File size: 33,590 Bytes
d76dce2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
"""BitNet b1.58 ternary drug-interaction classifier.

A clean-room Python implementation of a BitNet b1.58-style ternary linear
classifier for drug-drug-interaction (DDI) severity. The forward pass is
pure integer arithmetic over Q16.16 fixed-point activations and ternary
weights ∈ {-1, 0, +1}, which makes the output **bit-identical across
architectures** (ARM, x86_64, CUDA, NPU) β€” the same reproducibility
guarantee the FDA expects for production AI / ML SaMD.

Reference: Ma, Wang, Wang et al., "The Era of 1-bit LLMs: All Large Language
Models are in 1.58 Bits," arXiv:2402.17764, 2024.

Why this layer matters in ClinicalMem
─────────────────────────────────────
The 4-tier interaction pipeline already catches known pairs deterministically
and verifies novel ones via 5-LLM US-based consensus. The BitNet layer sits at "Layer
4.5": a determinism-checked, FDA-grade classifier that:

  1. Reproduces the deterministic table's outputs bit-identically.
  2. Emits a Q16.16-scaled severity logit vector that the audit chain can
     hash into the per-decision preimage (TAG_v1 schema).
  3. Returns a `repro_hash` that any auditor with this Python file and the
     ternary weights bundle can verify against without floating-point math.

The classifier is intentionally small (200-pair training corpus, 64-dim
hidden) β€” accuracy is bounded by the deterministic table the weights are
fit to. The load-bearing claim is the *architecture*, not the absolute
accuracy: bit-identical integer arithmetic across hardware.

Public scope
────────────
This file is Apache-2.0 licensed alongside the rest of ClinicalMem. It does NOT
vendor any source from the STARGA proprietary toolchain (MindLLM,
rfn-mind, mind-runtime, mind-flow are commercial-licensed and live in
private repositories). The BitNet b1.58 architecture is described in the
public arXiv paper above; this file implements it in pure Python.

Copyright 2026 STARGA, Inc. β€” Apache-2.0 License.
"""
from __future__ import annotations

import hashlib
import json
import logging
import os
from dataclasses import dataclass
from pathlib import Path
from typing import Any

logger = logging.getLogger(__name__)

# Q16.16 fixed-point: 16 integer bits, 16 fractional bits, signed 32-bit.
# Range: [-32768.0, 32767.99999847412].
Q16_ONE: int = 1 << 16     # 65536  β€” represents 1.0
Q16_HALF: int = 1 << 15    # 32768  β€” represents 0.5
Q16_ZERO: int = 0
_Q16_MIN: int = -(1 << 31)
_Q16_MAX: int = (1 << 31) - 1

# Severity classes β€” must match the deterministic table in clinical_scoring.py
SEVERITY_NONE: int = 0
SEVERITY_MINOR: int = 1
SEVERITY_MODERATE: int = 2
SEVERITY_MAJOR: int = 3
SEVERITY_CONTRAINDICATED: int = 4

# Iter-275 v8 promotion: vocab aligned with the corpus / cache /
# trainer (`retrain_runpod/train_bitnet_v8_h256.py:39 SEV_NAMES`).
# Pre-v8 (cfadb4f6) used `(none, minor, moderate, major,
# contraindicated)` β€” the engine's first-era vocab β€” but v3+ trainers
# all use the corpus vocab `(none, moderate, serious, major,
# contraindicated)`. Engine output now matches the cache ground-truth
# vocabulary directly: a class-2 logit emits "serious" (cache match),
# not "moderate" (vocab-skewed v1 mapping).
_SEVERITY_NAMES: tuple[str, ...] = (
    "none",
    "moderate",
    "serious",
    "major",
    "contraindicated",
)


@dataclass(frozen=True)
class BitNetResult:
    """Result of a BitNet b1.58 forward pass on a drug-pair input.

    Every field is integer-valued so the audit chain can record the result
    without any float-to-string conversion ambiguity.
    """

    severity: int                      # 0..4 (see SEVERITY_* constants)
    severity_name: str                 # e.g. "major"
    logits_q16: tuple[int, ...]        # Q16.16 logit per class, in canonical class order
    feature_hash: str                  # Hex SHA-256 over the canonical input encoding
    repro_hash: str                    # Hex SHA-256 over (feature_hash, logits_q16, severity, weights_id)
    weights_id: str                    # The bundle hash recorded at load time
    deterministic_table_match: bool    # True if the weights reproduce a row in the deterministic table


# ─── Q16.16 arithmetic primitives (bit-identical across architectures) ─────

def _q16_clamp(value: int) -> int:
    """Saturating clamp into the signed 32-bit Q16.16 range."""
    if value > _Q16_MAX:
        return _Q16_MAX
    if value < _Q16_MIN:
        return _Q16_MIN
    return value


def _q16_relu(value: int) -> int:
    """Clamp negative values to zero. Pure integer compare; no float."""
    return value if value > 0 else 0


def _q16_dot_ternary(activations_q16: list[int], ternary_weights: list[int]) -> int:
    """Dot product of a Q16.16 activation vector and a ternary weight row.

    Ternary weights are one of {-1, 0, +1}. The product is the activation
    itself (or its negation, or zero) β€” no multiplication required, only
    addition and subtraction. The result is the canonical Q16.16 sum
    accumulated in row-major left-to-right order (same reduction order
    as the rest of the MIND ecosystem's deterministic kernels).

    Bit-identical guarantee: this function uses only Python's
    arbitrary-precision integers; the output is independent of the
    underlying CPU/GPU architecture, FMA ordering, and tensor-core
    accumulate semantics.
    """
    if len(activations_q16) != len(ternary_weights):
        raise ValueError(
            f"shape mismatch: act={len(activations_q16)} ternary={len(ternary_weights)}"
        )
    acc: int = 0
    for activation, weight in zip(activations_q16, ternary_weights, strict=True):
        if weight == 1:
            acc += activation
        elif weight == -1:
            acc -= activation
        # weight == 0 contributes nothing β€” skipped by design
    return _q16_clamp(acc)


# ─── Deterministic feature encoding ────────────────────────────────────────

def _encode_drug_token(rxcui_or_name: str) -> list[int]:
    """Encode a drug identifier as a 64-dim ternary feature vector ∈ {-1, 0, +1}.

    The encoding is purely deterministic: the input string is canonicalised
    (lowercased, whitespace-collapsed) and hashed with BLAKE2b. Each pair
    of bits in the digest produces one ternary feature value via a
    distribution-balanced trit table. Same string β†’ same vector on every
    machine.

    The 64-dim feature size is small enough that the full DrugBank
    interaction matrix can be linearly separated by a 64Γ—5 ternary
    classifier head; large enough that two distinct drug names hash to
    distinct vectors with negligible collision probability.
    """
    canonical = " ".join(rxcui_or_name.strip().lower().split())
    digest = hashlib.blake2b(canonical.encode("utf-8"), digest_size=16).digest()
    # 16 bytes Γ— 4 trits/byte = 64 trits. 4 trits encoded per byte using the
    # 2-bit window mapping below (50/50/25/25 distribution biased toward 0
    # so most features stay sparse β€” important for the ternary linear
    # classifier's effective rank).
    _TRIT_LOOKUP: tuple[int, ...] = (0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, -1, -1, -1, -1)
    out: list[int] = []
    for byte in digest:
        out.append(_TRIT_LOOKUP[(byte >> 0) & 0xF])
        out.append(_TRIT_LOOKUP[(byte >> 4) & 0xF])
        out.append(_TRIT_LOOKUP[byte & 0xF])
        out.append(_TRIT_LOOKUP[(byte >> 2) & 0xF])
    return out[:64]


def _q16_scale_features(ternary_features: list[int]) -> list[int]:
    """Lift a ternary feature vector to Q16.16 activations.

    Ternary {-1, 0, +1} β†’ Q16.16 {-Q16_ONE, 0, +Q16_ONE}. The scale is
    canonical so the dot products in the first linear layer accumulate
    in the standard Q16.16 range.
    """
    return [v * Q16_ONE for v in ternary_features]


# ─── Weights bundle ────────────────────────────────────────────────────────

_SCHEMA_V1 = "bitnet_classifier_v1"
_SCHEMA_V3_ATC = "bitnet_classifier_v3_atc_flags"


@dataclass(frozen=True)
class BitNetWeights:
    """Loaded ternary-weights bundle.

    Layout (matches `engine/bitnet_weights.json`):

      schema    : one of ``bitnet_classifier_v1`` (128-dim hash-only
                  encoding, hidden=64) or ``bitnet_classifier_v3_atc_flags``
                  (193-dim hash + 26 ATC flag + 13 pair-derived encoding,
                  hidden=256). Drives encoder dispatch in ``classify``.
      hidden_w  : ``hidden_features`` Γ— ``in_features`` ternary matrix.
      hidden_b  : Q16.16 bias vector (length ``hidden_features``)
      output_w  : ``out_features`` Γ— ``hidden_features`` ternary matrix.
      output_b  : Q16.16 bias vector (length ``out_features``, = 5 for
                  the 5-severity classifier)
      bundle_id : SHA-256 over the canonical JSON encoding of the four
                  matrices above (stable across loads β€” the audit chain
                  records this as the "weights_id" so a verifier can
                  pin the exact bundle a decision was made under).
    """

    hidden_w: list[list[int]]
    hidden_b: list[int]
    output_w: list[list[int]]
    output_b: list[int]
    bundle_id: str
    schema: str = _SCHEMA_V1
    in_features: int = 128
    hidden_features: int = 64
    out_features: int = 5


def _bundle_id(payload: dict[str, Any]) -> str:
    """SHA-256 over the canonical-JSON encoding of the four weight matrices.

    Stable across loads on every machine; same payload β†’ same hash.
    """
    canonical = json.dumps(
        {
            "hidden_w": payload["hidden_w"],
            "hidden_b": payload["hidden_b"],
            "output_w": payload["output_w"],
            "output_b": payload["output_b"],
        },
        sort_keys=True,
        separators=(",", ":"),
    )
    return hashlib.sha256(canonical.encode("utf-8")).hexdigest()


def load_weights(path: str | os.PathLike[str] | None = None) -> BitNetWeights:
    """Load the ternary-weights bundle from disk.

    Defaults to `engine/bitnet_weights.json` next to this file.
    """
    if path is None:
        path = Path(__file__).parent / "bitnet_weights.json"
    # DEBUG entry β€” visibility into when the bundle gets parsed
    # (rotation, first-load, etc.). Mirrors the fda_label_search_start
    # convention (PHI-safe: only path basename + size).
    raw = Path(path).read_text(encoding="utf-8")
    logger.debug(
        "bitnet_load_weights_start",
        extra={
            "path_basename": Path(path).name,
            "raw_size_bytes": len(raw),
        },
    )
    payload = json.loads(raw)

    hidden_w = [list(row) for row in payload["hidden_w"]]
    hidden_b = list(payload["hidden_b"])
    output_w = [list(row) for row in payload["output_w"]]
    output_b = list(payload["output_b"])

    meta = payload.get("_meta", {})
    schema = meta.get("schema", _SCHEMA_V1)
    if schema not in (_SCHEMA_V1, _SCHEMA_V3_ATC):
        logger.error(
            "bitnet_weights_unknown_schema",
            extra={"schema": schema, "path": str(path)},
        )
        raise ValueError(
            f"Unknown bitnet schema {schema!r}; expected one of "
            f"{_SCHEMA_V1!r}, {_SCHEMA_V3_ATC!r}"
        )

    hidden_features = len(hidden_w)
    in_features = len(hidden_w[0]) if hidden_w else 0
    out_features = len(output_w)

    meta_in = meta.get("in_features", in_features)
    meta_hidden = meta.get("hidden_features", hidden_features)
    meta_out = meta.get("out_features", out_features)

    for field, observed, declared in (
        ("in_features", in_features, meta_in),
        ("hidden_features", hidden_features, meta_hidden),
        ("out_features", out_features, meta_out),
    ):
        if observed != declared:
            logger.error(
                "bitnet_weights_meta_mismatch",
                extra={
                    "field": field,
                    "matrix_dim": observed,
                    "meta_dim": declared,
                    "path": str(path),
                },
            )
            raise ValueError(
                f"{field}: matrix dim {observed} != _meta declaration {declared}"
            )

    if any(len(row) != in_features for row in hidden_w):
        logger.error(
            "bitnet_weights_shape_mismatch",
            extra={
                "field": "hidden_w",
                "expected_cols": in_features,
                "path": str(path),
            },
        )
        raise ValueError(
            f"hidden_w rows must all be length {in_features} (drug-pair feature dim)"
        )
    if len(hidden_b) != hidden_features:
        logger.error(
            "bitnet_weights_shape_mismatch",
            extra={
                "field": "hidden_b",
                "expected_len": hidden_features,
                "actual_len": len(hidden_b),
                "path": str(path),
            },
        )
        raise ValueError(
            f"hidden_b must have {hidden_features} entries; got {len(hidden_b)}"
        )
    if out_features != 5:
        logger.error(
            "bitnet_weights_shape_mismatch",
            extra={
                "field": "output_w",
                "expected_rows": 5,
                "actual_rows": out_features,
                "path": str(path),
            },
        )
        raise ValueError(
            f"output_w must have 5 rows (one per severity class); got {out_features}"
        )
    if any(len(row) != hidden_features for row in output_w):
        logger.error(
            "bitnet_weights_shape_mismatch",
            extra={
                "field": "output_w",
                "expected_cols": hidden_features,
                "path": str(path),
            },
        )
        raise ValueError(
            f"output_w rows must all be length {hidden_features} (hidden dim)"
        )
    if len(output_b) != out_features:
        logger.error(
            "bitnet_weights_shape_mismatch",
            extra={
                "field": "output_b",
                "expected_len": out_features,
                "actual_len": len(output_b),
                "path": str(path),
            },
        )
        raise ValueError(f"output_b must have {out_features} entries; got {len(output_b)}")

    expected_in = 128 if schema == _SCHEMA_V1 else 193
    if in_features != expected_in:
        logger.error(
            "bitnet_weights_schema_dim_mismatch",
            extra={
                "schema": schema,
                "expected_in_features": expected_in,
                "actual_in_features": in_features,
                "path": str(path),
            },
        )
        raise ValueError(
            f"schema {schema!r} expects in_features={expected_in}, got {in_features}"
        )

    for matrix_name, matrix in (("hidden_w", hidden_w), ("output_w", output_w)):
        for i, row in enumerate(matrix):
            for j, weight in enumerate(row):
                if weight not in (-1, 0, 1):
                    logger.error(
                        "bitnet_weights_non_ternary",
                        extra={"matrix": matrix_name, "row": i, "col": j, "value": weight},
                    )
                    raise ValueError(
                        f"{matrix_name}[{i}][{j}] = {weight!r}; weights must be ternary"
                    )

    weights = BitNetWeights(
        hidden_w=hidden_w,
        hidden_b=hidden_b,
        output_w=output_w,
        output_b=output_b,
        bundle_id=_bundle_id(payload),
        schema=schema,
        in_features=in_features,
        hidden_features=hidden_features,
        out_features=out_features,
    )
    logger.info(
        "bitnet_weights_loaded",
        extra={
            "bundle_id": weights.bundle_id,
            "path": str(path),
            "schema": schema,
            "in_features": in_features,
            "hidden_features": hidden_features,
        },
    )
    return weights


# ─── Forward pass ──────────────────────────────────────────────────────────

def load_weights_b(path: str | os.PathLike[str] | None = None) -> BitNetWeights | None:
    """Load the optional Path B tier-2 specialist bundle (iter-421).

    Returns None if the bundle file is absent β€” callers must treat single-
    bundle mode (A-only) as the default. The specialist is trained ONLY on
    the 95 non-contra samples (4 major + 69 serious + 22 moderate); engine
    dispatch applies a constrained argmax over {moderate, serious, major}
    so it can never emit ``contraindicated`` (class 4) or ``none`` (class 0).
    """
    if path is None:
        path = Path(__file__).parent / "bitnet_weights_b_specialist.json"
    p = Path(path)
    if not p.exists():
        return None
    return load_weights(p)


def _classify_constrained_b(
    a_canonical: str,
    b_canonical: str,
    weights_b: BitNetWeights,
) -> tuple[int, tuple[int, ...]]:
    """Forward pass through B with constrained argmax over {1, 2, 3}.

    Returns ``(severity_int, logits_q16)`` where severity_int is in
    {1, 2, 3} = {moderate, serious, major}. Classes 0 (none) and 4
    (contraindicated) are masked because B was never trained on them.
    The same Q16.16 ternary kernels as ``classify`` are reused so B's
    forward pass is bit-identical across architectures alongside A's.
    """
    if weights_b.schema == _SCHEMA_V3_ATC:
        from engine.bitnet_features_v8 import encode_pair_v8
        pair_features = encode_pair_v8(a_canonical, b_canonical)
    else:
        feature_a = _encode_drug_token(a_canonical)
        feature_b = _encode_drug_token(b_canonical)
        pair_features = feature_a + feature_b

    activations_q16 = _q16_scale_features(pair_features)
    hidden_pre_q16 = [
        _q16_clamp(_q16_dot_ternary(activations_q16, weights_b.hidden_w[j]) + weights_b.hidden_b[j])
        for j in range(weights_b.hidden_features)
    ]
    hidden_q16 = [_q16_relu(v) for v in hidden_pre_q16]
    logits_q16 = [
        _q16_clamp(_q16_dot_ternary(hidden_q16, weights_b.output_w[k]) + weights_b.output_b[k])
        for k in range(weights_b.out_features)
    ]
    # Constrained argmax over classes {1, 2, 3} only. Ties broken by
    # lower index (consistent with the unconstrained argmax in classify).
    severity = 1
    best_logit = logits_q16[1]
    for k in (2, 3):
        if logits_q16[k] > best_logit:
            best_logit = logits_q16[k]
            severity = k
    return severity, tuple(logits_q16)


def classify(
    drug_a: str,
    drug_b: str,
    weights: BitNetWeights,
    *,
    deterministic_table_severity: int | None = None,
    weights_b: BitNetWeights | None = None,
) -> BitNetResult:
    """Ternary classifier forward pass: drug-pair -> severity class.

    The pair is order-canonicalised (lex sort) so {warfarin, ibuprofen} and
    {ibuprofen, warfarin} produce the same feature vector and the same
    output. The logits and severity decision are reproducible bit-for-bit
    on any platform that runs Python.

    If `deterministic_table_severity` is provided (a known result from
    `engine.clinical_scoring`'s 4-tier deterministic table), the result's
    `deterministic_table_match` flag records whether the BitNet output
    agrees. Disagreement is a release-blocking event β€” surfaces in the
    `tests/test_engine/test_bitnet_classifier.py` regression set.
    """
    a_canonical, b_canonical = sorted((drug_a, drug_b))
    if weights.schema == _SCHEMA_V3_ATC:
        from engine.bitnet_features_v8 import encode_pair_v8
        pair_features = encode_pair_v8(a_canonical, b_canonical)
    else:
        feature_a = _encode_drug_token(a_canonical)
        feature_b = _encode_drug_token(b_canonical)
        pair_features = feature_a + feature_b
    if len(pair_features) != weights.in_features:
        raise RuntimeError(
            f"internal error: pair features length {len(pair_features)} != "
            f"weights.in_features {weights.in_features}"
        )

    activations_q16 = _q16_scale_features(pair_features)
    feature_hash = hashlib.sha256(
        bytes((v + 1) for v in pair_features)        # ternary -> {0,1,2}
    ).hexdigest()

    # First linear layer: in_features -> hidden_features
    hidden_pre_q16 = [
        _q16_clamp(_q16_dot_ternary(activations_q16, weights.hidden_w[j]) + weights.hidden_b[j])
        for j in range(weights.hidden_features)
    ]
    hidden_q16 = [_q16_relu(v) for v in hidden_pre_q16]

    # Second linear layer: hidden_features -> out_features (5 severity classes)
    logits_q16 = [
        _q16_clamp(_q16_dot_ternary(hidden_q16, weights.output_w[k]) + weights.output_b[k])
        for k in range(weights.out_features)
    ]

    # Argmax β€” pure integer compare; ties broken by lower-index class.
    severity = 0
    best_logit = logits_q16[0]
    for k in range(1, 5):
        if logits_q16[k] > best_logit:
            best_logit = logits_q16[k]
            severity = k

    # iter-421 Path B cascade: when a tier-2 specialist bundle is supplied,
    # A's contraindicated verdict ALWAYS wins (frozen FDA-grade contra
    # gate, 100% recall + 0 FP). For all non-contra A predictions, B's
    # constrained argmax over {moderate, serious, major} replaces A's
    # raw argmax. B was trained without contra anchors, so its capacity
    # is fully spent on the non-contra discrimination v8 historically
    # under-fit (84% serious / 91% moderate standalone).
    weights_id_for_audit = weights.bundle_id
    logits_q16_b: tuple[int, ...] | None = None
    if weights_b is not None and severity != 4:
        # 4 = contraindicated; preserve A's contra verdict.
        b_severity, logits_q16_b = _classify_constrained_b(
            a_canonical, b_canonical, weights_b
        )
        severity = b_severity
        # Audit-chain: composite weights_id captures both bundle hashes
        # so a verifier can replay the cascade decision exactly.
        weights_id_for_audit = f"{weights.bundle_id}+{weights_b.bundle_id}"

    repro_hash_payload = {
        "feature_hash": feature_hash,
        "logits_q16": logits_q16,
        "severity": severity,
        "weights_id": weights_id_for_audit,
    }
    if logits_q16_b is not None:
        repro_hash_payload["logits_q16_b"] = list(logits_q16_b)
        repro_hash_payload["bundle_id_b"] = weights_b.bundle_id  # type: ignore[union-attr]
    repro_hash = hashlib.sha256(
        json.dumps(
            repro_hash_payload,
            sort_keys=True,
            separators=(",", ":"),
        ).encode("utf-8")
    ).hexdigest()

    deterministic_match = True
    if deterministic_table_severity is not None:
        deterministic_match = (severity == deterministic_table_severity)

    # Audit-grade trace: structured DEBUG log so production INFO-level
    # surfaces stay quiet but a reviewer can opt in by raising verbosity.
    # iter-309 PHI fix: replace raw drug_a/drug_b with 16-char SHA-256
    # pair_hash_prefix (lex-sorted canonical form). Same iter-291 /
    # iter-284 / iter-279 PHI discipline class β€” drug-pair identity stays
    # grep-able for forensic correlation but raw names never reach handlers.
    # Pre-iter-309 this event leaked drug_a + drug_b on EVERY classification
    # (live since the iter-72-era classifier landing); caught by audit
    # because both keys are absent from the iter-240 forbidden-extras-keys
    # list (which is now extended in iter-309 to catch this regression class).
    _pair_hash_prefix = hashlib.sha256(
        f"{a_canonical}+{b_canonical}".encode("utf-8")
    ).hexdigest()[:16]
    # iter-432 observability ratchet: categorical `ensemble_path` field
    # disambiguates the 3 dispatch states a forensic reader otherwise
    # has to reverse-engineer from `weights_id` length + `ensemble_active`:
    #   - "cascade_fired"        : A predicted non-contra AND B was loaded;
    #                              B's constrained argmax replaced A's class.
    #   - "a_only_contra_veto"   : A predicted contra (severity=4); B was
    #                              available but bypassed by the safety
    #                              contract (A's contra ALWAYS wins).
    #   - "a_only_no_b"          : B was not loaded (single-bundle mode);
    #                              ensemble cascade unreachable for this
    #                              classification regardless of A's output.
    # Strict subset of the existing `ensemble_active` bool β€” preserved
    # alongside for backwards compat with parsers built pre-iter-432.
    if logits_q16_b is not None:
        _ensemble_path = "cascade_fired"
    elif weights_b is None:
        _ensemble_path = "a_only_no_b"
    else:
        # weights_b supplied AND severity == 4 (contra) AND no logits_q16_b:
        # cascade was bypassed by the contra-veto safety contract.
        _ensemble_path = "a_only_contra_veto"
    logger.debug(
        "bitnet_classified",
        extra={
            "pair_hash_prefix": _pair_hash_prefix,
            "severity": severity,
            "severity_name": _SEVERITY_NAMES[severity],
            "repro_hash": repro_hash,
            "weights_id": weights_id_for_audit,
            "deterministic_match": deterministic_match,
            "ensemble_active": logits_q16_b is not None,
            "ensemble_path": _ensemble_path,
        },
    )

    return BitNetResult(
        severity=severity,
        severity_name=_SEVERITY_NAMES[severity],
        logits_q16=tuple(logits_q16),
        feature_hash=feature_hash,
        repro_hash=repro_hash,
        weights_id=weights_id_for_audit,
        deterministic_table_match=deterministic_match,
    )


# ─── Convenience wrapper for the consensus pipeline ────────────────────────

import threading

_CACHED_WEIGHTS: BitNetWeights | None = None
_PINNED_BUNDLE_ID: str | None = None
# iter-421 Path B: tier-2 specialist cache + pin (parallel to A's cache).
# When the bundle file is absent the slot stays None and the engine falls
# back to single-bundle mode automatically.
_CACHED_WEIGHTS_B: BitNetWeights | None = None
_PINNED_BUNDLE_ID_B: str | None = None
_B_LOAD_ATTEMPTED: bool = False
_CACHE_LOCK = threading.Lock()


class WeightsTamperError(RuntimeError):
    """Raised when the on-disk weights bundle's bundle_id no longer matches
    the value pinned at first load. Indicates the file was swapped under
    the running process β€” a release-blocking integrity violation."""


def reload_weights() -> BitNetWeights:
    """Force a fresh load + re-pin. Use after a confirmed weights rotation."""
    global _CACHED_WEIGHTS, _PINNED_BUNDLE_ID
    global _CACHED_WEIGHTS_B, _PINNED_BUNDLE_ID_B, _B_LOAD_ATTEMPTED
    with _CACHE_LOCK:
        previous_id = _PINNED_BUNDLE_ID
        previous_id_b = _PINNED_BUNDLE_ID_B
        weights = load_weights()
        _CACHED_WEIGHTS = weights
        _PINNED_BUNDLE_ID = weights.bundle_id
        # iter-421 Path B: re-pin tier-2 specialist alongside A. If the
        # bundle disappears between rotations, ensemble drops to A-only.
        _B_LOAD_ATTEMPTED = True
        _CACHED_WEIGHTS_B = load_weights_b()
        _PINNED_BUNDLE_ID_B = (
            _CACHED_WEIGHTS_B.bundle_id if _CACHED_WEIGHTS_B is not None else None
        )
    logger.warning(
        "bitnet_weights_reloaded",
        extra={
            "previous_bundle_id": previous_id,
            "new_bundle_id": weights.bundle_id,
            "previous_bundle_id_b": previous_id_b,
            "new_bundle_id_b": _PINNED_BUNDLE_ID_B,
        },
    )
    return weights


def classifier_layer(drug_a: str, drug_b: str) -> BitNetResult:
    """Layer-4.5 entry point used by `engine.consensus_engine`.

    The first call loads the weights bundle and pins its `bundle_id`.
    **Every subsequent call re-loads the bundle and verifies the
    `bundle_id` still matches** β€” an inexpensive SHA-256 compare that
    closes the security gap where a tampered `bitnet_weights.json` swapped
    on disk would silently produce wrong severity verdicts for the entire
    process lifetime. A mismatch raises `WeightsTamperError`, which the
    pipeline must treat as release-blocking.

    The cache is guarded by a `threading.Lock` so high-throughput clinical
    deployments (10K+ pairs/min, multi-threaded) cannot trigger the
    thundering-herd race that would otherwise re-parse the JSON on every
    contended call.
    """
    global _CACHED_WEIGHTS, _PINNED_BUNDLE_ID
    global _CACHED_WEIGHTS_B, _PINNED_BUNDLE_ID_B, _B_LOAD_ATTEMPTED
    with _CACHE_LOCK:
        if _CACHED_WEIGHTS is None:
            _CACHED_WEIGHTS = load_weights()
            _PINNED_BUNDLE_ID = _CACHED_WEIGHTS.bundle_id
            # iter-421 Path B: opportunistic tier-2 load + pin (audit-clean
            # β€” same SHA-256-canonical-JSON integrity primitive as A). Absent
            # bundle leaves the slot None and engine falls back to A-only.
            if not _B_LOAD_ATTEMPTED:
                _B_LOAD_ATTEMPTED = True
                _CACHED_WEIGHTS_B = load_weights_b()
                if _CACHED_WEIGHTS_B is not None:
                    _PINNED_BUNDLE_ID_B = _CACHED_WEIGHTS_B.bundle_id
                    logger.info(
                        "bitnet_classifier_b_load_pinned",
                        extra={
                            "bundle_id_b_prefix": _PINNED_BUNDLE_ID_B[:16],
                            "hidden_features_b": len(_CACHED_WEIGHTS_B.hidden_w),
                        },
                    )
            # First-load pinning event β€” fires ONCE per process. Lets
            # auditors correlate every BitNetResult emitted in the
            # process to the bundle_id that was pinned at startup.
            # bundle_id is SHA-256-of-canonical-JSON, NOT secret material;
            # safe to log in full (it IS the integrity primitive).
            logger.info(
                "bitnet_classifier_first_load_pinned",
                extra={
                    "bundle_id_prefix": _PINNED_BUNDLE_ID[:16],
                    "hidden_features": len(_CACHED_WEIGHTS.hidden_w),
                    "in_features": (
                        len(_CACHED_WEIGHTS.hidden_w[0])
                        if _CACHED_WEIGHTS.hidden_w else 0
                    ),
                    "out_features": len(_CACHED_WEIGHTS.output_w),
                    "ensemble_active": _CACHED_WEIGHTS_B is not None,
                },
            )
        else:
            # Re-load + verify the pinned bundle_id on every call. The full
            # JSON parse is ~1 ms; the alternative is a class of FDA-blocking
            # silent-tampering bugs.
            current = load_weights()
            if current.bundle_id != _PINNED_BUNDLE_ID:
                # Pre-raise structured CRITICAL β€” this is a release-blocking
                # FDA SaMD integrity violation. Bundle IDs are SHA-256
                # prefixes of the canonical-JSON weights file, safe to log
                # (they ARE the integrity primitive, not secret material).
                logger.critical(
                    "bitnet_weights_tamper_detected",
                    extra={
                        "pinned_bundle_id": (
                            _PINNED_BUNDLE_ID[:16] if _PINNED_BUNDLE_ID else None
                        ),
                        "on_disk_bundle_id": current.bundle_id[:16],
                    },
                )
                raise WeightsTamperError(
                    f"bitnet_weights.json bundle_id changed under the running "
                    f"process: pinned {_PINNED_BUNDLE_ID[:16]}... "
                    f"on-disk {current.bundle_id[:16]}... β€” call "
                    f"reload_weights() after a deliberate rotation."
                )
            _CACHED_WEIGHTS = current
            # iter-421 Path B: same tamper check on B if it was pinned.
            if _PINNED_BUNDLE_ID_B is not None:
                current_b = load_weights_b()
                if current_b is None or current_b.bundle_id != _PINNED_BUNDLE_ID_B:
                    logger.critical(
                        "bitnet_weights_b_tamper_detected",
                        extra={
                            "pinned_bundle_id_b": _PINNED_BUNDLE_ID_B[:16],
                            "on_disk_bundle_id_b": (
                                current_b.bundle_id[:16] if current_b else None
                            ),
                        },
                    )
                    raise WeightsTamperError(
                        f"bitnet_weights_b_specialist.json bundle_id changed under "
                        f"the running process: pinned {_PINNED_BUNDLE_ID_B[:16]}... "
                        f"β€” call reload_weights() after a deliberate rotation."
                    )
                _CACHED_WEIGHTS_B = current_b
        return classify(drug_a, drug_b, _CACHED_WEIGHTS, weights_b=_CACHED_WEIGHTS_B)


__all__ = [
    "BitNetResult",
    "BitNetWeights",
    "Q16_ONE",
    "Q16_HALF",
    "Q16_ZERO",
    "SEVERITY_NONE",
    "SEVERITY_MINOR",
    "SEVERITY_MODERATE",
    "SEVERITY_MAJOR",
    "SEVERITY_CONTRAINDICATED",
    "WeightsTamperError",
    "classify",
    "classifier_layer",
    "load_weights",
    "load_weights_b",
    "reload_weights",
]