102
  // Recognised sites carry structural evidence that perceived sub-regions
64
103
  // miss: a word crossing a W-boundary is split into chunks whose partial
65
104
  // gists may not resonate distinctively, but the SITE (content-addressed,
@@ -78,20 +117,34 @@ export async function computeAttention(ctx, query, k, mode) {
78
117
  known: true, // a recognised site IS a stored form
79
118
  });
80
119
  }
81
- if (regions.length === 0)
120
+ // The trace draft (spec §9): allocated ONLY when a trace was requested —
121
+ // every downstream consumer gates its own writes on `td?` / `if (td)`, so
122
+ // an untraced climb pays zero allocation for this instrumentation.
123
+ const td = ctx.trace
124
+ ? newTraceDraft(perceivedCount)
125
+ : undefined;
126
+ const cfg0 = {
127
+ k,
128
+ mode,
129
+ perceivedCount,
130
+ totalRegions: regions.length,
131
+ };
132
+ if (regions.length === 0) {
133
+ traceAttention(ctx, [], [], [], undefined, td, cfg0);
82
134
  return { roots: [], ranked: [] };
135
+ }
83
136
  const N = corpusN(ctx);
84
137
  // One climb per distinct anchor for the WHOLE query: regions sharing a
85
138
  // chunk, and canonicalChunkId's prefix probes, all hit this memo instead of
86
139
  // re-reading the anchor's full edge fan-out from the store.
87
140
  const reachMemo = new Map();
88
- const rvs = await voteRegions(ctx, query, regions, k, mode, N, reachMemo);
141
+ const rvs = await voteRegions(ctx, query, regions, k, mode, N, reachMemo, td);
89
142
  // ── Cross-region: DIRECT region-to-region interaction ─────────────────
90
143
  // Two regions whose individual climbs land on DIFFERENT contexts leave
91
144
  // their JOINT context — the learnt whole that contains BOTH — with no
92
145
  // vote. crossRegionVotes recovers it by the bridge's content-addressed
93
146
  // junction ascent (see the note above the function).
94
- const cross = await crossRegionVotes(ctx, query, regions, rvs, k, N, reachMemo);
147
+ const cross = await crossRegionVotes(ctx, query, regions, rvs, k, N, reachMemo, td);
95
148
  // A vote SUPERSEDED by exact joint evidence (its bytes literally live
96
149
  // inside the joint container, yet it climbed elsewhere — grid aliasing)
97
150
  // is dropped, not down-weighted: the joint container explains it away.
@@ -101,14 +154,33 @@ export async function computeAttention(ctx, query, k, mode) {
101
154
  ...cross.votes,
102
155
  ]
103
156
  : rvs.votes;
157
+ // Mark, on the per-region trace, the source region of every superseded
158
+ // ordinary vote (spec §4's final rule) — an explicit pass over the exact
159
+ // set crossRegionVotes' explaining-away logic removed, never inferred
160
+ // from `absorbed`.
161
+ if (td && cross.superseded.size > 0) {
162
+ for (const rv of cross.superseded) {
163
+ const region = td.regions.find((r) => r.span[0] === rv.start && r.span[1] === rv.end);
164
+ if (region)
165
+ region.superseded = true;
166
+ }
167
+ }
104
168
  // ──────────────────────────────────────────────────────────────────────
169
+ const cfg = { ...cfg0, N, reachMemo };
105
170
  if (allVotes.length === 0) {
106
- traceAttention(ctx, regions, rvs.voters, []);
171
+ traceAttention(ctx, regions, rvs.voters, [], undefined, td, cfg);
107
172
  return { roots: [], ranked: [] };
108
173
  }
109
174
  const sat = detectSaturated(ctx, regions, rvs.saturated);
110
- const pooled = poolVotes(ctx, allVotes, sat, N);
111
- return commitVotes(ctx, pooled, sat, regions, rvs.voters, N);
175
+ if (td) {
176
+ td.saturation = {
177
+ regionIntervals: sat.intervals.map((iv) => ({ ...iv })),
178
+ hasLeading: sat.hasLeading,
179
+ leadingEnd: sat.leadingEnd,
180
+ };
181
+ }
182
+ const pooled = poolVotes(ctx, allVotes, sat, N, td);
183
+ return commitVotes(ctx, pooled, sat, regions, rvs.voters, N, td, cfg);
112
184
  }
113
185
  export function collectRegions(ctx, query) {
114
186
  const regions = [];
@@ -140,12 +212,41 @@ export function collectRegions(ctx, query) {
140
212
  });
141
213
  return regions;
142
214
  }
143
- export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo) {
215
+ export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo, td) {
144
216
  const regionSaturated = new Array(regions.length).fill(false);
145
217
  const regionVotes = [];
146
218
  const regionVoter = ctx.trace ? regions.map(() => null) : [];
147
219
  for (let ri = 0; ri < regions.length; ri++) {
148
220
  const { v, start, end, chunk, known } = regions[ri];
221
+ // Trace-only bookkeeping for this region — allocated only under `td`
222
+ // (i.e. only when ctx.trace is set); see ConsensusRegionTrace/
223
+ // RegionOutcome (spec §4). `examinedIds` tracks distinct ANN hits
224
+ // whose edgeAncestors reach was actually CONSULTED here (not merely
225
+ // returned by resonate) — the fallback/margin loops below add to it.
226
+ const examinedIds = td ? new Set() : undefined;
227
+ let annQueried = false;
228
+ let fallbackKind;
229
+ const recordRegion = (outcome, extra = {}) => {
230
+ if (!td)
231
+ return;
232
+ td.regions[ri] = {
233
+ index: ri,
234
+ source: ri < td.perceivedCount ? "perceived" : "recognised",
235
+ span: [start, end],
236
+ chunk,
237
+ known,
238
+ canonicalId: canonicalId ?? undefined,
239
+ canonicalUsable,
240
+ canonicalFailed,
241
+ annQueried,
242
+ annHitsReturned: hits ? hits.length : 0,
243
+ annHitsExamined: examinedIds ? examinedIds.size : 0,
244
+ outcome,
245
+ ordinaryVoteProduced: outcome === "voted",
246
+ superseded: false,
247
+ ...extra,
248
+ };
249
+ };
149
250
  // EXACT-FIRST: a chunk whose canonical anchor is content-addressed needs
150
251
  // no estimator — identity is exact, so its score is 1 BY DEFINITION (the
151
252
  // estimated cosine of a form with itself, minus quantisation noise, and
@@ -164,23 +265,35 @@ export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo) {
164
265
  (ctx.store.hasParents(canonicalId) ||
165
266
  ctx.store.hasContainers(canonicalId));
166
267
  let hits = null;
167
- const ensureHits = async () => hits ??= await ctx.store.resonate(v, k);
268
+ const ensureHits = async () => {
269
+ if (hits === null) {
270
+ hits = await ctx.store.resonate(v, k);
271
+ annQueried = true;
272
+ }
273
+ return hits;
274
+ };
168
275
  const canonicalFailed = chunk && canonicalId === null;
169
276
  let voterId;
170
277
  let score;
171
278
  let scoreId; // the node the score was measured against
279
+ let selectedSource;
172
280
  if (canonicalUsable) {
173
281
  voterId = canonicalId;
174
282
  score = 1;
175
283
  scoreId = canonicalId;
284
+ selectedSource = "canonical";
176
285
  }
177
286
  else {
178
287
  const h = await ensureHits();
179
- if (h.length === 0)
288
+ if (h.length === 0) {
289
+ recordRegion("no-ann-hit");
180
290
  continue;
291
+ }
181
292
  voterId = h[0].id;
182
293
  score = h[0].score;
183
294
  scoreId = h[0].id;
295
+ selectedSource = "ann";
296
+ examinedIds?.add(voterId);
184
297
  }
185
298
  let reach = edgeAncestors(ctx, voterId, N, reachMemo);
186
299
  // A region's vote must not die with the TOP hit: `hits[1..k]` were
@@ -194,12 +307,15 @@ export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo) {
194
307
  if (h.id === voterId)
195
308
  continue;
196
309
  const r2 = edgeAncestors(ctx, h.id, N, reachMemo);
310
+ examinedIds?.add(h.id);
197
311
  if (r2.saturated || r2.roots.length > 0) {
198
312
  ctx.trace?.step("anchorFallback", [rNode(ctx, voterId, "orphan-anchor", score)], [rNode(ctx, h.id, "anchor", h.score)], "the top-ranked anchor climbs to no context — a lower-ranked hit votes instead");
199
313
  reach = r2;
200
314
  voterId = h.id;
201
315
  score = h.score;
202
316
  scoreId = h.id;
317
+ selectedSource = "ann";
318
+ fallbackKind = "orphan";
203
319
  break;
204
320
  }
205
321
  }
@@ -226,28 +342,56 @@ export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo) {
226
342
  if (h.score < score - band)
227
343
  break; // hits are nearest-first
228
344
  const r2 = edgeAncestors(ctx, h.id, N, reachMemo);
345
+ examinedIds?.add(h.id);
229
346
  if (!r2.saturated && r2.roots.length > 0) {
230
347
  ctx.trace?.step("anchorFallback", [rNode(ctx, voterId, "saturated-anchor", score)], [rNode(ctx, h.id, "anchor", h.score)], "the top-ranked anchor is a saturated hub tied within estimator noise — the tied hit votes instead");
231
348
  reach = r2;
232
349
  voterId = h.id;
233
350
  score = h.score;
234
351
  scoreId = h.id;
352
+ selectedSource = "ann";
353
+ fallbackKind = "saturated-tie";
235
354
  break;
236
355
  }
237
356
  }
238
357
  }
239
358
  regionSaturated[ri] = reach.saturated;
240
- if (reach.roots.length === 0)
359
+ const selected = !td
360
+ ? undefined
361
+ : (() => {
362
+ const rank = selectedSource === "ann"
363
+ ? hits?.findIndex((h) => h.id === voterId)
364
+ : undefined;
365
+ return {
366
+ source: selectedSource,
367
+ node: voterId,
368
+ score,
369
+ ...(rank !== undefined ? { rank } : {}),
370
+ ...(fallbackKind ? { fallback: fallbackKind } : {}),
371
+ };
372
+ })();
373
+ if (reach.roots.length === 0) {
374
+ recordRegion("no-structural-reach", { selected, reachNode: voterId });
241
375
  continue;
242
- if (reach.saturated)
376
+ }
377
+ if (reach.saturated) {
378
+ recordRegion("saturated-abstention", { selected, reachNode: voterId });
243
379
  continue;
380
+ }
244
381
  // One IDF per region — dfWeight() and the focus weight used to compute
245
382
  // the same logarithm independently.
246
383
  const idf = Math.log(N / Math.max(1, reach.contextsReached));
247
384
  const df = Math.log(1 + reach.contextsReached);
248
385
  const wf = mode === "direct" ? df : mode === "combined" ? idf + df : idf;
249
- if (wf <= 0)
386
+ if (wf <= 0) {
387
+ recordRegion("nonpositive-df-weight", {
388
+ selected,
389
+ reachNode: voterId,
390
+ idf,
391
+ dfWeight: wf,
392
+ });
250
393
  continue;
394
+ }
251
395
  // CONTRASTIVE-MARGIN GATE — the compensation the linear (byte-proportional)
252
396
  // fold demands, applied to APPROXIMATE evidence only. Under the linear
253
397
  // fold a resonance score reads "fraction of aligned shared bytes", so a
@@ -271,12 +415,14 @@ export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo) {
271
415
  // (reordered / paraphrased queries) below the floor so they grounded
272
416
  // nothing. Gating at the noise floor keeps frame-echo suppression (a frame
273
417
  // region's margin ≈ 0 is gated out) without penalising honest evidence.
418
+ let contrastiveMargin;
274
419
  if (!known) {
275
420
  let margin = score;
276
421
  for (const h of await ensureHits()) {
277
422
  if (h.id === voterId)
278
423
  continue;
279
424
  const r2 = edgeAncestors(ctx, h.id, N, reachMemo);
425
+ examinedIds?.add(h.id);
280
426
  if (r2.saturated || r2.roots.length === 0)
281
427
  continue; // concludes nothing
282
428
  if (sameRoots(r2.roots, reach.roots))
@@ -284,8 +430,19 @@ export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo) {
284
430
  margin = score - h.score; // hits are nearest-first: the best rival
285
431
  break;
286
432
  }
287
- if (margin <= estimatorNoise(ctx.store.D))
433
+ contrastiveMargin = margin;
434
+ const noiseFloor = estimatorNoise(ctx.store.D);
435
+ if (margin <= noiseFloor) {
436
+ recordRegion("contrastive-margin-rejection", {
437
+ selected,
438
+ reachNode: voterId,
439
+ idf,
440
+ dfWeight: wf,
441
+ contrastiveMargin: margin,
442
+ contrastiveNoiseFloor: noiseFloor,
443
+ });
288
444
  continue;
445
+ }
289
446
  }
290
447
  // MUTUAL-EXPLANATION WEIGHT (angle + magnitude). Under the linear fold
291
448
  // cos = shared/(‖r‖·‖h‖) with ‖·‖² = content bytes, so the old score²
@@ -322,6 +479,21 @@ export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo) {
322
479
  if (ctx.trace) {
323
480
  regionVoter[ri] = { id: voterId, score, w: wf };
324
481
  }
482
+ recordRegion("voted", {
483
+ selected,
484
+ reachNode: voterId,
485
+ idf,
486
+ dfWeight: wf,
487
+ ...(contrastiveMargin !== undefined
488
+ ? {
489
+ contrastiveMargin,
490
+ contrastiveNoiseFloor: estimatorNoise(ctx.store.D),
491
+ }
492
+ : {}),
493
+ mutualWeight: mutual,
494
+ voteWeightPerRoot: w,
495
+ focusWeightPerRoot: wFocus,
496
+ });
325
497
  }
326
498
  return {
327
499
  votes: regionVotes,
@@ -341,7 +513,7 @@ export async function voteRegions(ctx, query, regions, k, mode, N, reachMemo) {
341
513
  * merely logs alongside it. `votesIdf`/`support` are the same two
342
514
  * read-outs {@link commitVotes} always gated on; only how they accumulate
343
515
  * changed. */
344
- export function poolVotes(ctx, regionVotes, sat, N) {
516
+ export function poolVotes(ctx, regionVotes, sat, N, td) {
345
517
  const eligible = [];
346
518
  for (let ri = 0; ri < regionVotes.length; ri++) {
347
519
  const rv = regionVotes[ri];
@@ -351,6 +523,13 @@ export function poolVotes(ctx, regionVotes, sat, N) {
351
523
  }
352
524
  eligible.push(ri);
353
525
  }
526
+ if (td) {
527
+ td.pooling = {
528
+ inputVotes: regionVotes.length,
529
+ eligibleVotes: eligible.length,
530
+ saturationMaskedVotes: regionVotes.length - eligible.length,
531
+ };
532
+ }
354
533
  const key = (it) => it.kind === "region"
355
534
  ? `r${it.ri}`
356
535
  : it.kind === "anchor"
@@ -495,10 +674,10 @@ function countClusters(spans, W) {
495
674
  }
496
675
  return clusters;
497
676
  }
498
- export function commitVotes(ctx, pooled, sat, regions, regionVoter, N) {
677
+ export function commitVotes(ctx, pooled, sat, regions, regionVoter, N, td, cfg) {
499
678
  const { votes, votesIdf, support, regionSupport, regionSpans, steps } = pooled;
500
679
  if (votes.size === 0) {
501
- traceAttention(ctx, regions, regionVoter, [], steps);
680
+ traceAttention(ctx, regions, regionVoter, [], steps, td, cfg);
502
681
  return { roots: [], ranked: [] };
503
682
  }
504
683
  // SCALE-INVARIANT confidence — see Attention.breadth's doc. regions.length
@@ -534,23 +713,92 @@ export function commitVotes(ctx, pooled, sat, regions, regionVoter, N) {
534
713
  const floor = consensusFloor(N);
535
714
  const placed = [];
536
715
  const roots = [];
537
- for (const point of ranked) {
716
+ const recordAnchor = (point, rank, status, dominant, passesNaturalBreak, passesConsensusFloor, pastLeadingSaturation, rejectionReasons) => {
717
+ if (!td)
718
+ return;
719
+ td.anchors.push({
720
+ anchor: point.anchor,
721
+ rank,
722
+ pooledVote: point.vote,
723
+ idfVote: votesIdf.get(point.anchor) ?? 0,
724
+ candidateBreadth: regions.length,
725
+ contributingVotes: regionSpans.get(point.anchor)?.length ?? 0,
726
+ contributingEvidence: regionSupport.get(point.anchor) ?? 0,
727
+ breadth: point.breadth,
728
+ contributingSpans: regionSpans.get(point.anchor) ?? [],
729
+ clusters: point.clusters,
730
+ commit: {
731
+ status,
732
+ dominant,
733
+ passesNaturalBreak,
734
+ passesConsensusFloor,
735
+ pastLeadingSaturation,
736
+ rejectionReasons,
737
+ },
738
+ });
739
+ };
740
+ for (let rank = 0; rank < ranked.length; rank++) {
741
+ const point = ranked[rank];
538
742
  const absorbed = placed.some((p) => overlaps(point, p));
539
- if (!absorbed) {
743
+ // Commit decisions are recorded LIVE, inside this loop, in the exact
744
+ // shape the gates below apply them — never reconstructed afterward from
745
+ // the final `roots` (spec §8's explicit requirement).
746
+ let status;
747
+ let dominant = false;
748
+ let passesNaturalBreak;
749
+ let passesConsensusFloor;
750
+ let pastLeadingSaturation;
751
+ const rejectionReasons = [];
752
+ if (absorbed) {
753
+ status = "overlap";
754
+ }
755
+ else {
540
756
  const pastLeading = !sat.hasLeading ||
541
757
  roots.length === 0 || point.start >= sat.leadingEnd;
758
+ pastLeadingSaturation = pastLeading;
542
759
  const vote = votesIdf.get(point.anchor) ?? 0;
543
- if ((roots.length === 0 || (vote >= rootCut && vote >= floor)) &&
544
- pastLeading) {
760
+ if (roots.length === 0) {
761
+ // The first non-overlapping root is DOMINANT and bypasses the two
762
+ // vote thresholds (it always grounds) — only the leading-saturation
763
+ // gate still applies to it.
764
+ dominant = true;
765
+ if (pastLeading) {
766
+ status = "root";
767
+ }
768
+ else {
769
+ status = "rejected";
770
+ rejectionReasons.push("leading-saturation");
771
+ }
772
+ }
773
+ else {
774
+ passesNaturalBreak = vote >= rootCut;
775
+ passesConsensusFloor = vote >= floor;
776
+ if (passesNaturalBreak && passesConsensusFloor && pastLeading) {
777
+ status = "root";
778
+ }
779
+ else {
780
+ status = "rejected";
781
+ if (!passesNaturalBreak)
782
+ rejectionReasons.push("below-natural-break");
783
+ if (!passesConsensusFloor) {
784
+ rejectionReasons.push("below-consensus-floor");
785
+ }
786
+ if (!pastLeading)
787
+ rejectionReasons.push("leading-saturation");
788
+ }
789
+ }
790
+ if (status === "root") {
545
791
  roots.push(point);
546
792
  }
547
793
  else {
794
+ recordAnchor(point, rank, status, dominant, passesNaturalBreak, passesConsensusFloor, pastLeadingSaturation, rejectionReasons);
548
795
  continue;
549
796
  }
550
797
  }
798
+ recordAnchor(point, rank, status, dominant, passesNaturalBreak, passesConsensusFloor, pastLeadingSaturation, rejectionReasons);
551
799
  placed.push(point);
552
800
  }
553
- traceAttention(ctx, regions, regionVoter, roots, steps);
801
+ traceAttention(ctx, regions, regionVoter, roots, steps, td, cfg ? { ...cfg, naturalBreak: rootCut, consensusFloor: floor } : undefined, ranked);
554
802
  return { roots, ranked };
555
803
  }
556
804
  export function detectSaturated(ctx, regions, saturated) {
@@ -653,63 +901,186 @@ export function naturalBreak(votes) {
653
901
  }
654
902
  return votes[breakAt - 1];
655
903
  }
656
- // ═══════════════════════════════════════════════════════════════════════════
657
- // Cross-region attention — DIRECT region-to-region interaction.
658
- //
659
- // voteRegions climbs each region INDEPENDENTLY; poolVotes then ADDS those
660
- // independent votes. Additive pooling is a soft conjunction, but it can only
661
- // ever surface a context at least one region already votes for. Two regions
662
- // whose individual climbs land on DIFFERENT contexts leave their JOINT context
663
- // the learnt whole that contains BOTH with no vote at all, and no amount
664
- // of pooling can recover it. ("red" climbs to `red square`, "circle" to
665
- // `circle`; nothing votes for `red circle`, the only fact holding both.)
666
- //
667
- // This is the attention counterpart of the bridge, and it ascends by the SAME
668
- // content-addressed junction walk (junction.ts): "which learnt whole contains
669
- // region A then region B?" is a bounded DAG ascent from the two forms'
670
- // canonical identities NOT a resonance guess on a synthesised gist. Folding
671
- // two region vectors cannot even reconstruct the stored joint form: Sema builds
672
- // a multi-word gist from BYTE-chunk folds, so isolated word vectors superpose
673
- // into a different direction and resonate to `red circle` and `red square`
674
- // indistinguishably. The ascent sidesteps this by matching BYTES, not vectors.
675
- //
676
- // A joint container is EXACT evidence (it literally holds both forms), so it
677
- // votes at full strength — the exact-first discipline voteRegions gives a
678
- // content-addressed chunk. Each junction search is a bounded walk with NO
679
- // ANN query, and searches are capped at k. Three further disciplines make
680
- // the composition ORDER-FREE, N-ARY, and CORPUS-INDEPENDENT:
681
- //
682
- // ORDER-FREE a junction is evidence the forms were LEARNT TOGETHER;
683
- // which one the query mentioned first is a fact about the query, not the
684
- // learnt whole. The walk tests both byte orders at no extra walk cost
685
- // (see junctionContainersFrom's `unordered`).
686
- // • N-ARY binding is not intrinsically pairwise. A pair's containers are
687
- // FILTERED by the remaining candidate forms: the container covering the
688
- // MOST of the query's composable forms wins, so three cross-cutting
689
- // attributes (each pair ambiguous) still resolve to their unique triple —
690
- // at the cost of one cached byte read + indexOf per (container, extra),
691
- // never an extra walk.
692
- // • CORPUS-INDEPENDENT — candidates are ANY voted region, not just
693
- // recognised sites. A word never learnt standalone has no site, but its
694
- // stored chunks still vote and their BYTES still compose: the ascent
695
- // matches byte containment, so a fragment pair evidences the same joint
696
- // container the whole word would. Contiguous shards of one word cannot
697
- // pair (the adjacency skip), and a pair covered by a single KNOWN region
698
- // is skipped — that whole form already votes directly, and re-deriving it
699
- // from its own pieces would only double-count.
700
- //
701
- // EXPLAINING AWAY (the aliasing complement of corpus independence): a chunk
702
- // of the query can straddle the byte grid so that it exists verbatim in the
703
- // WRONG deposit (" cir" of "red then circle" is a stored chunk of `blue
704
- // circle`, never of `red circle` — a pure alignment accident) and its
705
- // independent climb then votes for a context the query gives no reason to
706
- // believe. When a junction binds, any individual vote whose bytes the joint
707
- // container LITERALLY CONTAINS yet whose climb disagrees with the junction's
708
- // is superseded: the exact joint evidence explains those bytes, so their
709
- // disagreeing vote is grid aliasing, not signal. Votes whose bytes the
710
- // container does not hold (a genuine second topic) are untouched.
711
- // ═══════════════════════════════════════════════════════════════════════════
712
- async function crossRegionVotes(ctx, query, regions, rvs, k, N, reachMemo) {
904
+ const VARIANT_KIND_ORDER = {
905
+ "exact-exact": -1,
906
+ "left-synonym": 0,
907
+ "right-synonym": 1,
908
+ "double-synonym": 2,
909
+ };
910
+ /** A node's structural gist, read directly from its own stored bytes — the
911
+ * repository's node gist accessor: content is immutable and perception is
912
+ * a pure function of bytes, so re-perceiving a node's full stored bytes
913
+ * reproduces exactly the gist it was interned with. Never concatenates
914
+ * the sibling's bytes with anything else. */
915
+ function storedNodeGist(ctx, id) {
916
+ return gistOf(ctx, read(ctx, id));
917
+ }
918
+ /** A halo sibling's structural part, occupying the ORIGINAL query-region
919
+ * slot length the sibling replaces only the DIRECTION, never the query's
920
+ * own bytes, position, or gap (see §6 of the spec this implements). */
921
+ function siblingPart(ctx, sibling, originalRegion) {
922
+ return {
923
+ v: storedNodeGist(ctx, sibling.id),
924
+ len: originalRegion.end - originalRegion.start,
925
+ };
926
+ }
927
+ /** Build, bound and order every mandatory structural variant (§7-8): the
928
+ * exact/exact composition is always kept; up to `ctx.cfg.haloQueryK`
929
+ * synonym variants (single- and double-synonym combined, one shared
930
+ * budget) are appended, ordered by confidence, then kind, then sibling id. */
931
+ export function buildStructuralVariants(ctx, ra, rb, sides) {
932
+ const exactLeft = { v: ra.v, len: ra.end - ra.start };
933
+ const exactRight = { v: rb.v, len: rb.end - rb.start };
934
+ const synonymVariants = [];
935
+ for (const ls of sides.leftSiblings) {
936
+ synonymVariants.push({
937
+ left: siblingPart(ctx, ls, ra),
938
+ right: exactRight,
939
+ kind: "left-synonym",
940
+ semanticConfidence: ls.score,
941
+ leftSiblingId: ls.id,
942
+ });
943
+ }
944
+ for (const rs of sides.rightSiblings) {
945
+ synonymVariants.push({
946
+ left: exactLeft,
947
+ right: siblingPart(ctx, rs, rb),
948
+ kind: "right-synonym",
949
+ semanticConfidence: rs.score,
950
+ rightSiblingId: rs.id,
951
+ });
952
+ }
953
+ for (const ls of sides.leftSiblings) {
954
+ for (const rs of sides.rightSiblings) {
955
+ synonymVariants.push({
956
+ left: siblingPart(ctx, ls, ra),
957
+ right: siblingPart(ctx, rs, rb),
958
+ kind: "double-synonym",
959
+ semanticConfidence: Math.min(ls.score, rs.score),
960
+ leftSiblingId: ls.id,
961
+ rightSiblingId: rs.id,
962
+ });
963
+ }
964
+ }
965
+ // §8: semantic confidence desc, then kind (left-synonym, right-synonym,
966
+ // double-synonym), then left sibling id asc, then right sibling id asc.
967
+ synonymVariants.sort((a, b) => b.semanticConfidence - a.semanticConfidence ||
968
+ VARIANT_KIND_ORDER[a.kind] - VARIANT_KIND_ORDER[b.kind] ||
969
+ (a.leftSiblingId ?? -1) - (b.leftSiblingId ?? -1) ||
970
+ (a.rightSiblingId ?? -1) - (b.rightSiblingId ?? -1));
971
+ const variants = [
972
+ {
973
+ left: exactLeft,
974
+ right: exactRight,
975
+ kind: "exact-exact",
976
+ semanticConfidence: 1,
977
+ },
978
+ ...synonymVariants.slice(0, ctx.cfg.haloQueryK),
979
+ ];
980
+ return { variants, exactLeft, exactRight };
981
+ }
982
+ /** Deterministic best-of tie-break for two proposals ranked for the SAME
983
+ * candidate id — effectiveScore, then annScore, then semanticConfidence,
984
+ * then variant kind, then sibling ids (§10). */
985
+ function betterProposal(a, b) {
986
+ if (a.effectiveScore !== b.effectiveScore) {
987
+ return a.effectiveScore > b.effectiveScore;
988
+ }
989
+ if (a.annScore !== b.annScore)
990
+ return a.annScore > b.annScore;
991
+ if (a.semanticConfidence !== b.semanticConfidence) {
992
+ return a.semanticConfidence > b.semanticConfidence;
993
+ }
994
+ if (VARIANT_KIND_ORDER[a.variant] !== VARIANT_KIND_ORDER[b.variant]) {
995
+ return VARIANT_KIND_ORDER[a.variant] < VARIANT_KIND_ORDER[b.variant];
996
+ }
997
+ if ((a.leftSiblingId ?? -1) !== (b.leftSiblingId ?? -1)) {
998
+ return (a.leftSiblingId ?? -1) < (b.leftSiblingId ?? -1);
999
+ }
1000
+ return (a.rightSiblingId ?? -1) < (b.rightSiblingId ?? -1);
1001
+ }
1002
+ /** The final approximate tier: compose every retained structural variant,
1003
+ * ANN-query each, merge proposals by candidate id, and validate the winner
1004
+ * through the SAME structural gates every other tier answers to (saturation,
1005
+ * roots, IDF, contrastive margin). Returns null when nothing survives. */
1006
+ export async function structuralResonance(ctx, query, ra, rb, sides, k, N, reachMemo,
1007
+ /** Each side's OWN individual climb roots (from voteRegions), when it cast
1008
+ * one — the self-evidence backstop structural-resonance needs and the
1009
+ * exact tier gets for free from literal byte containment (§11's whole
1010
+ * premise: recover a JOINT context neither side votes for alone). A
1011
+ * candidate whose reach is exactly one side's own conclusion is not new
1012
+ * evidence of a joint whole; it is that side's resonance rediscovering
1013
+ * itself through a synthetic gist still dominated by its own direction. */
1014
+ ownRootsA, ownRootsB) {
1015
+ const { variants } = buildStructuralVariants(ctx, ra, rb, sides);
1016
+ const middleBytes = query.subarray(ra.end, rb.start);
1017
+ const middlePart = middleBytes.length === 0
1018
+ ? null
1019
+ : { v: perceive(ctx, middleBytes).v, len: middleBytes.length };
1020
+ const proposals = new Map();
1021
+ for (const variant of variants) {
1022
+ const parts = [variant.left];
1023
+ if (middlePart)
1024
+ parts.push(middlePart);
1025
+ parts.push(variant.right);
1026
+ const synthetic = composeStructuralGist(ctx.space, parts);
1027
+ const hits = await ctx.store.resonate(synthetic, k);
1028
+ for (const hit of hits) {
1029
+ const candidate = {
1030
+ id: hit.id,
1031
+ annScore: hit.score,
1032
+ semanticConfidence: variant.semanticConfidence,
1033
+ effectiveScore: hit.score * variant.semanticConfidence,
1034
+ variant: variant.kind,
1035
+ leftSiblingId: variant.leftSiblingId,
1036
+ rightSiblingId: variant.rightSiblingId,
1037
+ };
1038
+ const prev = proposals.get(hit.id);
1039
+ if (prev === undefined || betterProposal(candidate, prev)) {
1040
+ proposals.set(hit.id, candidate);
1041
+ }
1042
+ }
1043
+ }
1044
+ if (proposals.size === 0)
1045
+ return null;
1046
+ const sorted = [...proposals.values()].sort((a, b) => b.effectiveScore - a.effectiveScore || a.id - b.id);
1047
+ let selected = null;
1048
+ let selectedReach = null;
1049
+ let selectedIdf = 0;
1050
+ let rival = null;
1051
+ for (const p of sorted) {
1052
+ const reach = edgeAncestors(ctx, p.id, N, reachMemo);
1053
+ if (reach.saturated || reach.roots.length === 0)
1054
+ continue;
1055
+ const idf = Math.log(N / Math.max(1, reach.contextsReached));
1056
+ if (idf <= 0)
1057
+ continue;
1058
+ // Self-evidence backstop (see the param doc above): a candidate that is
1059
+ // exactly one side's own already-voted conclusion carries no JOINT
1060
+ // evidence — skip it as if it never survived.
1061
+ if ((ownRootsA && sameRoots(reach.roots, ownRootsA)) ||
1062
+ (ownRootsB && sameRoots(reach.roots, ownRootsB)))
1063
+ continue;
1064
+ if (selected === null) {
1065
+ selected = p;
1066
+ selectedReach = reach;
1067
+ selectedIdf = idf;
1068
+ }
1069
+ else if (!sameRoots(reach.roots, selectedReach.roots)) {
1070
+ rival = p;
1071
+ break;
1072
+ }
1073
+ }
1074
+ if (selected === null || selectedReach === null)
1075
+ return null;
1076
+ const margin = rival
1077
+ ? selected.effectiveScore - rival.effectiveScore
1078
+ : selected.effectiveScore;
1079
+ if (margin <= estimatorNoise(ctx.store.D))
1080
+ return null;
1081
+ return { proposal: selected, reach: selectedReach, idf: selectedIdf };
1082
+ }
1083
+ async function crossRegionVotes(ctx, query, regions, rvs, k, N, reachMemo, td) {
713
1084
  // Candidate regions: every region that ALREADY CAST ITS OWN VOTE in
714
1085
  // voteRegions — individually idf > 0, genuinely discriminative on its own,
715
1086
  // just not necessarily for the SAME context as its partner. This is the
@@ -759,6 +1130,14 @@ async function crossRegionVotes(ctx, query, regions, rvs, k, N, reachMemo) {
759
1130
  regions[x].end <= regions[y].end &&
760
1131
  regions[y].end - regions[y].start > regions[x].end - regions[x].start));
761
1132
  const none = { votes: [], superseded: new Set() };
1133
+ if (td) {
1134
+ td.crossRegionSummary = {
1135
+ eligibleRegions: eligible.length,
1136
+ maximalRegions: cand.length,
1137
+ probeLimit: k,
1138
+ probesAttempted: 0, // updated below as probes accrue
1139
+ };
1140
+ }
762
1141
  if (cand.length < 2)
763
1142
  return none;
764
1143
  cand.sort((x, y) => regions[x].start - regions[y].start || regions[x].end - regions[y].end);
@@ -823,92 +1202,174 @@ async function crossRegionVotes(ctx, query, regions, rvs, k, N, reachMemo) {
823
1202
  if (regions.some((r) => r.known && r.start <= ra.start && rb.end <= r.end))
824
1203
  continue;
825
1204
  probes++;
1205
+ if (td?.crossRegionSummary) {
1206
+ td.crossRegionSummary.probesAttempted = probes;
1207
+ }
826
1208
  const left = query.subarray(ra.start, ra.end);
827
1209
  const right = query.subarray(rb.start, rb.end);
828
1210
  // Phrase-scale contract, exactly as the bridge: the glue between the two
829
1211
  // forms may be up to W× the content it joins.
830
1212
  const maxInterior = (left.length + right.length) * ctx.space.maxGroup;
831
1213
  const cap = left.length + right.length + maxInterior;
832
- // Tier 1 exact containers (both forms as substrings, either order, by
833
- // DAG ascent). Exact evidence first; only falls through to synonyms
834
- // when the exact ascent finds nothing the SAME graded ladder the
835
- // bridge uses.
1214
+ // The graded ladder (spec §1): exact DAG junction, then single-synonym,
1215
+ // then double-synonym, then only when every DAG tier found nothing —
1216
+ // structural-resonance. `sides` (the two halo sibling lists) is loaded
1217
+ // ONCE and reused by junctionSynonyms AND structural-resonance, so no
1218
+ // ladder rung repeats a halo ANN query an earlier rung already paid for.
1219
+ const sides = await loadJunctionSynonymSides(ctx, left, right);
1220
+ let tier = "exact";
836
1221
  let containers = junctionContainersFrom(ctx, left, right, cap, seedsOf(cand[a]), seedsOf(cand[b]), undefined, true);
837
1222
  if (containers.length === 0) {
838
- // Tier 2.5 — synonym containers (halo sibling of one side + the other).
839
- containers = await junctionSynonyms(ctx, left, right, maxInterior, true);
1223
+ // Tiers 2-4 — synonym containers (junctionSynonyms itself runs
1224
+ // single-synonym first, falling to double-synonym only when
1225
+ // single-synonym found nothing — see junction.ts).
1226
+ const syn = await junctionSynonyms(ctx, left, right, maxInterior, true, sides);
1227
+ if (syn.length > 0) {
1228
+ containers = syn;
1229
+ tier = syn[0].tier;
1230
+ }
1231
+ }
1232
+ // Tier 5 — structural-resonance ANN, the FINAL approximate proposal
1233
+ // path. Only reached when every DAG tier found NOTHING, and only when
1234
+ // there is no already-corroborated region between the endpoints (a
1235
+ // between-region with its own vote is evidence the gap already means
1236
+ // something specific — an ANN guess must not override it).
1237
+ let structuralPick = null;
1238
+ if (containers.length === 0) {
1239
+ // Structural-resonance composes each side's OWN gist directly (no
1240
+ // byte-containment truth backs it, unlike the DAG tiers) — so, unlike
1241
+ // the DAG ladder (which tolerates one approximate side because byte
1242
+ // containment cannot lie), the ANN tier requires BOTH sides to be
1243
+ // KNOWN (content-addressed, exact identities): an approximate chunk
1244
+ // fragment's own resonance is noise at any tier, and composing noise
1245
+ // into a synthetic gist only manufactures a plausible-looking but
1246
+ // spurious ANN neighbour, not evidence of a genuine joint whole.
1247
+ // PHRASE-SCALE CONTRACT — the same one the DAG tiers hold their glue
1248
+ // to (see maxInterior above): a junction, exact or approximate, is a
1249
+ // whole the two forms nearly exhaust, not two arbitrary landmarks
1250
+ // anywhere in a long, multi-topic query. Without this, structural-
1251
+ // resonance would pair opposite ends of an unrelated scaffolding-
1252
+ // dominated query and manufacture a plausible-looking ANN neighbour
1253
+ // for a "gap" that never was a phrase.
1254
+ // BOTH sides must be independently DISCRIMINATIVE (individually
1255
+ // voted — `strong`, not merely a content-addressed `known` chunk):
1256
+ // a shared, non-discriminative scaffolding run (a repeated system
1257
+ // preamble) can be `known` without ever being distinctive evidence
1258
+ // of anything, and composing its own gist into a synthetic query
1259
+ // manufactures a plausible-looking but spurious ANN neighbour. The
1260
+ // DAG tiers can tolerate one merely-`known` side because byte
1261
+ // containment cannot lie; structural-resonance has no such
1262
+ // backstop, so both sides earn their place here the same way an
1263
+ // ordinary approximate region earns its individual vote.
1264
+ const gap = rb.start - ra.end;
1265
+ if (between.length === 0 &&
1266
+ strong.has(cand[a]) && strong.has(cand[b]) &&
1267
+ ra.known && rb.known &&
1268
+ gap <= maxInterior) {
1269
+ const ownRootsA = rvs.votes.find((v) => v.start === ra.start && v.end === ra.end)?.roots;
1270
+ const ownRootsB = rvs.votes.find((v) => v.start === rb.start && v.end === rb.end)?.roots;
1271
+ structuralPick = await structuralResonance(ctx, query, ra, rb, sides, k, N, reachMemo, ownRootsA, ownRootsB);
1272
+ }
1273
+ if (structuralPick === null)
1274
+ continue;
1275
+ tier = "structural-resonance";
840
1276
  }
841
- if (containers.length === 0)
842
- continue;
843
- // N-ARY selection: the container covering the MOST remaining candidate
844
- // forms wins (then tightest interior, then lowest id). Reads are
845
- // cache hits — every container's bytes were already read by the walk.
846
- //
847
- // SELF-EVIDENCE GUARD: a junction is BINDING evidence only when the
848
- // container joins forms the query mentions APART. When the container's
849
- // own joined occurrence (left..right including its interior) is a
850
- // literal substring of the query, the query already spells that phrase
851
- // out contiguously — perception already voted with it, and grid shards
852
- // of one phrase pairing "around" a gap chunk would merely rediscover
853
- // the phrase they are shards of, then explain away its rivals.
854
1277
  let best = null;
855
1278
  let bestExtras = [];
856
1279
  let bestCov = -1;
857
- for (const c of containers) {
858
- const bytes = cachedRead(ctx, cache, c.id, cap);
859
- const li = indexOf(bytes, left, 0);
860
- const ri = indexOf(bytes, right, 0);
861
- if (li >= 0 && ri >= 0) {
862
- const joined = bytes.subarray(Math.min(li, ri), Math.max(li + left.length, ri + right.length));
863
- if (indexOf(query, joined, 0) >= 0)
864
- continue; // query says it itself
865
- }
866
- // CONTRADICTION GUARD: a between-region already carrying its own
867
- // vote must actually recur in this container's bytes — otherwise
868
- // the container is a different learnt whole that happens to share
869
- // ra/rb, and letting it stand in for the gap would silently
870
- // override evidence the query itself already resolved there.
871
- if (between.some((bi) => indexOf(bytes, query.subarray(regions[bi].start, regions[bi].end), 0) < 0)) {
872
- continue;
873
- }
874
- let cov = left.length + right.length;
875
- const extras = [];
876
- for (const ei of cand) {
877
- if (ei === cand[a] || ei === cand[b] || consumed.has(ei))
878
- continue;
879
- const e = regions[ei];
880
- if (overlapsSpan(e, ra) || overlapsSpan(e, rb))
1280
+ let reach;
1281
+ let idf;
1282
+ let confidence;
1283
+ if (structuralPick !== null) {
1284
+ // A resonance proposal is NOT a Junction — there is no container to
1285
+ // read bytes from, so the self-evidence/contradiction/N-ary
1286
+ // machinery below (byte-verified against a real container) does not
1287
+ // apply; per spec §13, no N-ary extra-region coverage for resonance
1288
+ // proposals.
1289
+ best = { id: structuralPick.proposal.id, interior: new Uint8Array(0) };
1290
+ bestExtras = [];
1291
+ bestCov = rb.end - ra.start;
1292
+ reach = structuralPick.reach;
1293
+ idf = structuralPick.idf;
1294
+ confidence = structuralPick.proposal.effectiveScore;
1295
+ }
1296
+ else {
1297
+ // N-ARY selection: the container covering the MOST remaining candidate
1298
+ // forms wins (then tightest interior, then lowest id). Reads are
1299
+ // cache hits every container's bytes were already read by the walk.
1300
+ //
1301
+ // SELF-EVIDENCE GUARD: a junction is BINDING evidence only when the
1302
+ // container joins forms the query mentions APART. When the container's
1303
+ // own joined occurrence (left..right including its interior) is a
1304
+ // literal substring of the query, the query already spells that phrase
1305
+ // out contiguously — perception already voted with it, and grid shards
1306
+ // of one phrase pairing "around" a gap chunk would merely rediscover
1307
+ // the phrase they are shards of, then explain away its rivals.
1308
+ for (const c of containers) {
1309
+ const bytes = cachedRead(ctx, cache, c.id, cap);
1310
+ const li = indexOf(bytes, left, 0);
1311
+ const ri = indexOf(bytes, right, 0);
1312
+ if (li >= 0 && ri >= 0) {
1313
+ const joined = bytes.subarray(Math.min(li, ri), Math.max(li + left.length, ri + right.length));
1314
+ if (indexOf(query, joined, 0) >= 0)
1315
+ continue; // query says it itself
1316
+ }
1317
+ // CONTRADICTION GUARD: a between-region already carrying its own
1318
+ // vote must actually recur in this container's bytes — otherwise
1319
+ // the container is a different learnt whole that happens to share
1320
+ // ra/rb, and letting it stand in for the gap would silently
1321
+ // override evidence the query itself already resolved there.
1322
+ if (between.some((bi) => indexOf(bytes, query.subarray(regions[bi].start, regions[bi].end), 0) < 0)) {
881
1323
  continue;
882
- const eb = query.subarray(e.start, e.end);
883
- if (indexOf(bytes, eb, 0) >= 0) {
884
- extras.push(ei);
885
- cov += eb.length;
1324
+ }
1325
+ let cov = left.length + right.length;
1326
+ const extras = [];
1327
+ for (const ei of cand) {
1328
+ if (ei === cand[a] || ei === cand[b] || consumed.has(ei))
1329
+ continue;
1330
+ const e = regions[ei];
1331
+ if (overlapsSpan(e, ra) || overlapsSpan(e, rb))
1332
+ continue;
1333
+ const eb = query.subarray(e.start, e.end);
1334
+ if (indexOf(bytes, eb, 0) >= 0) {
1335
+ extras.push(ei);
1336
+ cov += eb.length;
1337
+ }
1338
+ }
1339
+ if (cov > bestCov ||
1340
+ (cov === bestCov && best !== null &&
1341
+ (c.interior.length < best.interior.length ||
1342
+ (c.interior.length === best.interior.length && c.id < best.id)))) {
1343
+ best = c;
1344
+ bestExtras = extras;
1345
+ bestCov = cov;
886
1346
  }
887
1347
  }
888
- if (cov > bestCov ||
889
- (cov === bestCov && best !== null &&
890
- (c.interior.length < best.interior.length ||
891
- (c.interior.length === best.interior.length && c.id < best.id)))) {
892
- best = c;
893
- bestExtras = extras;
894
- bestCov = cov;
895
- }
1348
+ if (best === null)
1349
+ continue; // every container was self-evidence
1350
+ const r = edgeAncestors(ctx, best.id, N, reachMemo);
1351
+ if (r.saturated || r.roots.length === 0)
1352
+ continue;
1353
+ const df = Math.log(N / Math.max(1, r.contextsReached));
1354
+ if (df <= 0)
1355
+ continue;
1356
+ reach = r;
1357
+ idf = df;
1358
+ // Confidence used by voting (spec §13): exact junction = 1;
1359
+ // single/double-synonym = the sibling(s)' score(s), carried on the
1360
+ // SynonymJunction the ladder selected.
1361
+ confidence = "confidence" in best ? best.confidence : 1;
896
1362
  }
897
- if (best === null)
898
- continue; // every container was self-evidence
899
- const reach = edgeAncestors(ctx, best.id, N, reachMemo);
900
- if (reach.saturated || reach.roots.length === 0)
901
- continue;
902
- const idf = Math.log(N / Math.max(1, reach.contextsReached));
903
- if (idf <= 0)
904
- continue;
905
- // EXACT joint evidence (score = 1): the container literally contains
906
- // every composed form. Mutual-explanation weight over their COMBINED
907
- // byte length — the same magnitude reading voteRegions uses, here with
908
- // the estimator collapsed to certainty, so mutual = min(ratio, 1/ratio).
1363
+ // MUTUAL-EXPLANATION WEIGHT — the same formula for every tier, with
1364
+ // `confidence` collapsed to certainty (1) for exact evidence: under
1365
+ // that collapse this is byte-for-byte the old exact-only formula
1366
+ // (min(1,ratio)·min(1,1/ratio)). For structural-resonance,
1367
+ // `confidence` is already annScore·semanticConfidence — never
1368
+ // multiplied a second time.
909
1369
  const lenR = Math.max(1, bestCov);
910
1370
  const ratio = Math.sqrt(Math.max(1, ctx.store.contentLen(best.id, lenR * ctx.store.D)) / lenR);
911
- const mutual = Math.min(1, ratio) * Math.min(1, 1 / ratio);
1371
+ const mutual = Math.min(1, confidence * ratio) *
1372
+ Math.min(1, confidence / ratio);
912
1373
  const w = (mutual * idf) / reach.roots.length;
913
1374
  let spanStart = ra.start;
914
1375
  let spanEnd = rb.end;
@@ -930,16 +1391,33 @@ async function crossRegionVotes(ctx, query, regions, rvs, k, N, reachMemo) {
930
1391
  // for, not evidence lost — `absorbed` (RegionVote's breadth-accounting
931
1392
  // field) must credit the junction with all of it, not just the ONE
932
1393
  // pooled axiom it collapses to.
933
- const containerBytes = cachedRead(ctx, cache, best.id, cap);
934
- const jointRoots = new Set(reach.roots);
1394
+ // Only EXACT DAG evidence may explain away ordinary votes (spec §15).
1395
+ // Single-synonym, double-synonym, and structural-resonance may ADD
1396
+ // supporting evidence but never remove it: their evidence is itself
1397
+ // approximate (a sibling substitution, or an ANN guess), so treating
1398
+ // their byte-containment the way exact containment is treated would
1399
+ // let an approximation override a genuine, independently-voted region.
935
1400
  let explainedAway = 0;
936
- for (const rv of rvs.votes) {
937
- if (rv.roots.some((r) => jointRoots.has(r)))
938
- continue;
939
- const bytes = query.subarray(rv.start, rv.end);
940
- if (indexOf(containerBytes, bytes, 0) >= 0 && !superseded.has(rv)) {
941
- superseded.add(rv);
942
- explainedAway++;
1401
+ // Exact set of ORIGINAL region indices this junction explained away —
1402
+ // recorded live as `superseded.add` fires (spec §3's explicit rule:
1403
+ // never inferred from `absorbed` afterward).
1404
+ const explainedAwayIndices = [];
1405
+ if (tier === "exact") {
1406
+ const containerBytes = cachedRead(ctx, cache, best.id, cap);
1407
+ const jointRoots = new Set(reach.roots);
1408
+ for (const rv of rvs.votes) {
1409
+ if (rv.roots.some((r) => jointRoots.has(r)))
1410
+ continue;
1411
+ const bytes = query.subarray(rv.start, rv.end);
1412
+ if (indexOf(containerBytes, bytes, 0) >= 0 && !superseded.has(rv)) {
1413
+ superseded.add(rv);
1414
+ explainedAway++;
1415
+ if (td) {
1416
+ const idx = regions.findIndex((r) => r.start === rv.start && r.end === rv.end);
1417
+ if (idx >= 0)
1418
+ explainedAwayIndices.push(idx);
1419
+ }
1420
+ }
943
1421
  }
944
1422
  }
945
1423
  out.push({
@@ -951,28 +1429,61 @@ async function crossRegionVotes(ctx, query, regions, rvs, k, N, reachMemo) {
951
1429
  wFocus: w,
952
1430
  absorbed: 1 + explainedAway,
953
1431
  });
1432
+ if (td) {
1433
+ td.crossRegionJunctionVotes.push({
1434
+ container: best.id,
1435
+ span: [spanStart, spanEnd],
1436
+ roots: [...reach.roots],
1437
+ sourceRegionIndices: [cand[a], cand[b], ...bestExtras],
1438
+ explainedAwayRegionIndices: explainedAwayIndices,
1439
+ absorbed: 1 + explainedAway,
1440
+ tier,
1441
+ });
1442
+ }
954
1443
  const label = [cand[a], cand[b], ...bestExtras]
955
1444
  .sort((x, y) => regions[x].start - regions[y].start)
956
1445
  .map((ri) => dec(query.subarray(regions[ri].start, regions[ri].end)))
957
1446
  .join(" ▸ ");
1447
+ const tierNote = tier === "exact"
1448
+ ? `junction node ${best.id}` +
1449
+ (best.interior.length === 0
1450
+ ? " (adjacent)"
1451
+ : ` (interior "${dec(best.interior)}")`) +
1452
+ ", by content-addressed ascent"
1453
+ : tier === "structural-resonance"
1454
+ ? `structurally-composed ANN proposal, node ${best.id} — the query ` +
1455
+ `structurally composed the endpoint regions, the real middle-` +
1456
+ `query structure, and the selected halo-sibling endpoint ` +
1457
+ `direction(s) (variant ${structuralPick.proposal.variant}, ` +
1458
+ `annScore ${structuralPick.proposal.annScore.toFixed(3)} × ` +
1459
+ `semanticConfidence ${structuralPick.proposal.semanticConfidence.toFixed(3)} = effectiveScore ${structuralPick.proposal.effectiveScore.toFixed(3)}); it did not concatenate endpoint bytes or rewrite the query`
1460
+ : `${tier} junction node ${best.id}` +
1461
+ (best.interior.length === 0
1462
+ ? " (adjacent)"
1463
+ : ` (interior "${dec(best.interior)}")`) +
1464
+ `, by halo-sibling DAG ascent (confidence ${confidence.toFixed(3)})`;
958
1465
  ctx.trace?.step("crossRegion", [{ text: label, role: "pair" }], reach.roots.map((r) => ({
959
1466
  text: dec(read(ctx, r)).slice(0, 60),
960
1467
  node: r,
961
1468
  role: "joint-context",
962
- })), `${label} → junction node ${best.id}` +
963
- (best.interior.length === 0
964
- ? " (adjacent)"
965
- : ` (interior "${dec(best.interior)}")`) +
966
- ` → ${reach.roots.length} context(s), by content-addressed ascent` +
1469
+ })), `${label} → ${tierNote} ${reach.roots.length} context(s)` +
967
1470
  (superseded.size > 0
968
1471
  ? `; ${superseded.size} aliasing vote(s) explained away`
969
1472
  : ""));
970
1473
  break; // ra is consumed — move to the next unconsumed candidate
971
1474
  }
972
1475
  }
1476
+ if (td)
1477
+ td.supersededOrdinaryVotes = superseded.size;
973
1478
  return { votes: out, superseded };
974
1479
  }
975
- export function traceAttention(ctx, regions, regionVoter, roots, steps = []) {
1480
+ /** Emit the "climbConsensus" step the human-readable note this always
1481
+ * produced, now paired (when `ctx.trace` and `cfg` are both present) with
1482
+ * the structured {@link ClimbConsensusData} payload on the SAME step's
1483
+ * `data` field. Every exit of {@link computeAttention} funnels through
1484
+ * here, so instrumentation and the existing rationale text can never drift
1485
+ * apart — see the instrumentation spec's §9 "every exit path". */
1486
+ export function traceAttention(ctx, regions, regionVoter, roots, steps = [], td, cfg, ranked = roots) {
976
1487
  if (!ctx.trace)
977
1488
  return;
978
1489
  const voters = [];
@@ -989,9 +1500,50 @@ export function traceAttention(ctx, regions, regionVoter, roots, steps = []) {
989
1500
  // shape {@link GraphSearch}'s own cover steps take (see traceDerivation).
990
1501
  if (steps.length > 0)
991
1502
  traceDerivation(ctx, steps);
1503
+ const data = (td && cfg)
1504
+ ? {
1505
+ version: 1,
1506
+ cache: { hit: false, detailAvailable: true },
1507
+ config: {
1508
+ annK: cfg.k,
1509
+ crossRegionProbeLimit: cfg.k,
1510
+ mode: cfg.mode,
1511
+ ...(cfg.N !== undefined ? { corpusN: cfg.N } : {}),
1512
+ dimension: ctx.store.D,
1513
+ ...(cfg.N !== undefined ? { hubBound: hubBound(ctx) } : {}),
1514
+ estimatorNoise: estimatorNoise(ctx.store.D),
1515
+ ...(cfg.naturalBreak !== undefined
1516
+ ? { naturalBreak: cfg.naturalBreak }
1517
+ : {}),
1518
+ ...(cfg.consensusFloor !== undefined
1519
+ ? { consensusFloor: cfg.consensusFloor }
1520
+ : {}),
1521
+ },
1522
+ candidates: {
1523
+ perceived: cfg.perceivedCount,
1524
+ recognised: cfg.totalRegions - cfg.perceivedCount,
1525
+ total: cfg.totalRegions,
1526
+ },
1527
+ ...(td.regions.length > 0 ? { regions: td.regions } : {}),
1528
+ ...(cfg.reachMemo ? { reaches: serialiseReaches(cfg.reachMemo) } : {}),
1529
+ ...(td.crossRegionSummary
1530
+ ? {
1531
+ crossRegion: {
1532
+ ...td.crossRegionSummary,
1533
+ junctionVotes: td.crossRegionJunctionVotes,
1534
+ supersededOrdinaryVotes: td.supersededOrdinaryVotes,
1535
+ },
1536
+ }
1537
+ : {}),
1538
+ ...(td.saturation ? { saturation: td.saturation } : {}),
1539
+ ...(td.pooling ? { pooling: td.pooling } : {}),
1540
+ ...(td.anchors.length > 0 ? { anchors: td.anchors } : {}),
1541
+ result: { roots: [...roots], ranked: [...ranked] },
1542
+ }
1543
+ : undefined;
992
1544
  t.done(roots.map((r) => rNode(ctx, r.anchor, "anchor", r.vote)), roots.length === 0
993
1545
  ? `${regions.length} sub-regions climbed the DAG, but none agreed on a context`
994
1546
  : roots.length === 1
995
1547
  ? `${voters.length} of ${regions.length} sub-regions voted; IDF-weighted consensus picked one context (vote ${roots[0].vote.toFixed(2)})`
996
- : `${voters.length} of ${regions.length} sub-regions voted; consensus ordered ${roots.length} INDEPENDENT points of attention (votes ${roots.map((r) => r.vote.toFixed(2)).join(", ")})`);
1548
+ : `${voters.length} of ${regions.length} sub-regions voted; consensus ordered ${roots.length} INDEPENDENT points of attention (votes ${roots.map((r) => r.vote.toFixed(2)).join(", ")})`, data);
997
1549
  }