@hviana/sema 0.5.2 → 0.5.4
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/AGENTS.md +115 -53
- package/HOW_IT_WORKS.md +291 -184
- package/jsr.json +6 -0
- package/package.json +1 -1
- package/src/mind/bridge.ts +10 -104
- package/src/mind/match.ts +497 -7
- package/src/mind/{prefix-completion.ts → mechanisms/prefix-completion.ts} +66 -92
- package/src/mind/mechanisms/recall.ts +38 -129
- package/src/mind/mechanisms/reference.ts +343 -0
- package/src/mind/mind.ts +12 -8
- package/src/mind/pipeline-mechanism.ts +120 -3
- package/src/mind/pipeline.ts +16 -2
- package/src/mind/traverse.ts +92 -1
- package/src/store.ts +13 -4
- package/test/33-multi-candidate.test.mjs +21 -11
- package/test/70-prefix-completion.test.mjs +1 -1
- package/test/72-prefix-candidate-supply.test.mjs +7 -9
- package/test/74-prefix-trap-not-sprung-early.test.mjs +1 -1
- package/test/76-reference-binding.test.mjs +543 -0
- package/dist/example/demo.d.ts +0 -1
- package/dist/example/demo.js +0 -39
- package/dist/example/train_base.d.ts +0 -93
- package/dist/example/train_base.js +0 -2269
- package/dist/src/alphabet.d.ts +0 -7
- package/dist/src/alphabet.js +0 -33
- package/dist/src/alu/src/alu.d.ts +0 -185
- package/dist/src/alu/src/alu.js +0 -440
- package/dist/src/alu/src/expr.d.ts +0 -61
- package/dist/src/alu/src/expr.js +0 -318
- package/dist/src/alu/src/index.d.ts +0 -11
- package/dist/src/alu/src/index.js +0 -19
- package/dist/src/alu/src/kernel-arith.d.ts +0 -16
- package/dist/src/alu/src/kernel-arith.js +0 -264
- package/dist/src/alu/src/kernel-bits.d.ts +0 -19
- package/dist/src/alu/src/kernel-bits.js +0 -152
- package/dist/src/alu/src/kernel-logic.d.ts +0 -4
- package/dist/src/alu/src/kernel-logic.js +0 -60
- package/dist/src/alu/src/kernel-nd.d.ts +0 -3
- package/dist/src/alu/src/kernel-nd.js +0 -208
- package/dist/src/alu/src/kernel-numeric.d.ts +0 -54
- package/dist/src/alu/src/kernel-numeric.js +0 -366
- package/dist/src/alu/src/operation.d.ts +0 -168
- package/dist/src/alu/src/operation.js +0 -189
- package/dist/src/alu/src/parser.d.ts +0 -221
- package/dist/src/alu/src/parser.js +0 -577
- package/dist/src/alu/src/resonance.d.ts +0 -55
- package/dist/src/alu/src/resonance.js +0 -126
- package/dist/src/alu/src/text.d.ts +0 -31
- package/dist/src/alu/src/text.js +0 -73
- package/dist/src/alu/src/value.d.ts +0 -109
- package/dist/src/alu/src/value.js +0 -300
- package/dist/src/alu/test/alu.test.d.ts +0 -1
- package/dist/src/alu/test/alu.test.js +0 -764
- package/dist/src/bytes.d.ts +0 -14
- package/dist/src/bytes.js +0 -59
- package/dist/src/canon.d.ts +0 -45
- package/dist/src/canon.js +0 -85
- package/dist/src/config.d.ts +0 -111
- package/dist/src/config.js +0 -91
- package/dist/src/derive/src/deduction.d.ts +0 -136
- package/dist/src/derive/src/deduction.js +0 -159
- package/dist/src/derive/src/index.d.ts +0 -8
- package/dist/src/derive/src/index.js +0 -11
- package/dist/src/derive/src/priority-queue.d.ts +0 -20
- package/dist/src/derive/src/priority-queue.js +0 -73
- package/dist/src/derive/src/rewrite.d.ts +0 -56
- package/dist/src/derive/src/rewrite.js +0 -100
- package/dist/src/derive/src/trie.d.ts +0 -90
- package/dist/src/derive/src/trie.js +0 -217
- package/dist/src/derive/test/derive.test.d.ts +0 -1
- package/dist/src/derive/test/derive.test.js +0 -122
- package/dist/src/extension.d.ts +0 -37
- package/dist/src/extension.js +0 -7
- package/dist/src/geometry.d.ts +0 -233
- package/dist/src/geometry.js +0 -1089
- package/dist/src/index.d.ts +0 -17
- package/dist/src/index.js +0 -19
- package/dist/src/ingest-cache.d.ts +0 -41
- package/dist/src/ingest-cache.js +0 -165
- package/dist/src/meter.d.ts +0 -176
- package/dist/src/meter.js +0 -274
- package/dist/src/mind/articulation.d.ts +0 -6
- package/dist/src/mind/articulation.js +0 -99
- package/dist/src/mind/attention.d.ts +0 -430
- package/dist/src/mind/attention.js +0 -2418
- package/dist/src/mind/bridge.d.ts +0 -39
- package/dist/src/mind/bridge.js +0 -998
- package/dist/src/mind/canonical.d.ts +0 -34
- package/dist/src/mind/canonical.js +0 -93
- package/dist/src/mind/frame-filler.d.ts +0 -15
- package/dist/src/mind/frame-filler.js +0 -535
- package/dist/src/mind/graph-search.d.ts +0 -294
- package/dist/src/mind/graph-search.js +0 -996
- package/dist/src/mind/index.d.ts +0 -9
- package/dist/src/mind/index.js +0 -5
- package/dist/src/mind/junction.d.ts +0 -137
- package/dist/src/mind/junction.js +0 -358
- package/dist/src/mind/learning.d.ts +0 -75
- package/dist/src/mind/learning.js +0 -265
- package/dist/src/mind/match.d.ts +0 -192
- package/dist/src/mind/match.js +0 -666
- package/dist/src/mind/mechanisms/alu.d.ts +0 -4
- package/dist/src/mind/mechanisms/alu.js +0 -36
- package/dist/src/mind/mechanisms/cast.d.ts +0 -89
- package/dist/src/mind/mechanisms/cast.js +0 -1186
- package/dist/src/mind/mechanisms/confluence.d.ts +0 -24
- package/dist/src/mind/mechanisms/confluence.js +0 -271
- package/dist/src/mind/mechanisms/cover.d.ts +0 -6
- package/dist/src/mind/mechanisms/cover.js +0 -232
- package/dist/src/mind/mechanisms/extraction.d.ts +0 -33
- package/dist/src/mind/mechanisms/extraction.js +0 -327
- package/dist/src/mind/mechanisms/recall.d.ts +0 -16
- package/dist/src/mind/mechanisms/recall.js +0 -558
- package/dist/src/mind/mind.d.ts +0 -386
- package/dist/src/mind/mind.js +0 -754
- package/dist/src/mind/pipeline-mechanism.d.ts +0 -183
- package/dist/src/mind/pipeline-mechanism.js +0 -521
- package/dist/src/mind/pipeline.d.ts +0 -49
- package/dist/src/mind/pipeline.js +0 -331
- package/dist/src/mind/prefix-completion.d.ts +0 -59
- package/dist/src/mind/prefix-completion.js +0 -270
- package/dist/src/mind/primitives.d.ts +0 -85
- package/dist/src/mind/primitives.js +0 -333
- package/dist/src/mind/rationale.d.ts +0 -139
- package/dist/src/mind/rationale.js +0 -163
- package/dist/src/mind/reasoning.d.ts +0 -42
- package/dist/src/mind/reasoning.js +0 -305
- package/dist/src/mind/recognition.d.ts +0 -20
- package/dist/src/mind/recognition.js +0 -696
- package/dist/src/mind/resonance.d.ts +0 -35
- package/dist/src/mind/resonance.js +0 -365
- package/dist/src/mind/trace.d.ts +0 -15
- package/dist/src/mind/trace.js +0 -73
- package/dist/src/mind/traverse.d.ts +0 -158
- package/dist/src/mind/traverse.js +0 -716
- package/dist/src/mind/types.d.ts +0 -418
- package/dist/src/mind/types.js +0 -179
- package/dist/src/rabitq-ivf/src/database.d.ts +0 -113
- package/dist/src/rabitq-ivf/src/database.js +0 -201
- package/dist/src/rabitq-ivf/src/index.d.ts +0 -7
- package/dist/src/rabitq-ivf/src/index.js +0 -4
- package/dist/src/rabitq-ivf/src/ivf.d.ts +0 -200
- package/dist/src/rabitq-ivf/src/ivf.js +0 -1165
- package/dist/src/rabitq-ivf/src/prng.d.ts +0 -19
- package/dist/src/rabitq-ivf/src/prng.js +0 -36
- package/dist/src/rabitq-ivf/src/rabitq.d.ts +0 -95
- package/dist/src/rabitq-ivf/src/rabitq.js +0 -313
- package/dist/src/sema.d.ts +0 -41
- package/dist/src/sema.js +0 -77
- package/dist/src/store-sqlite.d.ts +0 -184
- package/dist/src/store-sqlite.js +0 -942
- package/dist/src/store.d.ts +0 -762
- package/dist/src/store.js +0 -1823
- package/dist/src/vec.d.ts +0 -31
- package/dist/src/vec.js +0 -109
- package/src/mind/frame-filler.ts +0 -604
- package/test/69-frame-filler.test.mjs +0 -115
|
@@ -1,85 +0,0 @@
|
|
|
1
|
-
import { Vec } from "../vec.js";
|
|
2
|
-
import { Sema } from "../sema.js";
|
|
3
|
-
import type { Input, MindContext } from "./types.js";
|
|
4
|
-
/** The content key of a byte span — one latin1 char per byte, an exact,
|
|
5
|
-
* collision-free encoding. Spans on the perception path are query-scale
|
|
6
|
-
* (windows, regions, candidate spans), so key construction is far cheaper
|
|
7
|
-
* than the river fold it deduplicates. */
|
|
8
|
-
export declare function latin1Key(bytes: Uint8Array): string;
|
|
9
|
-
/** The {@link perceive} memo key: the span's content PLUS the boundary set it
|
|
10
|
-
* was folded under. The tree is a function of BOTH — the same bytes fold
|
|
11
|
-
* plainly with no boundaries and into a left-nested stable-prefix shape with
|
|
12
|
-
* them — so a content-only key returns whichever shape was computed first.
|
|
13
|
-
* That is exactly what happened: a conversation seeded its cumulative context
|
|
14
|
-
* under the content key, and every later plain `perceive` of those bytes was
|
|
15
|
-
* served the boundary tree instead (measured: respondTurn answered where
|
|
16
|
-
* respond() on byte-identical input did not). NUL separates the two parts —
|
|
17
|
-
* the boundary rendering is digits and commas, so no content byte can forge
|
|
18
|
-
* the split. */
|
|
19
|
-
export declare function perceiveKey(bytes: Uint8Array, boundaries?: readonly number[]): string;
|
|
20
|
-
/** Perceive input into a content-defined tree (the river fold).
|
|
21
|
-
* Deterministic — identical bytes always produce an identical tree.
|
|
22
|
-
*
|
|
23
|
-
* `boundaries` is an optional sorted list of proper byte offsets where the
|
|
24
|
-
* fold must split so that each prefix segment folds identically to how it
|
|
25
|
-
* folded when it was learned (§10.3 stable-prefix contract). Only the
|
|
26
|
-
* CALLER — who assembled the multi-turn context — knows where those
|
|
27
|
-
* boundaries are; the geometry never guesses them from the bytes. */
|
|
28
|
-
export declare function perceive(ctx: MindContext, input: Input, leafAt?: (i: number) => number | null, lookup?: (ids: number[]) => number | null, boundaries?: readonly number[]): Sema;
|
|
29
|
-
/** The DEPOSIT-shaped perceive. Folds over the stream's own content cuts —
|
|
30
|
-
* bit-identical to what inference computes for the same bytes. That
|
|
31
|
-
* train/inference agreement is the whole contract: the trained context node
|
|
32
|
-
* and the node `resolve(query)` reaches must be the SAME node, and the only
|
|
33
|
-
* way to guarantee it is to give this function nothing extra to say. It
|
|
34
|
-
* imposes no boundaries, knows nothing about turns, and reads no convention
|
|
35
|
-
* out of the bytes.
|
|
36
|
-
*
|
|
37
|
-
* An input that EXTENDS a previously deposited one — a conversation context
|
|
38
|
-
* grown by a turn, or a resumed replay — reuses that deposit's already-folded
|
|
39
|
-
* content segments ({@link contentFoldIncremental}), so it costs O(new bytes)
|
|
40
|
-
* instead of O(context). The reuse is TRANSPARENT by construction: a segment
|
|
41
|
-
* is a pure function of its own bytes, so a reused one is bit-identical to a
|
|
42
|
-
* refolded one. Nothing has to prove that the extending deposit is "really"
|
|
43
|
-
* a next turn — a coincidental byte prefix reuses the same segments and gets
|
|
44
|
-
* the same tree it would have got anyway. (It used to matter: while this
|
|
45
|
-
* path imposed turn BOUNDARIES, a wrong guess changed the tree, so the cache
|
|
46
|
-
* needed a continuation-bytes proof to gate it. Nothing is imposed now, so
|
|
47
|
-
* there is nothing to gate.) */
|
|
48
|
-
export declare function perceiveDeposit(ctx: MindContext, bytes: Uint8Array, conversational?: boolean): Sema;
|
|
49
|
-
/** The raw bytes of an input — modality-neutral conversion. */
|
|
50
|
-
export declare function inputBytes(ctx: MindContext, input: Input): Uint8Array;
|
|
51
|
-
/** Convenience: the gist vector of a byte span. */
|
|
52
|
-
export declare function gistOf(ctx: MindContext, bytes: Uint8Array): Vec;
|
|
53
|
-
/** Fold a perceived tree bottom-up against the store's content-addressed maps:
|
|
54
|
-
* every leaf is named by findLeaf, every branch by findBranch over its kids'
|
|
55
|
-
* ids (null the moment any child is unknown). `visit`, when given, sees each
|
|
56
|
-
* node with its byte span and resolved id. Returns the node's byte end and
|
|
57
|
-
* resolved id. */
|
|
58
|
-
export declare function foldTree(ctx: MindContext, n: Sema, start: number, visit?: (n: Sema, start: number, end: number, node: number | null) => void): {
|
|
59
|
-
end: number;
|
|
60
|
-
node: number | null;
|
|
61
|
-
};
|
|
62
|
-
/** The canonical node id of a byte span: perceive it in isolation — the way
|
|
63
|
-
* training did — and recover its root bottom-up. Returns null if any part is
|
|
64
|
-
* unknown. */
|
|
65
|
-
export declare function resolve(ctx: MindContext, bytes: Uint8Array): number | null;
|
|
66
|
-
/** Equivalence-class resolution: when the exact content-addressed lookup
|
|
67
|
-
* misses, find a stored node whose CANONICAL key equals the span's — the
|
|
68
|
-
* store's canon index proposes candidates by key hash, and each is verified
|
|
69
|
-
* by re-canonicalizing its bytes (hash-then-verify, like every content
|
|
70
|
-
* lookup). Among verified candidates, one that leads somewhere (has a
|
|
71
|
-
* continuation edge) is preferred; ties break to the lowest id — a corpus
|
|
72
|
-
* property, not a seed property. Null when the response carries no
|
|
73
|
-
* canonicalizer, the store has no canon index, or nothing verifies. */
|
|
74
|
-
export declare function canonResolve(ctx: MindContext, bytes: Uint8Array): number | null;
|
|
75
|
-
/** Walk a perceived tree in POST-ORDER with byte offsets — children before
|
|
76
|
-
* their parent, `visit(node, start, end)` for every node including leaves.
|
|
77
|
-
* Returns the byte end. The one shared traversal the offset-carrying tree
|
|
78
|
-
* readers (recognition via foldTree's richer variant, attention's region
|
|
79
|
-
* collection, resonance's branch counting) build on, so each does not
|
|
80
|
-
* re-derive the offset bookkeeping. (recognition.segment keeps its own
|
|
81
|
-
* walk: its flush semantics need PRE-order decisions at leaf-parents, which
|
|
82
|
-
* a post-order visitor cannot express.) */
|
|
83
|
-
export declare function walkTree(n: Sema, start: number, visit: (node: Sema, start: number, end: number) => void): number;
|
|
84
|
-
/** Reconstruct a node's byte content from the DAG, up to `maxLen` bytes. */
|
|
85
|
-
export declare function read(ctx: MindContext, id: number, maxLen?: number): Uint8Array;
|
|
@@ -1,333 +0,0 @@
|
|
|
1
|
-
// primitives.ts — Address + Read primitives (Section 1 of the mind).
|
|
2
|
-
//
|
|
3
|
-
// Address — bytes → node (perceive, foldTree, resolve)
|
|
4
|
-
// Read — node → bytes (read)
|
|
5
|
-
import { bytesToTree, contentFoldIncremental, gridToTree, hilbertBytes, stackGrids, } from "../geometry.js";
|
|
6
|
-
import { canonHash } from "../canon.js";
|
|
7
|
-
import { bytesEqual } from "../bytes.js";
|
|
8
|
-
import { ALL } from "./types.js";
|
|
9
|
-
// ── Address: bytes → node ──────────────────────────────────────────────
|
|
10
|
-
/** The content key of a byte span — one latin1 char per byte, an exact,
|
|
11
|
-
* collision-free encoding. Spans on the perception path are query-scale
|
|
12
|
-
* (windows, regions, candidate spans), so key construction is far cheaper
|
|
13
|
-
* than the river fold it deduplicates. */
|
|
14
|
-
export function latin1Key(bytes) {
|
|
15
|
-
// Batched String.fromCharCode — avoids the O(n²) cost of repeated += on
|
|
16
|
-
// potentially-large query spans, and stays well under the ~65536 arg limit.
|
|
17
|
-
const n = bytes.length;
|
|
18
|
-
let s = "";
|
|
19
|
-
for (let i = 0; i < n; i += 4096) {
|
|
20
|
-
s += String.fromCharCode(...bytes.subarray(i, Math.min(i + 4096, n)));
|
|
21
|
-
}
|
|
22
|
-
return s;
|
|
23
|
-
}
|
|
24
|
-
/** The {@link perceive} memo key: the span's content PLUS the boundary set it
|
|
25
|
-
* was folded under. The tree is a function of BOTH — the same bytes fold
|
|
26
|
-
* plainly with no boundaries and into a left-nested stable-prefix shape with
|
|
27
|
-
* them — so a content-only key returns whichever shape was computed first.
|
|
28
|
-
* That is exactly what happened: a conversation seeded its cumulative context
|
|
29
|
-
* under the content key, and every later plain `perceive` of those bytes was
|
|
30
|
-
* served the boundary tree instead (measured: respondTurn answered where
|
|
31
|
-
* respond() on byte-identical input did not). NUL separates the two parts —
|
|
32
|
-
* the boundary rendering is digits and commas, so no content byte can forge
|
|
33
|
-
* the split. */
|
|
34
|
-
export function perceiveKey(bytes, boundaries) {
|
|
35
|
-
const k = latin1Key(bytes);
|
|
36
|
-
return boundaries === undefined || boundaries.length === 0
|
|
37
|
-
? k
|
|
38
|
-
: k + "\u0000" + boundaries.join(",");
|
|
39
|
-
}
|
|
40
|
-
/** Perceive input into a content-defined tree (the river fold).
|
|
41
|
-
* Deterministic — identical bytes always produce an identical tree.
|
|
42
|
-
*
|
|
43
|
-
* `boundaries` is an optional sorted list of proper byte offsets where the
|
|
44
|
-
* fold must split so that each prefix segment folds identically to how it
|
|
45
|
-
* folded when it was learned (§10.3 stable-prefix contract). Only the
|
|
46
|
-
* CALLER — who assembled the multi-turn context — knows where those
|
|
47
|
-
* boundaries are; the geometry never guesses them from the bytes. */
|
|
48
|
-
export function perceive(ctx, input, leafAt, lookup, boundaries) {
|
|
49
|
-
if (typeof input === "string" || input instanceof Uint8Array) {
|
|
50
|
-
const bytes = typeof input === "string"
|
|
51
|
-
? new TextEncoder().encode(input)
|
|
52
|
-
: input;
|
|
53
|
-
if (leafAt === undefined && lookup === undefined) {
|
|
54
|
-
// Per-response memo (see MindContext.perceiveMemo): only the plain
|
|
55
|
-
// inference shape — raw bytes, no store capabilities — is memoised,
|
|
56
|
-
// keyed by CONTENT so byte-identical spans in fresh arrays still hit.
|
|
57
|
-
// The tree is shared by reference; Sema nodes are never mutated.
|
|
58
|
-
const memo = ctx.perceiveMemo;
|
|
59
|
-
if (memo) {
|
|
60
|
-
const key = perceiveKey(bytes, boundaries);
|
|
61
|
-
const hit = memo.get(key);
|
|
62
|
-
if (hit !== undefined) {
|
|
63
|
-
if (ctx.meter)
|
|
64
|
-
ctx.meter.perceiveHits++;
|
|
65
|
-
return hit;
|
|
66
|
-
}
|
|
67
|
-
if (ctx.meter) {
|
|
68
|
-
ctx.meter.perceptions++;
|
|
69
|
-
ctx.meter.perceivedBytes += bytes.length;
|
|
70
|
-
}
|
|
71
|
-
const tree = bytesToTree(ctx.space, ctx.alphabet, bytes, undefined, undefined, boundaries);
|
|
72
|
-
memo.set(key, tree);
|
|
73
|
-
return tree;
|
|
74
|
-
}
|
|
75
|
-
if (ctx.meter) {
|
|
76
|
-
ctx.meter.perceptions++;
|
|
77
|
-
ctx.meter.perceivedBytes += bytes.length;
|
|
78
|
-
}
|
|
79
|
-
return bytesToTree(ctx.space, ctx.alphabet, bytes, undefined, undefined, boundaries);
|
|
80
|
-
}
|
|
81
|
-
return bytesToTree(ctx.space, ctx.alphabet, bytes, leafAt, lookup);
|
|
82
|
-
}
|
|
83
|
-
if (Array.isArray(input)) {
|
|
84
|
-
return gridToTree(ctx.space, ctx.alphabet, stackGrids(input));
|
|
85
|
-
}
|
|
86
|
-
return gridToTree(ctx.space, ctx.alphabet, input);
|
|
87
|
-
}
|
|
88
|
-
/** The DEPOSIT-shaped perceive. Folds over the stream's own content cuts —
|
|
89
|
-
* bit-identical to what inference computes for the same bytes. That
|
|
90
|
-
* train/inference agreement is the whole contract: the trained context node
|
|
91
|
-
* and the node `resolve(query)` reaches must be the SAME node, and the only
|
|
92
|
-
* way to guarantee it is to give this function nothing extra to say. It
|
|
93
|
-
* imposes no boundaries, knows nothing about turns, and reads no convention
|
|
94
|
-
* out of the bytes.
|
|
95
|
-
*
|
|
96
|
-
* An input that EXTENDS a previously deposited one — a conversation context
|
|
97
|
-
* grown by a turn, or a resumed replay — reuses that deposit's already-folded
|
|
98
|
-
* content segments ({@link contentFoldIncremental}), so it costs O(new bytes)
|
|
99
|
-
* instead of O(context). The reuse is TRANSPARENT by construction: a segment
|
|
100
|
-
* is a pure function of its own bytes, so a reused one is bit-identical to a
|
|
101
|
-
* refolded one. Nothing has to prove that the extending deposit is "really"
|
|
102
|
-
* a next turn — a coincidental byte prefix reuses the same segments and gets
|
|
103
|
-
* the same tree it would have got anyway. (It used to matter: while this
|
|
104
|
-
* path imposed turn BOUNDARIES, a wrong guess changed the tree, so the cache
|
|
105
|
-
* needed a continuation-bytes proof to gate it. Nothing is imposed now, so
|
|
106
|
-
* there is nothing to gate.) */
|
|
107
|
-
export function perceiveDeposit(ctx, bytes, conversational = false) {
|
|
108
|
-
// Longest cached PROPER prefix first — the most segments to reuse.
|
|
109
|
-
let prev;
|
|
110
|
-
const lens = [...ctx._depositLens]
|
|
111
|
-
.filter((L) => L >= 2 && L < bytes.length)
|
|
112
|
-
.sort((a, b) => b - a);
|
|
113
|
-
for (const L of lens) {
|
|
114
|
-
const hit = ctx._depositTrees.get(latin1Key(bytes.subarray(0, L)));
|
|
115
|
-
if (hit !== undefined) {
|
|
116
|
-
prev = hit.content;
|
|
117
|
-
break;
|
|
118
|
-
}
|
|
119
|
-
}
|
|
120
|
-
const folded = contentFoldIncremental(ctx.space, ctx.alphabet, bytes, prev);
|
|
121
|
-
// Only a CONVERSATIONAL deposit writes the cache: reuse is sound for any
|
|
122
|
-
// deposit, but the budget is 8 entries and a corpus of unrelated facts would
|
|
123
|
-
// evict the live chains for nothing. Purely a cost decision now, not a
|
|
124
|
-
// correctness one.
|
|
125
|
-
if (conversational && bytes.length >= 2) {
|
|
126
|
-
// The lengths set drifts as the map evicts; past the probe budget the
|
|
127
|
-
// drift itself becomes the cost (each stale length is an O(len) key
|
|
128
|
-
// build), so both reset together — losing only warm-up on live chains.
|
|
129
|
-
if (ctx._depositLens.size > 64) {
|
|
130
|
-
ctx._depositLens.clear();
|
|
131
|
-
ctx._depositTrees.clear();
|
|
132
|
-
}
|
|
133
|
-
ctx._depositTrees.set(latin1Key(bytes), { content: folded.fold });
|
|
134
|
-
ctx._depositLens.add(bytes.length);
|
|
135
|
-
}
|
|
136
|
-
return folded.tree;
|
|
137
|
-
}
|
|
138
|
-
/** The raw bytes of an input — modality-neutral conversion. */
|
|
139
|
-
export function inputBytes(ctx, input) {
|
|
140
|
-
if (typeof input === "string")
|
|
141
|
-
return new TextEncoder().encode(input);
|
|
142
|
-
if (input instanceof Uint8Array)
|
|
143
|
-
return input;
|
|
144
|
-
if (Array.isArray(input))
|
|
145
|
-
return hilbertBytes(stackGrids(input));
|
|
146
|
-
return hilbertBytes(input);
|
|
147
|
-
}
|
|
148
|
-
/** Convenience: the gist vector of a byte span. */
|
|
149
|
-
export function gistOf(ctx, bytes) {
|
|
150
|
-
return perceive(ctx, bytes).v;
|
|
151
|
-
}
|
|
152
|
-
/** Fold a perceived tree bottom-up against the store's content-addressed maps:
|
|
153
|
-
* every leaf is named by findLeaf, every branch by findBranch over its kids'
|
|
154
|
-
* ids (null the moment any child is unknown). `visit`, when given, sees each
|
|
155
|
-
* node with its byte span and resolved id. Returns the node's byte end and
|
|
156
|
-
* resolved id. */
|
|
157
|
-
export function foldTree(ctx, n, start, visit) {
|
|
158
|
-
// Subtree already resolved (from a previous conversation turn or an earlier
|
|
159
|
-
// recognition pass). The pyramid reuses prefix subtrees as identical Sema
|
|
160
|
-
// objects, so a conversation's prefix is warm from its second turn on.
|
|
161
|
-
// Without a visitor that makes foldTree O(suffix) instead of O(context);
|
|
162
|
-
// with one it stays O(context) and saves the per-node store probes instead
|
|
163
|
-
// (see below for why the distinction is not negotiable).
|
|
164
|
-
//
|
|
165
|
-
// WHAT THE CACHE KNOWS, AND WHAT IT DOES NOT. An entry records this
|
|
166
|
-
// subtree's id and byte length — nothing about its DESCENDANTS' spans.
|
|
167
|
-
// Returning here therefore emits ONE visit() where a cold walk emits one per
|
|
168
|
-
// node, and `visit` is not instrumentation: recognise() emits its sites from
|
|
169
|
-
// it (recognition.ts) and attention's collectRegions votes over what it
|
|
170
|
-
// yields (attention.ts). Skipping the descent silently shrinks the evidence
|
|
171
|
-
// those mechanisms see, purely because the cache happened to be warm.
|
|
172
|
-
//
|
|
173
|
-
// That is not hypothetical and not an edge case — it is every conversation
|
|
174
|
-
// turn after the first. `contentFoldIncremental` deliberately shares prefix
|
|
175
|
-
// segment OBJECTS across turns (~99% reuse), so by turn 2 the prefix is
|
|
176
|
-
// warm; meanwhile recogniseMemo/climbMemo are keyed on exact query BYTES,
|
|
177
|
-
// which a growing context never repeats. Warm subtrees + missed memos is
|
|
178
|
-
// the unprotected quadrant. Measured over real trained conversations,
|
|
179
|
-
// recognising the same context with a warm prefix lost 67-92% of its leaves
|
|
180
|
-
// (772->204, 589->47, 872->291, 377->37) with `sites` unchanged, so the loss
|
|
181
|
-
// is invisible to the coarse counts; a direct foldTree probe on identical
|
|
182
|
-
// bytes and an identical tree object fired visit() 661 times cold and 37
|
|
183
|
-
// warm. respond() is immune only because it never sets _resolvedSubtrees
|
|
184
|
-
// (mind.ts) — the degradation was unique to the multi-turn API.
|
|
185
|
-
//
|
|
186
|
-
// So the fast path is taken only when NOBODY IS WATCHING. With a visitor
|
|
187
|
-
// present we still walk, and the cache degrades to the thing it soundly is:
|
|
188
|
-
// an elision of the store probes (findLeaf/findBranch) at each node, not an
|
|
189
|
-
// elision of the traversal. Ids still come from the cache, so a warm walk
|
|
190
|
-
// is cheaper than a cold one; it is no longer *different* from one.
|
|
191
|
-
const cached = ctx._resolvedSubtrees?.get(n);
|
|
192
|
-
if (cached !== undefined && visit === undefined) {
|
|
193
|
-
return { end: start + cached.len, node: cached.id };
|
|
194
|
-
}
|
|
195
|
-
if (n.kids === null) {
|
|
196
|
-
const b = n.leaf ?? new Uint8Array(0);
|
|
197
|
-
const end = start + b.length;
|
|
198
|
-
const node = cached !== undefined ? cached.id : ctx.store.findLeaf(b);
|
|
199
|
-
visit?.(n, start, end, node);
|
|
200
|
-
if (node !== null && ctx._resolvedSubtrees) {
|
|
201
|
-
ctx._resolvedSubtrees.set(n, { id: node, len: b.length });
|
|
202
|
-
}
|
|
203
|
-
return { end, node };
|
|
204
|
-
}
|
|
205
|
-
let pos = start;
|
|
206
|
-
let known = true;
|
|
207
|
-
const kids = [];
|
|
208
|
-
for (const k of n.kids) {
|
|
209
|
-
const r = foldTree(ctx, k, pos, visit);
|
|
210
|
-
if (r.node === null)
|
|
211
|
-
known = false;
|
|
212
|
-
else if (known)
|
|
213
|
-
kids.push(r.node);
|
|
214
|
-
pos = r.end;
|
|
215
|
-
}
|
|
216
|
-
// Same store-probe elision as the leaf case: a cached entry already names
|
|
217
|
-
// this subtree, so the descent above was for `visit`'s benefit alone and the
|
|
218
|
-
// id need not be re-derived. Using it also keeps a warm walk's ids
|
|
219
|
-
// bit-identical to a cold walk's rather than re-deriving them from children
|
|
220
|
-
// that may themselves have come from cache.
|
|
221
|
-
const node = cached !== undefined
|
|
222
|
-
? cached.id
|
|
223
|
-
: known
|
|
224
|
-
? ctx.store.findBranch(kids)
|
|
225
|
-
: null;
|
|
226
|
-
visit?.(n, start, pos, node);
|
|
227
|
-
if (node !== null && ctx._resolvedSubtrees) {
|
|
228
|
-
ctx._resolvedSubtrees.set(n, { id: node, len: pos - start });
|
|
229
|
-
}
|
|
230
|
-
return { end: pos, node };
|
|
231
|
-
}
|
|
232
|
-
/** The canonical node id of a byte span: perceive it in isolation — the way
|
|
233
|
-
* training did — and recover its root bottom-up. Returns null if any part is
|
|
234
|
-
* unknown. */
|
|
235
|
-
export function resolve(ctx, bytes) {
|
|
236
|
-
if (bytes.length === 0)
|
|
237
|
-
return null;
|
|
238
|
-
if (ctx.meter)
|
|
239
|
-
ctx.meter.resolves++;
|
|
240
|
-
const exact = foldTree(ctx, perceive(ctx, bytes), 0).node;
|
|
241
|
-
if (exact !== null)
|
|
242
|
-
return exact;
|
|
243
|
-
return canonResolve(ctx, bytes);
|
|
244
|
-
}
|
|
245
|
-
/** Equivalence-class resolution: when the exact content-addressed lookup
|
|
246
|
-
* misses, find a stored node whose CANONICAL key equals the span's — the
|
|
247
|
-
* store's canon index proposes candidates by key hash, and each is verified
|
|
248
|
-
* by re-canonicalizing its bytes (hash-then-verify, like every content
|
|
249
|
-
* lookup). Among verified candidates, one that leads somewhere (has a
|
|
250
|
-
* continuation edge) is preferred; ties break to the lowest id — a corpus
|
|
251
|
-
* property, not a seed property. Null when the response carries no
|
|
252
|
-
* canonicalizer, the store has no canon index, or nothing verifies. */
|
|
253
|
-
export function canonResolve(ctx, bytes) {
|
|
254
|
-
const canon = ctx.canon;
|
|
255
|
-
const store = ctx.store;
|
|
256
|
-
if (canon === null || !store.canonFind)
|
|
257
|
-
return null;
|
|
258
|
-
if (bytes.length < 2)
|
|
259
|
-
return null;
|
|
260
|
-
const memo = ctx.canonMemo;
|
|
261
|
-
const memoKey = memo ? latin1Key(bytes) : "";
|
|
262
|
-
if (memo) {
|
|
263
|
-
const hit = memo.get(memoKey);
|
|
264
|
-
if (hit !== undefined)
|
|
265
|
-
return hit;
|
|
266
|
-
}
|
|
267
|
-
const set = (v) => {
|
|
268
|
-
memo?.set(memoKey, v);
|
|
269
|
-
return v;
|
|
270
|
-
};
|
|
271
|
-
const key = canon(bytes);
|
|
272
|
-
if (key.length === 0)
|
|
273
|
-
return set(null);
|
|
274
|
-
// A stored form that IS canonical is not in the index (buildCanonIndex
|
|
275
|
-
// skips identity rows) — the exact content-addressed lookup of the
|
|
276
|
-
// canonical bytes finds it directly.
|
|
277
|
-
if (key.length !== bytes.length || !bytesEqual(key, bytes)) {
|
|
278
|
-
const direct = foldTree(ctx, perceive(ctx, key), 0).node;
|
|
279
|
-
if (direct !== null)
|
|
280
|
-
return set(direct);
|
|
281
|
-
}
|
|
282
|
-
if (ctx.meter)
|
|
283
|
-
ctx.meter.canonLookups++;
|
|
284
|
-
const candidates = store.canonFind(canonHash(key));
|
|
285
|
-
if (candidates.length === 0)
|
|
286
|
-
return set(null);
|
|
287
|
-
let best = null;
|
|
288
|
-
let bestLeads = false;
|
|
289
|
-
for (const id of candidates) {
|
|
290
|
-
const bytesOf = read(ctx, id);
|
|
291
|
-
const stored = canon(bytesOf);
|
|
292
|
-
if (stored.length !== key.length || !bytesEqual(stored, key))
|
|
293
|
-
continue;
|
|
294
|
-
// The index stores FLAT content twins; the id the exact path would have
|
|
295
|
-
// resolved for these bytes is their FOLD — the deposit-shaped node that
|
|
296
|
-
// carries the edges and halos. Re-folding the candidate's bytes lands
|
|
297
|
-
// on exactly the node the canonical-case query would have found.
|
|
298
|
-
const folded = foldTree(ctx, perceive(ctx, bytesOf), 0).node;
|
|
299
|
-
const use = folded ?? id;
|
|
300
|
-
const leads = store.hasNext(use) || store.haloMass(use) > 0;
|
|
301
|
-
if (best === null || (leads && !bestLeads) ||
|
|
302
|
-
(leads === bestLeads && use < best)) {
|
|
303
|
-
best = use;
|
|
304
|
-
bestLeads = leads;
|
|
305
|
-
}
|
|
306
|
-
}
|
|
307
|
-
return set(best);
|
|
308
|
-
}
|
|
309
|
-
/** Walk a perceived tree in POST-ORDER with byte offsets — children before
|
|
310
|
-
* their parent, `visit(node, start, end)` for every node including leaves.
|
|
311
|
-
* Returns the byte end. The one shared traversal the offset-carrying tree
|
|
312
|
-
* readers (recognition via foldTree's richer variant, attention's region
|
|
313
|
-
* collection, resonance's branch counting) build on, so each does not
|
|
314
|
-
* re-derive the offset bookkeeping. (recognition.segment keeps its own
|
|
315
|
-
* walk: its flush semantics need PRE-order decisions at leaf-parents, which
|
|
316
|
-
* a post-order visitor cannot express.) */
|
|
317
|
-
export function walkTree(n, start, visit) {
|
|
318
|
-
if (n.kids === null) {
|
|
319
|
-
const end = start + (n.leaf?.length ?? 0);
|
|
320
|
-
visit(n, start, end);
|
|
321
|
-
return end;
|
|
322
|
-
}
|
|
323
|
-
let pos = start;
|
|
324
|
-
for (const k of n.kids)
|
|
325
|
-
pos = walkTree(k, pos, visit);
|
|
326
|
-
visit(n, start, pos);
|
|
327
|
-
return pos;
|
|
328
|
-
}
|
|
329
|
-
// ── Read: node → bytes ──────────────────────────────────────────────────
|
|
330
|
-
/** Reconstruct a node's byte content from the DAG, up to `maxLen` bytes. */
|
|
331
|
-
export function read(ctx, id, maxLen = ALL) {
|
|
332
|
-
return ctx.store.bytesPrefix(id, maxLen);
|
|
333
|
-
}
|
|
@@ -1,139 +0,0 @@
|
|
|
1
|
-
import type { Vec } from "../vec.js";
|
|
2
|
-
/** One element of a step's input or output vector.
|
|
3
|
-
*
|
|
4
|
-
* Modality-neutral and deliberately partial: an element might be a byte span of
|
|
5
|
-
* the query, a resolved graph node, a resonance hit with its score, a spliced
|
|
6
|
-
* connector — so every descriptive field is optional and a mechanism fills only
|
|
7
|
-
* the ones that carry meaning for what it did. `text` is always present (the
|
|
8
|
-
* human-readable rendering); the rest is provenance a debugger can lean on. */
|
|
9
|
-
export interface RationaleItem {
|
|
10
|
-
/** Human-readable rendering — decoded text for a byte span, else a label like
|
|
11
|
-
* "‹none›" or an operator name. Always set, so a step always reads. */
|
|
12
|
-
text: string;
|
|
13
|
-
/** The graph node this element is, or resolved to, when known — the handle to
|
|
14
|
-
* point back at the exact stored fact in the content-addressed DAG. */
|
|
15
|
-
node?: number;
|
|
16
|
-
/** The `[start, end)` span this element occupies in its step's frame of
|
|
17
|
-
* reference (usually the query or the answer being composed). */
|
|
18
|
-
span?: [number, number];
|
|
19
|
-
/** The resonance / cosine score that selected this element, when it was chosen
|
|
20
|
-
* by similarity rather than by exact structure. */
|
|
21
|
-
score?: number;
|
|
22
|
-
/** A short role tag — "query", "leaf", "form", "hit", "connector", "answer",
|
|
23
|
-
* … — naming what KIND of element this is within the step. */
|
|
24
|
-
role?: string;
|
|
25
|
-
/** The gist vector, only when the element fundamentally IS a vector and a
|
|
26
|
-
* caller asked to carry it (off by default — a D-float array per item would
|
|
27
|
-
* bury the reasoning it is meant to explain). */
|
|
28
|
-
v?: Vec;
|
|
29
|
-
}
|
|
30
|
-
/** A single completed act of inference — one mechanism, run once.
|
|
31
|
-
*
|
|
32
|
-
* Steps are emitted in COMPLETION order (a sub-mechanism finishes, and is
|
|
33
|
-
* reported, before the mechanism that called it), while `index` is assigned in
|
|
34
|
-
* ENTRY order (a parent reserves its index before its children run). So a
|
|
35
|
-
* parent's `index` is always lower than its children's, and the two orderings
|
|
36
|
-
* together give a valid topological reading of the dependency graph. */
|
|
37
|
-
export interface RationaleStep {
|
|
38
|
-
/** This step's index, assigned when the mechanism was ENTERED — a strict,
|
|
39
|
-
* incremental ordering over the whole inference. */
|
|
40
|
-
index: number;
|
|
41
|
-
/** The mechanism and its enclosing mechanisms, outermost → innermost, e.g.
|
|
42
|
-
* `["respond", "think", "recognise"]`. The last entry is this step; the
|
|
43
|
-
* prefix is the nest of sub-mechanisms it ran inside. */
|
|
44
|
-
mechanism: string[];
|
|
45
|
-
/** The enclosing mechanism's step index, or -1 at the root — the NESTING edge
|
|
46
|
-
* of the dependency graph (which step this one is a part of). */
|
|
47
|
-
parent: number;
|
|
48
|
-
/** The earlier steps whose OUTPUTS became this step's inputs — the DATA-FLOW
|
|
49
|
-
* edges of the dependency graph. Defaults to the previous sibling (the step
|
|
50
|
-
* run just before this one inside the same mechanism), or the parent when
|
|
51
|
-
* this is the first sub-step; a mechanism that fuses several earlier results
|
|
52
|
-
* names them all explicitly. */
|
|
53
|
-
dependsOn: number[];
|
|
54
|
-
/** The vector of elements handed to the mechanism (one or more). */
|
|
55
|
-
inputs: RationaleItem[];
|
|
56
|
-
/** The vector of elements the mechanism produced (one or more) — longer than
|
|
57
|
-
* `inputs` when it decomposed, shorter when it combined. */
|
|
58
|
-
outputs: RationaleItem[];
|
|
59
|
-
/** A one-line, human account of what the mechanism did and why — the sentence
|
|
60
|
-
* that turns the data into an explanation. */
|
|
61
|
-
note?: string;
|
|
62
|
-
/** Optional structured payload — a mechanism-specific, plain-serialisable
|
|
63
|
-
* shape (no Map/Set/vectors/mutable internals) that carries more than the
|
|
64
|
-
* human-readable `note` can, for a debugger or downstream tool to consume
|
|
65
|
-
* programmatically. Never read by inference; purely additive. */
|
|
66
|
-
data?: unknown;
|
|
67
|
-
}
|
|
68
|
-
/** The callback {@link Mind.respond} / {@link Mind.respondText} accept. It is
|
|
69
|
-
* invoked once per completed step, AS the inference unfolds — never batched at
|
|
70
|
-
* the end — so a caller can stream the reasoning live or accumulate it. */
|
|
71
|
-
export type InspectRationale = (step: RationaleStep) => void;
|
|
72
|
-
/** Decode bytes to text for display, dropping the NUL padding the encoder uses
|
|
73
|
-
* (the same cleanup {@link Mind.respondText} does for its result). */
|
|
74
|
-
export declare function decodeText(bytes: Uint8Array): string;
|
|
75
|
-
/** The `[start, end)` gaps of `[0, queryLen)` NOT covered by `accounted` —
|
|
76
|
-
* the same union-of-spans reading think's grounding decider prices at PASS
|
|
77
|
-
* per byte, exposed here so a mechanism can turn it into a human label. */
|
|
78
|
-
export declare function unexplainedSpans(queryLen: number, accounted: ReadonlyArray<[number, number]>): Array<[number, number]>;
|
|
79
|
-
/** A human-readable label for the query bytes a mechanism's `accounted`
|
|
80
|
-
* spans leave unexplained — purely diagnostic (Task 2's negative evidence):
|
|
81
|
-
* it never changes a candidate's weight, only what the rationale trace
|
|
82
|
-
* says the mechanism left on the table. `""` when nothing is unexplained. */
|
|
83
|
-
export declare function unexplainedLabel(query: Uint8Array, accounted: ReadonlyArray<[number, number]>): string;
|
|
84
|
-
/** An open mechanism — the handle {@link Rationale.enter} returns. Hold it for
|
|
85
|
-
* the duration of the mechanism and call {@link Scope.done} with the outputs
|
|
86
|
-
* when it finishes; that emits the step and pops the nesting. */
|
|
87
|
-
export interface Scope {
|
|
88
|
-
/** The step index reserved for this mechanism at entry — pass it as an
|
|
89
|
-
* explicit dependency of a later step that consumes this one's output. */
|
|
90
|
-
readonly index: number;
|
|
91
|
-
/** Close the mechanism: emit its step with these outputs and pop it off the
|
|
92
|
-
* nesting stack. Idempotent — a second call is ignored, so a `finally` that
|
|
93
|
-
* closes after an early return is safe. */
|
|
94
|
-
done(outputs: RationaleItem[], note?: string, data?: unknown): void;
|
|
95
|
-
}
|
|
96
|
-
/** The live tracer: a stack of open mechanisms over one {@link Mind.respond}.
|
|
97
|
-
*
|
|
98
|
-
* Sema's inference is single-threaded and strictly sequential — every async
|
|
99
|
-
* step is awaited before the next begins, and `respond` holds no two thoughts
|
|
100
|
-
* at once — so a plain stack exactly tracks the current nesting: {@link enter}
|
|
101
|
-
* pushes, {@link Scope.done} pops, and {@link step} (a mechanism with no
|
|
102
|
-
* sub-steps) is the two fused. The tracer never branches the control flow; it
|
|
103
|
-
* only records it. */
|
|
104
|
-
export declare class Rationale {
|
|
105
|
-
private readonly sink;
|
|
106
|
-
private next;
|
|
107
|
-
/** Open mechanisms, outermost first. Each frame remembers the last child it
|
|
108
|
-
* has spawned so the next sibling can default its data-flow edge to it. */
|
|
109
|
-
private readonly stack;
|
|
110
|
-
/** The most recent step index emitted under each mechanism name — the handle
|
|
111
|
-
* a later step uses to name an EARLIER mechanism as its data-flow producer
|
|
112
|
-
* (e.g. cover depends on the latest recognise / computeExtensions). One tracer is
|
|
113
|
-
* built per response and inference is sequential, so "most recent" is exactly
|
|
114
|
-
* "the one that produced the inputs I am about to consume". */
|
|
115
|
-
private readonly lastByName;
|
|
116
|
-
constructor(sink: InspectRationale);
|
|
117
|
-
/** The index of the most recent step with this mechanism name, or undefined if
|
|
118
|
-
* none has run. Used to wire an explicit producer edge into {@link
|
|
119
|
-
* Scope.done} / {@link step}'s `deps`. */
|
|
120
|
-
lastIndex(name: string): number | undefined;
|
|
121
|
-
/** The mechanism names currently open, outermost → innermost. */
|
|
122
|
-
private path;
|
|
123
|
-
/** The default data-flow edge for a step entering now: the previous sibling
|
|
124
|
-
* inside the current mechanism, else the enclosing mechanism, else nothing
|
|
125
|
-
* (the root). An explicit `deps` overrides this. */
|
|
126
|
-
private defaultDeps;
|
|
127
|
-
/** Reserve this step's index and register it as the current mechanism's most
|
|
128
|
-
* recent child (so the NEXT sibling chains to it) and as the most recent step
|
|
129
|
-
* of its own NAME (so a later mechanism can name it as a producer). */
|
|
130
|
-
private reserve;
|
|
131
|
-
private emit;
|
|
132
|
-
/** Enter a mechanism that has sub-steps. Captures its inputs and the nesting
|
|
133
|
-
* now; the matching {@link Scope.done} supplies the outputs when it finishes.
|
|
134
|
-
* `deps` overrides the default data-flow edge (previous sibling / parent). */
|
|
135
|
-
enter(name: string, inputs: RationaleItem[], deps?: number[]): Scope;
|
|
136
|
-
/** Record a mechanism that has no sub-steps — its inputs and outputs are both
|
|
137
|
-
* known at the call site. Returns its index, for a later step to depend on. */
|
|
138
|
-
step(name: string, inputs: RationaleItem[], outputs: RationaleItem[], note?: string, deps?: number[], data?: unknown): number;
|
|
139
|
-
}
|