Repository navigation
Expand file tree
/
Copy pathinventory.yaml
More file actions
1044 lines (1040 loc) · 70.1 KB
/
Copy pathinventory.yaml
File metadata and controls
1044 lines (1040 loc) · 70.1 KB
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
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Constraint coverage inventory (issue #6, first deliverable). Descriptive, not a soundness claim,
# and not complete: it lists the elements named below, and `excluded` names the research directories
# it does not cover, with the reason. The checker fails if a research directory is neither cited nor
# excluded; for the shipped verifier, completeness rests on reading and on review.
#
# Each row maps one element of the relation that a FlashProve proof is about (an admitted operator,
# a range or rounding rule, a committed input or state, an output position) to the place in this
# repository that enforces it, and says what kind of enforcement that is:
#
# proof_constraint intended to be enforced inside the proof: a sub-proof's sumcheck, lookup,
# range check, or the hash chain itself. That the constraint is correct or
# sufficient is not established here.
# verifier_validation checked by the verifier program or the client's integration on public values,
# outside the proof's relation (parameter conditions, an ID comparison, a policy or
# range comparison with what the caller states); a complete proof of another
# statement passes the proof part and is stopped only here, if the caller states
# the right values
# prover_only done only by an honest committer or prover (cp_commit, cp_prove); nothing in a
# proof, key file or commitment shows a verifier that it happened
# not_enforced nothing in the repository enforces it; the proof says nothing about it
#
# scope: shipped (cp_verify, cp_commit, cp_prove, integrations/ on main) or research (the fork's operator
# prototypes under research/operators/, which are not in cp_verify).
#
# sites: file + symbol. `symbol` is a literal string of at least 12 characters that must occur in
# `file`. The checker (check_inventory.py, run in CI by claims-check) fails if a file is missing or a
# symbol no longer occurs in it, so a rename or deletion of a cited site is noticed. It does not see a
# row moved to a stronger kind, a check disabled behind unchanged text, or a row losing one of two
# sites. A site is where the enforcement lives, or, for prover_only and not_enforced rows, the place
# that documents the absence. A site says where; it does not show that the check there is right.
#
# Base: main ba1e3067f58b0142dd16a3df884626cfe6ccc585 for rows OP-01 .. OUT-07 (first version, PR #97
# head 655d70b2). Rows added after review (OP-15, OP-R8 .. OP-R14, RR-15 .. RR-21, RR-R1, RR-R2, IS-23, IS-24,
# OUT-08, OUT-09) were read at main 0bdf8326, where the cited files are unchanged from ba1e3067. Read
# from source and docs; only rows N1 to N3 of experiments/20261110-6-coverage-inventory/ were run.
# Rows OP-R15, OP-R16, RR-R3, RR-R4, IS-25 .. IS-27 (research/operators/qwen_conv_history, PR #101) were
# read at main 8d579337 from the code and experiments/20261111-35-conv-history-binding/RESULTS.md; nothing
# was rerun for them.
# Row OP-R19 (research/operators/qwen_ops_followups, PR #115, issue #94) was read at main 7c40dd9f from the
# code and experiments/20261111-94-op-followups/RESULTS.md.
schema_version: 1
base_commit: ba1e3067f58b0142dd16a3df884626cfe6ccc585
kinds: [proof_constraint, verifier_validation, prover_only, not_enforced]
categories: [operator, range_rounding, input_state, output]
rows:
# ---------------------------------------------------------------- operators, shipped integer model
- id: OP-01
category: operator
scope: shipped
element: "layer norm N(x; M, K, eps) before attention and feed-forward (root of the sum of squares, rounded reciprocal, rounded product, clamp)"
relation: "docs/model.md section 5; lookup segments N1, N2 (quotient), N1H, N2H (high part), CN1, CN2 (clamp), ROW (the norm's digit row of 9 + 8 equations)"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "N1" || sn == "N2") {'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "CN1" || sn == "CN2") {'}
- {file: verifier/src/vf_norm_digit_rows.h, symbol: 'template <class T> struct Row5 {'}
- {file: docs/model.md, symbol: '## 5 The norm'}
- id: OP-02
category: operator
scope: shipped
element: "final norm before the head (NF), with its clamp (CNF)"
relation: "docs/model.md section 8 aF_t; output head lookup segments NF, NFH, CNF"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_output_head_check.inc, symbol: 'if (sn == "NF")'}
- {file: verifier/src/vf_output_head_check.inc, symbol: 'if (sn == "CNF") {'}
- id: OP-03
category: operator
scope: shipped
element: "query and key projections with rotary embedding, rounding shift and clamp"
relation: "docs/model.md section 6 step 2; segments QK, QKH (quotient, high part), CQK (clamp); rotary table C, S"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "QK" || sn == "QKH") {'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "CQK")'}
- id: OP-04
category: operator
scope: shipped
element: "value projection with rounding shift and clamp"
relation: "docs/model.md section 6 step 3; segments V, VH, CV"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "V") return site('}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "CV") return'}
- id: OP-05
category: operator
scope: shipped
element: "causal attention: scores, row maximum, exponential table EX, sum Z, accumulator, rounded division o = floor((2 acc + Z) / (2 Z))"
relation: "docs/model.md section 6 step 4; the attention sub-proof (SC3, SCD sumchecks, its lookup forest, claim 5 the division)"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_attention_check.h, symbol: 'struct AttentionSubProof {'}
- {file: verifier/src/vf_attention_check.h, symbol: 'attention: claim 5: the division: Z o, its remainders'}
- id: OP-06
category: operator
scope: shipped
element: "attention output projection and residual add"
relation: "docs/model.md section 6 step 5; segments O, OH"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "O") return site('}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "OH") return high('}
- id: OP-07
category: operator
scope: shipped
element: "feed-forward gate: wide gate key, low byte g0, clamped high part gc, SiLU table lookup"
relation: "docs/model.md section 7 'A block of hidden units'; segments G, GH, G0, CG, F"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "G0") return'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "CG") return'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "F")'}
- id: OP-08
category: operator
scope: shipped
element: "feed-forward up projection u and the product s u"
relation: "docs/model.md section 7; segments U, UH, UT"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "UT") return S("u")'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "G" || sn == "U") {'}
- id: OP-09
category: operator
scope: shipped
element: "down projection (dense, or shared plus routed expert sums) and residual add"
relation: "docs/model.md section 7 dense layer and MoE step 6; segments D, DH"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "D")'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "DH") return high('}
- id: OP-10
category: operator
scope: shipped
element: "MoE router logits: wide logit, high part ly, low 3 bits l3, clamp lr = 8 clamp(ly) + l3"
relation: "docs/model.md section 7 MoE step 1; segments LR, LH, L3, CLR"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "LR") {'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "L3") {'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "CLR") {'}
- id: OP-11
category: operator
scope: shipped
element: "MoE top-1 expert choice: argmax of lr, ties to the lowest index, one-hot selector"
relation: "docs/model.md section 7 MoE step 2; segments LT, DD, SEL and the one-hot claim; the parameter condition one-hot keeps NE Ty + 1 below p"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "DD") {'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "SEL") {'}
- {file: verifier/src/cp_verify.cpp, symbol: 'need((u128)NE * Ty + 1 < P2, "one-hot");'}
note: "the third site (one-hot) is a verifier condition on public parameters (a verifier_validation site, as in RR-04); the selection itself is in the proof (DD, SEL)"
- id: OP-12
category: operator
scope: shipped
element: "MoE routed input scale: sigmoid table Sig[lr[e*]], xr = rq(sc m, 16)"
relation: "docs/model.md section 7 MoE step 3; segments MR, XR"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "MR") return'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "XR") {'}
- id: OP-13
category: operator
scope: shipped
element: "output head: logits lam_t = aF_t Wh, the margin of each checked token in three digits, the exponential key for clipped-gap costs"
relation: "docs/model.md section 8 (logit gap); docs/format/11-4-output-head.md; segments HD0, HD1, HD2, HE0, HE1"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_output_head_check.inc, symbol: 'if (sn == "HD0" || sn == "HD1" || sn == "HD2") {'}
- {file: verifier/src/vf_output_head_check.inc, symbol: 'if (sn == "HE0" || sn == "HE1") {'}
- id: OP-14
category: operator
scope: shipped
element: "every matrix product against committed weights W (Wq, Wk, Wv, Wo, Wg, Wu, Wd, Wr, experts, Wh): claims on W merged and tied to the leaf table"
relation: "docs/protocol.md 6.3 (g M e + mu = v; shared chain; second chain); docs/format/11-5-weights-in-the-table.md"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_weights_pending_claim.h, symbol: 'struct WsPending {'}
- {file: verifier/src/cp_verify.cpp, symbol: 'blk::check_proof_if_tested('}
- id: OP-15
category: operator
scope: shipped
element: "embedding lookup x^0_t = Emb[tok_t], and its value as the first residual stream that layer 0 reads"
relation: "group 0's claim on Emb (unit X = 12): v_E = sum_t eq(ztE, t) sum_k phiE[k] Emb[tok_t][k], with the token factor built from the request's tokens; then the claim that v_E is a read of xbI at Z_E"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: "// The embedding's claim (unit X = 12)"}
- {file: verifier/src/vf_layers_check.inc, symbol: "the embedding's value is a read of the first stream"}
note: "RR-02 covers only the range of Emb's digits; this row is the lookup itself (added after review of PR #97)"
# ---------------------------------------------------------------- operators, research prototypes
- id: OP-R1
category: operator
scope: research
element: "same-oracle sparse affine rows (research/operators/affine_rows)"
relation: "verifier-fixed affine relation over the committed product oracle; honest and two complete malicious proofs (README 'Two complete malicious proofs'); reached main through PR #20 (PR #15 was closed unmerged)"
enforcement: proof_constraint
sites:
- {file: research/operators/affine_rows/product.h, symbol: 'inline Row affine_row(OracleId oracle,const AffineRelation &a) {'}
- {file: research/operators/affine_rows/README.md, symbol: 'Trusted mutable in-memory runs/plans are not safe untrusted parsers'}
note: "the run and plan objects come from a trusted constructor; invalid constructions abort (SIGABRT), there is no untrusted statement decoder. Its substrates research/operators/public_binding and shared_wires are under excluded"
- id: OP-R2
category: operator
scope: research
element: "MoE bias-corrected top-k routing gadget (research/operators/moe_topk): selection, normalization, aggregation relations"
relation: "research/operators/moe_topk/README.md relations (2) to (6)"
enforcement: proof_constraint
sites:
- {file: research/operators/moe_topk/README.md, symbol: 'Why the selection is unique'}
- {file: research/operators/moe_topk/test.cpp, symbol: '// (2) one-hot selection: Boolean P[m,i]; P*K and P*s gathers.'}
- {file: research/operators/moe_topk/test.cpp, symbol: '// (6) threshold: u_i = (1 - sel_i) * (G_{k-1} - K_i - 1) in [0, 2^26).'}
- {file: research/operators/moe_topk/test.cpp, symbol: '// (8) aggregation: XG[m][d] = sum_i P[m,i] X_i[d]; y[d] = floor(sum_m w_m XG[m][d] / 2^12).'}
- id: OP-R3
category: operator
scope: research
element: "MoE top-k gadget input ranges: bias -2^15 <= b_i < 2^15 and expert output |X| < 2^15"
relation: "statement-side admission, not a relation"
enforcement: verifier_validation
sites:
- {file: research/operators/moe_topk/test.cpp, symbol: 'static bool admissible(const Fixture &z) {'}
- {file: research/operators/moe_topk/README.md, symbol: 'Two inputs are bounded by statement-side admission'}
- id: OP-R4
category: operator
scope: research
element: "Qwen Gated DeltaNet head M1b (one head, T = 4) with the RNG/LKP range check"
relation: "experiments/20261105-32-qwen-deltanet-m1b-lookup; lookup_verify"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_deltanet_lookup/lookup.h, symbol: 'inline bool lookup_verify('}
- id: OP-R5
category: operator
scope: research
element: "Qwen M2 derived-value admission rule (reject out-of-class recomputed k, q, a, beta, v, cg)"
relation: "experiments/20261105-32-qwen-deltanet-m2-design DESIGN.md sections 3.1, 4.5, 7"
enforcement: not_enforced
sites:
- {file: research/operators/qwen_deltanet_m2/README.md, symbol: 'therefore has **no C++ implementation here**'}
- id: OP-R6
category: operator
scope: research
element: "Qwen bounded operator relations (conv, nonlinear tables, norm, rotary): binding of the conv kernel, its history and every input to a model"
relation: "research/operators/qwen_ops/README.md: public values are operator, F, rotary position, outputs only"
enforcement: not_enforced
sites:
- {file: research/operators/qwen_ops/README.md, symbol: 'Committed, chosen freely by the prover'}
- id: OP-R7
category: operator
scope: research
element: "first top-k constraint prototype (research/operators/topk.py): explicit quadratic constraints checked on a disclosed witness"
relation: "research/operators/README.md exact relation v0; constraint satisfaction, no proof"
enforcement: not_enforced
sites:
- {file: research/operators/README.md, symbol: 'Status: executable **constraint satisfaction**, not a cryptographic proof'}
note: "checks a witness; there is no proof or verifier, so nothing is enforced against a prover"
- id: OP-R8
category: operator
scope: research
element: "GLM KDA algebra slice (research/operators/glm_kda_slice): two-token transition over field values, with the second token reading the first token's committed state cells"
relation: "30 product gates on the shared product oracle, equality rows, affine rows for sums, coordinate rows for public inputs, outputs and final state; research verifier, not cp_verify"
enforcement: proof_constraint
sites:
- {file: research/operators/glm_kda_slice/README.md, symbol: 'committed state cells: token two cannot substitute an unrelated state.'}
- {file: research/operators/glm_kda_slice/test.cpp, symbol: 'Wrong subtraction, disconnected second-token state, wrong second delta.'}
- id: OP-R9
category: operator
scope: research
element: "GLM KDA algebra slice: rounding, saturation, integer ranges and fixed-point scale"
relation: "none; the slice is exact field algebra"
enforcement: not_enforced
sites:
- {file: research/operators/glm_kda_slice/README.md, symbol: 'There is no rounding, saturation, range enforcement or fixed-point scale here.'}
- id: OP-R10
category: operator
scope: research
element: "GLM KDA signed staged-floor micro-slice (research/operators/glm_kda_fixedpoint): each rounded product a*b = 4q + r with signed q in [-8, 7], r in [0, 3]; signed four-bit range decomposition of every value"
relation: "56 product gates, 42 wire equalities, 57 affine relations on the shared product oracle; Boolean bits by b*b = b; research verifier, not cp_verify"
enforcement: proof_constraint
sites:
- {file: research/operators/glm_kda_fixedpoint/README.md, symbol: 'Each rounded multiplication binds `a*b = 4*q+r`'}
- {file: research/operators/glm_kda_fixedpoint/README.md, symbol: 'Signed range decomposition binds `x+8 = sum(2^i*b_i)`'}
- {file: research/operators/glm_kda_fixedpoint/test.cpp, symbol: 'signed staged-floor scalar KDA'}
note: "one fixed toy statement (public S = O = -3), not the publisher's numerics"
- id: OP-R11
category: operator
scope: research
element: "Qwen nonlinear tables (research/operators/qwen_ops): input floored to the table grid, 4096-way one-hot selector, six outputs read from one pinned table row"
relation: "c.floor(x, F - 8), selectors(key), lookup(sel, table) on the shared product oracle; the verifier rebuilds the relation list from its own operator definition"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'static Circuit nonlinear(const Input& in,bool witness) {'}
- {file: research/operators/qwen_ops/README.md, symbol: 'A 4096-way one-hot selector has Boolean entries'}
note: "the input is committed and chosen by the prover (OP-R6)"
- id: OP-R12
category: operator
scope: research
element: "Qwen depthwise convolution, one channel, T = 4: products of history/input cells with kernel cells, floored sum, output range"
relation: "qwen_ops conv(): mul, floor(sum, F), range(y, F + 4)"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'static Circuit conv(const Input& in,bool witness) {'}
note: "kernel, history and inputs are committed and chosen by the prover (OP-R6)"
- id: OP-R13
category: operator
scope: research
element: "Qwen L2 / RMS norm of one head: sum of squares, floored mean-square key, table coefficient, floored scaled outputs"
relation: "qwen_ops norm(): no gain, no epsilon"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'static Circuit norm(const Input& in,bool witness) {'}
note: "inputs committed and chosen by the prover (OP-R6); not the M2 isqrt norm"
- id: OP-R17
category: operator
scope: research
element: "Qwen q/k RMSNorm per head (research/operators/qwen_qk_norm, issue #35): sigma = isqrt(4^m sum x^2 + epsN), R = floor(2^K / sigma), Z = floor(x R / 2^s), y = floor(g Z / 2^F), learned per-channel gain g, epsilon, Hq = 2 q heads with gq and Hk = 1 k head with gk, D in {16, 256}, F = 12"
relation: "build(): two isqrt residual equations, R sigma + rr = 2^K with sigma - 1 - rr = gap, floor equations with bit-decomposed remainders; gains are verifier-side coefficients and y the targets; the verifier rebuilds the relation list from the statement"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: 'static Circuit build(const Stmt& st,const std::vector<HeadWit>* wit) {'}
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: '// 4^m E + epsN - sigma^2 = e1'}
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: '// R sigma + rr = 2^K'}
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: '// g Z - ry = 2^F y (g, y public)'}
note: "not the qwen_ops table norm (OP-R13) and not the M2 gated norm; sampled hostile classes in experiments/20261111-35-qk-rmsnorm, not exhaustive; research security 10 bits per round, toy coins; not in cp_verify"
- id: OP-R18
category: operator
scope: research
element: "qwen_qk_norm: binding of the pre-norm inputs to a q/k projection and of the gains to the model's q_norm / k_norm weights"
relation: "none; inputs and gains are public statement fields the caller supplies"
enforcement: not_enforced
sites:
- {file: research/operators/qwen_qk_norm/README.md, symbol: '**The inputs are public and pinned in-proof.**'}
- {file: research/operators/qwen_qk_norm/README.md, symbol: '**The gains are public statement fields with an admitted bound.**'}
- id: OP-R14
category: operator
scope: research
element: "Qwen rotary: first 64 of 256 coordinates rotated with pinned cos/sin at a public position, floored; the rest copied"
relation: "qwen_ops rotary(): coefficients from the pinned rotary table at the public position"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'static Circuit rotary(const Input& in,bool witness) {'}
note: "inputs committed and chosen by the prover (OP-R6)"
- id: OP-R15
category: operator
scope: research
element: "Qwen conv1d with carried history (research/operators/qwen_conv_history, issue #35): one channel, four taps, three history cells, four new inputs, y[t] = floor(sum_j w[j] * input[t + j] / 2^F); the PR #67 conv relation copied, with a bound and an unbound profile in one binary"
relation: "conv(): mul, floor(sum, F), range(y, F + 4), y pinned to the statement's outputs; the verifier rebuilds the relation list from its own statement (at F = 12: bound profile 294 columns, 579 relations; unbound profile 573 relations, as PR #67)"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'static Circuit conv(const CStmt& st,const std::vector<i64>* wit) {'}
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'for(int j=0;j<4;j++) products.push_back({c.mul(x[t+j],w[j]),1});'}
- {file: research/operators/qwen_conv_history/README.md, symbol: 'keeps the #67 relation unchanged'}
note: "a copy of the qwen_ops conv circuit, not the #67 binary (README lists the differences); the unbound profile is kept as the named negative control and accepts the #67 replaced-history construction by design. Kernel and inputs are free (OP-R16)"
- id: OP-R16
category: operator
scope: research
element: "qwen_conv_history: binding of the conv kernel (taps) and of the new inputs X[0..3] to a model"
relation: "none; taps and new inputs are committed cells the prover chooses. The taps (0, 0, 0, -S) construction (y[t] = -X[t]) still verifies; the tail opening only ties the next segment to this segment's own inputs"
enforcement: not_enforced
sites:
- {file: research/operators/qwen_conv_history/README.md, symbol: '**The kernel is still free.**'}
- {file: research/operators/qwen_conv_history/README.md, symbol: '**Inputs have no provenance.**'}
- id: OP-R19
category: operator
scope: research
element: "identity of the pinned Qwen operator tables (research/operators/qwen_ops_followups, issue #94): profile v2 of the PR #67 relations, with the two table manifest SHA-256 values and the SHA-256 of each table the verifier reads in the configuration text"
relation: "v2_text(): configuration text enters the configuration digest and so the transcript's first query; the per-table hashes are computed from the verifier's compiled-in arrays. v1 (PR #67) is unchanged and has no table line"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_ops_followups/follow.cpp, symbol: 'static std::string v2_text(const Circuit& c,const std::vector<NamedTable>& tabs,const Claim& claim) {'}
- {file: research/operators/qwen_ops_followups/README.md, symbol: '## Item 1: table identity in the configuration digest (profile v2)'}
note: "binds the verifier's own tables into the transcript; it does not make a table correct and does not bind inputs, kernel or history (the follow-up suite keeps those as accepted gap controls)"
# ---------------------------------------------------------------- range and rounding rules
- id: RR-01
category: range_rounding
scope: shipped
element: "every committed weight (all kinds but the embedding) in -2^12 .. 2^12 - 1"
relation: "key file range proof WF; WEIGHT_BOUND"
enforcement: proof_constraint
sites:
- {file: blocks/frame/check_key.h, symbol: 'the range proof WF, blocks/iop/wf.h'}
- {file: verifier/src/vf_widths_and_bounds.h, symbol: 'WEIGHT_BOUND = 1u << (WEIGHT_WIDTH_BITS - 1)'}
- id: RR-02
category: range_rounding
scope: shipped
element: "embedding weights: only their base-256 digits are range-checked; the head's no-overflow condition J-E assumes the digit bound"
relation: "docs/protocol.md 6.1 step 2; condition J-E"
enforcement: proof_constraint
sites:
- {file: docs/protocol.md, symbol: 'For the embedding, however, only the digits are checked.'}
- {file: verifier/src/cp_verify.cpp, symbol: 'need(TyX + pw(LBZW + 1) * 16843009ull < P2, "J-E");'}
note: "digits in the proof; J-E is a verifier condition on public parameters; docs/model.md section 9 argues the range conditions do not narrow Emb"
- id: RR-03
category: range_rounding
scope: shipped
element: "every committed trace cell small, |cell| < Ty = 2^18 (approximate range check, 'smallness')"
relation: "random -1/0/1 combinations Z bounded by 2^17 via a digit lookup; docs/protocol.md 6.3"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_smallness_test.h, symbol: 'struct HTailF2 {'}
- {file: verifier/src/vf_common.h, symbol: "LBZ: a request's range checks bound Z"}
note: "probabilistic and approximate: it shows a looser bound than the cells' intended range, except with small probability over the factors"
- id: RR-04
category: range_rounding
scope: shipped
element: "no-overflow conditions: field equations over p = 2^64 - 2^32 + 1 also hold over the integers"
relation: "tests on public parameters (no-wrap-*, af-no-wrap-N1..N9, one-hot, J-E)"
enforcement: verifier_validation
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'if (!ok) refuse("no-wrap-" + n);'}
- {file: verifier/src/vf_attention_shape.h, symbol: 'refuse("af-no-wrap-N1")'}
- {file: docs/protocol.md, symbol: "it is not proved that the programs' conditions are enough for the programs' equations"}
note: "the tests exist; that they are sufficient is argued from reading, not proved (open obligation G-OB1)"
- id: RR-05
category: range_rounding
scope: shipped
element: "rounding shift rq(x, k): quotient plus a remainder in range, at every site (G, U, QK, V, O, D, LR, norms)"
relation: "the site() form: quotient, remainder lookup; the high() form for the remainder's top part"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'auto site = [&](const Ex &w, const Aff &SQ, u32 k, int form, const std::function<Aff()> &Srh) {'}
- {file: verifier/src/vf_layers_check.inc, symbol: 'auto high = [&](const Ex &w, const Aff &Srh, u32 k, int form) {'}
- id: RR-06
category: range_rounding
scope: shipped
element: "clamp to -2047 .. 2047 (table FC) of norm outputs, q, k, v, the final norm"
relation: "key aw + 2^15 + 2^28 (a + 2^11) in table FC"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: "Lookup table FC's row for x"}
- {file: verifier/src/vf_output_head_check.inc, symbol: 'is meant to show a = clamp(aw; 2,047)'}
- id: RR-07
category: range_rounding
scope: shipped
element: "up projection |u| <= 2^15 (range condition 7)"
relation: "UT key u + 2^LBU in a 2^16-row table"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "UT") return S("u") + affC(pw(LBU) * w);'}
- {file: verifier/src/vf_widths_and_bounds.h, symbol: 'static constexpr u32 SMAXTOP = 4095, LBU = 15;'}
- id: RR-08
category: range_rounding
scope: shipped
element: "attention rounded-division remainders fit three bytes (lt + S + 1 <= 24) and softmax table parameters"
relation: "parameter condition af-policy-softmax; remainders in the attention lookup"
enforcement: verifier_validation
sites:
- {file: verifier/src/vf_attention_shape.h, symbol: 'lt + cf.S + 1 > 24) refuse("af-policy-softmax");'}
- {file: verifier/src/vf_attention_shape.h, symbol: 'remainders mod 2 Z must fit three bytes'}
note: "the size argument is 'argued from the sizes, not proved' (docs/model.md section 9)"
- id: RR-09
category: range_rounding
scope: shipped
element: "public tables SiLU, sigmoid, rotary, exponential: the verifier recomputes them and compares their digests with the parameter file's"
relation: "docs/model.md section 4 'Digests'"
enforcement: verifier_validation
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'tblHeld(0, hh, "table-digest-silu");'}
- {file: verifier/src/cp_verify.cpp, symbol: 'tblHeld(2, hh, "table-digest-rope");'}
- {file: verifier/src/cp_verify.cpp, symbol: 'tblHeld(3, hh, "table-digest-exp");'}
- id: RR-10
category: range_rounding
scope: shipped
element: "norm shift parameters 24 <= M <= 47, 17 <= K <= 32, K >= M - 14; eps in 1 .. 2^59"
relation: "docs/model.md section 9 last paragraph"
enforcement: verifier_validation
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'refuse("KL-norm");'}
- {file: verifier/src/cp_verify.cpp, symbol: 'refuse("policy-norm");'}
- id: RR-11
category: range_rounding
scope: shipped
element: "worst case over all prompts: no forward-pass value leaves its range for the committed weights and scales"
relation: "cp_commit's worst-case check from weights and scales; scales/worst-case-bounds.md"
enforcement: prover_only
sites:
- {file: prover/weights_commitment/worst_case.cuh, symbol: 'The worst-case check of the committer program cp_commit'}
- {file: docs/model.md, symbol: 'bounds, from the weights and scales alone, what any prompt can drive these values to'}
- id: RR-12
category: range_rounding
scope: shipped
element: "range conditions between the model's 'defined' interval and the provers' narrower limb interval (e.g. 32,640 .. 32,767)"
relation: "docs/model.md section 9 table, last column"
enforcement: prover_only
sites:
- {file: docs/model.md, symbol: 'The last column is narrower because the provers write these values in limbs'}
note: "the honest prover answers 'failed' there; nothing tells the client whether a refusal was this or a false claim"
- id: RR-13
category: range_rounding
scope: shipped
element: "conversion of the publisher's weights (GGUF blocks, floats) to the committed integers"
relation: "docs/model.md section 11"
enforcement: not_enforced
sites:
- {file: docs/model.md, symbol: 'checks this conversion: a proof is about the committed integers.'}
- id: RR-14
category: range_rounding
scope: shipped
element: "per-round soundness target of at least 128 bits (the verifier's own accounting)"
relation: "BAR; security-bits rejection; docs/protocol.md 6.4 'The target'"
enforcement: verifier_validation
sites:
- {file: verifier/src/vf_common.h, symbol: 'static constexpr double BAR = DW == 8 ? 128.0 : 136.0;'}
- {file: verifier/src/cp_verify.cpp, symbol: 'refuse("security-bits");'}
note: "a program's calculation, no theorem about the programs; not a security level. Not a range or rounding rule of the model: it is filed here for want of a fifth category (a parameter condition on the proof system)"
- id: RR-15
category: range_rounding
scope: shipped
element: "range condition 1: Vx + eps < 2^78 (five limbs; with four limbs Vx < 2^59), the sum of squares a norm reads"
relation: "docs/model.md section 9 condition 1; V is held only as byte digits of the norm's digit row, each digit range-checked by the row's lookup (mu x digit a byte)"
enforcement: proof_constraint
sites:
- {file: docs/model.md, symbol: '| 1 | Vx + eps | 5 | < 2^78 |'}
- {file: verifier/src/vf_norm_digit_rows.h, symbol: 'V exists only as digits'}
note: "by reading: the bound follows from the number of digits of V in the row; not separately tested here"
- id: RR-16
category: range_rounding
scope: shipped
element: "range condition 2: a norm's quotient w[k] in -32,768 .. 32,767"
relation: "the clamp key of CN1 / CN2: aw + 2^15 must be a row number of the 2^16-row table FC"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "CN1" || sn == "CN2") {'}
- {file: docs/model.md, symbol: "| 2 | a norm's quotient w[k] | 5 | -32,768 .. 32,767 |"}
note: "that the key pins the range rests on the radix conditions (guard-radix, RR-04); the provers' narrower interval is RR-12"
- id: RR-17
category: range_rounding
scope: shipped
element: "range condition 3: a wide query or key entry qw, kw in -32,768 .. 32,767"
relation: "the clamp key of CQK in table FC"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'return S("qkw") + (SBA(wh, 0, dq, "q", lam, Z, 0)'}
- {file: docs/model.md, symbol: '| 3 | a wide query or key entry, qw or kw | 6 |'}
- id: RR-18
category: range_rounding
scope: shipped
element: "range condition 4: a wide value entry vw in -32,768 .. 32,767"
relation: "the clamp key of CV in table FC"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: 'if (sn == "CV") return S("vw") + S("v") * pw(28)'}
- {file: docs/model.md, symbol: '| 4 | a wide value entry vw | 6 |'}
- id: RR-19
category: range_rounding
scope: shipped
element: "range condition 5: the high part ly of a wide router logit (with rcl 1: lw itself)"
relation: "the clamp key of CLR in table FC (rcl 1), L3 for the low bits"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: "CLR's key is FC's row for ly exactly if lr - l3 = 8 clamp(ly, +-2047)"}
- {file: docs/model.md, symbol: '| 5 | the high part ly of a wide router logit | 7 |'}
note: "MoE layers only; by reading, the form without rcl was not traced here"
- id: RR-20
category: range_rounding
scope: shipped
element: "range condition 6: the high part gh of a wide gate key in -16,384 .. 16,383 (with gcl 1: gw)"
relation: "the key of CG must be a row of the clamp half of table FS (gcl 1); without gcl, the gate key in F"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_layers_check.inc, symbol: "CG's must be a row of lookup table FS"}
- {file: verifier/src/cp_verify.cpp, symbol: "so that key = row forces gh + 49152 = t for a row of the clamp's half of table FS"}
- {file: docs/model.md, symbol: '| 6 | the high part gh of a wide gate key | 7 |'}
- id: RR-21
category: range_rounding
scope: shipped
element: "range condition 8: |x[k]| < 2^39 for a stream value a norm reads"
relation: "follows from condition 1 on the same norm (docs/model.md section 9); the stream is held in residualStreamLimbs signed-byte limbs"
enforcement: proof_constraint
sites:
- {file: docs/model.md, symbol: 'which follows from condition 1'}
- {file: verifier/src/vf_norm_digit_rows.h, symbol: 'static int residualStreamLimbs = 4;'}
note: "no check of its own: it holds if RR-15 holds, an argument from the docs, not tested here"
- id: RR-R1
category: range_rounding
scope: research
element: "Qwen conv, L2/RMS norm and rotary ranges (qwen_ops): every input cell in [-2^F, 2^F) (F + 1 bits, offset 2^F), outputs in F + 4 bits (conv) or F + 3 bits (norm, rotary), floor remainders bit-decomposed. Not the nonlinear operator: its input interval is wider, see RR-R2"
relation: "c.range(cell, f+1, S) on each input cell of conv/norm/rotary; c.range() on outputs; Circuit::floor range-checks its remainder"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'auto a=c.cell(witness?in.x.at(i):0); c.range(a,f+1,S);'}
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'auto v=c.cell(witness?in.x.at(j):0);c.range(v,f+1,S);x.push_back(v);'}
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'for(int j=0;j<256;j++){auto v=c.cell(witness?in.x.at(j):0);c.range(v,f+1,S);'}
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'c.range(y,f+4,8*S)'}
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'c.range(y,f+3,4*S)'}
note: "narrowed after review of PR #97 (astra round 2): this row used to name all qwen_ops inputs as [-2^F, 2^F), which the nonlinear operator does not enforce; the row's claim for conv/norm/rotary is unchanged"
- id: RR-R5
category: range_rounding
scope: research
element: "qwen_qk_norm ranges: sigma and R, the isqrt residuals e1, e2, the division remainder rr and gap = sigma - 1 - rr, Z in F + D2 + 2 signed bits, the floor remainders rz (s bits) and ry (F bits)"
relation: "range(): bit columns b*b = b and an affine reconstruction; an out-of-range hostile value still gets bits, so only the reconstruction fails"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: 'auto e1c=c.cell(e1); c.range(e1c,e1,E_BITS);'}
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: 'auto gpc=c.cell(gap); c.range(gpc,gap,SIGMA_BITS);'}
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: 'auto rzc=c.cell(rz); c.range(rzc,rz,unsigned(p.s));'}
note: "no-wrap argument in research/operators/qwen_qk_norm/README.md"
- id: RR-R6
category: range_rounding
scope: research
element: "qwen_qk_norm statement admission: F = 12, D in {16, 256}, head count, inputs in [-2^F, 2^F), gains |g| < 2^(F+3), outputs |y| < 2^(F + D2 + 4)"
relation: "admit() on the public statement before the circuit is rebuilt; outside the proof's relation"
enforcement: verifier_validation
sites:
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: 'static void admit(const Stmt& st) {'}
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: 'for(auto g:st.gq) need(g>-GB&&g<GB,"q gain bound");'}
- id: RR-R2
category: range_rounding
scope: research
element: "Qwen nonlinear input range (qwen_ops): X in [-8*2^F, 8*2^F), i.e. x in [-8, 8) on the 1/256 table grid. There is no range call on X itself; the interval follows from q = floor(X / 2^(F-8)) with a bit-decomposed remainder in [0, 2^(F-8)), key = q + 2048, and the 4096-way one-hot selector forcing key in [0, 4096)"
relation: "Circuit::floor (remainder range + X = q*2^(F-8) + rem), key = sum({q,1}, 2048), selectors(): sum b_i = 1 and sum i*b_i = key"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'auto q=c.floor(x,f-8,"grid"); auto key=c.sum({{q,1}},2048);'}
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'range(rem,bits); eq({term(a),term(out,-(i64(1)<<bits)),term(rem,-1)});'}
- {file: research/operators/qwen_ops/proof.cpp, symbol: 'eq({term(sum(count))},1); eq({term(sum(index)),term(key,-1)});'}
- {file: research/operators/qwen_ops/README.md, symbol: 'Integer input X maps to i=floor(X/2^(F-8))+2048. Domain [-8,8), fail closed outside.'}
note: "split from RR-R1 after review of PR #97 (astra round 2): astra proved 5 of 5 complete nonlinear proofs at F = 12 with X = 8192 .. 8196, outside [-2^F, 2^F) but inside this interval; these are valid proofs of the actual relation, not a soundness break (reviewer's descriptive figure, https://github.com/wakesync/flashprove/pull/97#issuecomment-6784831796). The narrower interval is not enforced for nonlinear and is not claimed"
- id: RR-R3
category: range_rounding
scope: research
element: "qwen_conv_history ranges: every history, input and tap cell in [-2^F, 2^F) (F + 1 bits, offset 2^F), outputs in F + 4 bits, floor remainders bit-decomposed in F bits"
relation: "range(): bit columns b*b = b and an affine reconstruction; for an out-of-range witness the bits take the low bits of n, every product gate still holds and only the affine reconstruction fails"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'auto a=c.cell(wit?wit->at(i):0); c.range(a,f+1,S);'}
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'c.range(y,f+4,8*S); c.eq({c.term(y)},st.y.at(t));'}
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'Hostile witnesses may be out of range'}
note: "the honest-side range gadget does not refuse out-of-range witnesses (so hostile witnesses can be proved, README); enforcement is the in-proof reconstruction. Research security 10 bits per round, toy coins"
- id: RR-R4
category: range_rounding
scope: research
element: "qwen_conv_history statement admission: F in {12, 14, 16}, four outputs in [-8 * 2^F, 8 * 2^F), history and tail opening words in [-2^F, 2^F)"
relation: "admit() on the public statement before the root comparison and before the circuit is rebuilt; outside the proof's relation"
enforcement: verifier_validation
sites:
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'static void admit(const CStmt& st) {'}
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'need(v>=-S&&v<S,"history opening class");'}
note: "in verify_bound, admit() runs before the held-root comparison; the history and tail opening class checks apply only to the bound profile (the unbound entry checks F and the outputs only)"
# ---------------------------------------------------------------- committed inputs and state
- id: IS-01
category: input_state
scope: shipped
element: "configuration (policy.txt, KL.bin, layout.txt), anchor root and shape, OOD seeds and answers, hashed into the commitment ID and compared with the caller's -id"
relation: "blocks/frame/heads.h configuration_digest, published_id; cp_verify key stage"
enforcement: verifier_validation
sites:
- {file: blocks/frame/heads.h, symbol: 'inline Digest configuration_digest('}
- {file: blocks/frame/heads.h, symbol: 'inline Digest published_id('}
- {file: verifier/src/cp_verify.cpp, symbol: 'refuse("model-statement");'}
note: "negative test N1: a complete honest proof of another commitment is rejected here only because the caller passes its own -id; substitution suite (experiments/20261111-6-substitution-suite, PR #108): another example's config with its own -policy (C-CFG1) and layout.txt with a 30-byte comment line appended, under the original ID (C-CFG3b), rejected here, NO: model-statement, 5 of 5 each"
- id: IS-02
category: input_state
scope: shipped
element: "which model the committed weights are (publisher, revision, signature)"
relation: "none; the ID covers integers and configuration only"
enforcement: not_enforced
sites:
- {file: blocks/frame/heads.h, symbol: 'The ID does not say whose weights the table contains.'}
- id: IS-03
category: input_state
scope: shipped
element: "anchor table well formed (close to a codeword, entries in range) by the key file's one-time test"
relation: "blk::check_key"
enforcement: proof_constraint
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'blk::check_key(*confH, file, tested, why, nullptr, &place);'}
note: "substitution suite (experiments/20261111-6-substitution-suite, PR #108): layout.txt with a 30-byte comment line appended, under its own recomputed ID (C-CFG3), and a verifier build with changed one-time/proof chain labels (C-DOM2) end NO: key-refused, 5 of 5 each"
- id: IS-04
category: input_state
scope: shipped
element: "leaf table encodes the same weights as the anchor (link proof)"
relation: "blk::check_entry_if_tested"
enforcement: proof_constraint
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'blk::check_entry_if_tested(*confH, first, file, why, nullptr, &place);'}
- id: IS-05
category: input_state
scope: shipped
element: "request tokens and range r, e in the first chain's first query"
relation: "bind_request_statement; PR #41 measured 20 of 20 first-challenge changes"
enforcement: proof_constraint
sites:
- {file: blocks/frame/check.h, symbol: 'inline void bind_request_statement('}
note: "substitution suite (experiments/20261111-6-substitution-suite, PR #108): complete honest proof under another example's request, one prompt token changed, two prompt tokens swapped, r moved 60 to 59, another reply token, prompt token 0 changed (C-REQ1..4, C-RPL1, C-PFX1) and another example's proof under this request (C-REQ5) all end NO: proof-refused, 5 of 5 each; a verifier build with a changed first-chain label also ends NO: proof-refused (C-DOM1)"
- id: IS-06
category: input_state
scope: shipped
element: "every token id (and y) below the vocabulary size V"
relation: "token-range"
enforcement: verifier_validation
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'if (so.y >= Vv) refuse("token-range");'}
- id: IS-07
category: input_state
scope: shipped
element: "request form: exactly T + 2 words, r < e <= T, y = 0 unless e = T"
relation: "statement-range"
enforcement: verifier_validation
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'if (so.r >= so.e || so.e > T || (so.e < T && so.y != 0)) refuse("statement-range");'}
- id: IS-08
category: input_state
scope: shipped
element: "acceptance policy (exact match or clipped gap with C, theta, cap, s, sh) matches what the caller expects"
relation: "cpPolicyHeld compares -policy (else exact match) with policy.txt"
enforcement: verifier_validation
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'static void cpPolicyHeld(const PolicyNamed &p, const string &pol, const vector<u32> &request) {'}
- {file: verifier/src/vf_common.h, symbol: 'static constexpr bool acceptancePolicyIsExactMatch('}
note: "negative test N2: a complete honest clipped-gap proof under its own ID is rejected for a caller that expects exact match; substitution suite (experiments/20261111-6-substitution-suite, PR #108): another example's config without -policy ends NO: policy (C-CFG1b), 5 of 5"
- id: IS-09
category: input_state
scope: shipped
element: "the request's seal, slot and batch tree (no splicing of parts of different runs)"
relation: "lin_binding_root; R7.18, R14.7"
enforcement: proof_constraint
sites:
- {file: blocks/frame/check.h, symbol: 'lin_binding_root(file, first_ch.get(x.binding)'}
- id: IS-23
category: input_state
scope: shipped
element: "stream links between segments: the stream that leaves layer group g - 1 (xbO) is the one that enters group g (xbI), and the last one (xbL) feeds the final norm"
relation: "xbI from group 1 on and xbL are not committed by their segment: they are the earlier segment's xbO array, kept by the frame; the reader's placement record must equal the owner's"
enforcement: proof_constraint
sites:
- {file: docs/format/11-2-feed-forward-segment.md, symbol: 'From group 1 on, xbI is not committed by this segment'}
- {file: docs/format/11-4-output-head.md, symbol: 'xbL is not committed by this segment'}
- {file: verifier/src/vf_layers_plan.inc, symbol: 'samePlaceF(hidPlans2[g - 1].tl.ownEntry("stream"), P.tl.plFor[0])'}
note: "structural (one committed array read by two segments), by reading; IS-04 is the leaf/anchor link, not this (added after review of PR #97)"
- id: IS-10
category: input_state
scope: shipped
element: "Fiat-Shamir challenges recomputed from the chain, proof of work on planned rounds"
relation: "every read and draw compared with the next planned round"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_transcript_channel.h, symbol: 'static Ex chainAnswer(blk::VerifierRounds &ch, RoundName round) {'}
- id: IS-11
category: input_state
scope: shipped
element: "leaf-table column counter (opened columns stay within the query bound)"
relation: "cp_prove adds the batch's columns to the counter on disk before opening"
enforcement: prover_only
sites:
- {file: docs/protocol.md, symbol: 'So every column is counted before it is opened'}
- id: IS-12
category: input_state
scope: shipped
element: "masks present in a proof or public file (hiding)"
relation: "none on the verifier side"
enforcement: not_enforced
sites:
- {file: docs/protocol.md, symbol: 'because the verifier does not check for masks'}
- id: IS-13
category: input_state
scope: shipped
element: "tokenizer and chat template"
relation: "none; the proof binds token ids only"
enforcement: not_enforced
sites:
- {file: docs/protocol.md, symbol: 'The verifier knows nothing about end tokens or the tokenizer'}
note: "substitution suite (experiments/20261111-6-substitution-suite, PR #108): bare cp_verify on the exact example takes no tokenizer and says YES (C-TOK5, 5 of 5, registered expected accept, gap G-04)"
- id: IS-14
category: input_state
scope: shipped
element: "who chose the prompt tokens (system prompt, template tokens, earlier turns)"
relation: "none; the request is whatever token ids the request file holds"
enforcement: not_enforced
sites:
- {file: research/security/threat-model/CODE-CHECKS.md, symbol: 'The proof binds the window'}
- id: IS-15
category: input_state
scope: shipped
element: "sampling parameters (temperature, top-p, seed)"
relation: "none"
enforcement: not_enforced
sites:
- {file: docs/protocol.md, symbol: 'Sampling is not checked under any policy'}
- id: IS-16
category: input_state
scope: shipped
element: "freshness, client identity, replay of a proof"
relation: "none"
enforcement: not_enforced
sites:
- {file: docs/limits.md, symbol: 'A proof does not name a client or a time'}
note: "substitution suite (experiments/20261111-6-substitution-suite, PR #108): the unaltered proof replayed with the unaltered request is accepted again (C-RPL2, 5 of 5, registered expected accept, gap G-05)"
- id: IS-17
category: input_state
scope: shipped
element: "identity of the verifier binary and its built-in constants"
relation: "none; a manual format version stands in"
enforcement: not_enforced
sites:
- {file: research/security/fiat-shamir-attacks.md, symbol: 'not cryptographically identified'}
note: "substitution suite (experiments/20261111-6-substitution-suite, PR #108): verifier builds with changed chain-label constants reject the unaltered example (C-DOM1 proof-refused, C-DOM2 key-refused, 5 of 5 each); this shows the labels enter the transcript, not that a client can identify its verifier binary"
- id: IS-18
category: input_state
scope: shipped
element: "the 32-byte file of cp_commit -public-weights (seed source)"
relation: "cp_commit accepts any 32-byte file; nothing shows that the option was used"
enforcement: prover_only
sites:
- {file: prover/common/public_weights.h, symbol: 'cp_commit accepts any 32-byte file.'}
- id: IS-19
category: input_state
scope: shipped
element: "what the scale file does to the model (a scale file within the admitted range can switch off blocks)"
relation: "KL.bin is bound by the ID and its ranges are checked; its meaning is not"
enforcement: not_enforced
sites:
- {file: guide/known-limits.md, symbol: 'a scale file within the admitted range can switch off'}
- id: IS-20
category: input_state
scope: research
element: "signed model manifest next to cp_verify (tools/manifest): upstream, files, tokenizer, template, system prompt"
relation: "client-side ed25519 manifest; 14 of 37 fields are the signer's word (PR #59)"
enforcement: verifier_validation
sites:
- {file: tools/manifest/manifest.py, symbol: 'def check_config(m, config_dir):'}
- {file: tools/manifest/README.md, symbol: "signer's word"}
note: "substitution suite (experiments/20261111-6-substitution-suite, PR #108): one-byte-changed tokenizer.json, another chat template, end tokens differing from the manifest, and a manifest re-signed by another key all REFUSED (C-TOK1, C-TOK2, C-TOK3, C-TOK6, 5 of 5 each); client-side only"
- id: IS-21
category: input_state
scope: research
element: "Qwen DeltaNet recurrent state at chunk boundaries (issue #34 harness)"
relation: "chunk proofs plus comparisons of SHA-256 of public words outside the proofs"
enforcement: verifier_validation
sites:
- {file: research/operators/qwen_state_commit/window.py, symbol: 'def chain_check(outdir, stmt, names, verify_chunk):'}
- {file: research/operators/qwen_state_commit/README.md, symbol: 'done in Python, outside the proofs, on public values'}
- id: IS-22
category: input_state
scope: research
element: "Qwen state-commit harness: replaced token inputs and a consistently changed conv history"
relation: "accepted by design of the harness (issue #34 RESULTS 'Still accepted after the fix')"
enforcement: not_enforced
sites:
- {file: experiments/20261109-34-qwen-state-commit/RESULTS.md, symbol: 'Still accepted after the fix, by design of this harness (open)'}
- id: IS-28
category: input_state
scope: research
element: "qwen_qk_norm: the committed input cells of each head equal the public pre-norm integers, and both operands of each square read the same cell"
relation: "one-term affine relation per input with the statement's value as target, plus a row-1 = row-0 tie"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_qk_norm/qk_norm.cpp, symbol: 'c.eq({{0,a,F{1}}},st.x[h][j]);'}
note: "binds the norm's input to the statement only; where the statement's inputs come from is OP-R18"
- id: IS-24
category: input_state
scope: research
element: "GLM KDA slices: where the public initial state comes from"
relation: "none; the initial state is a public value of the toy statement"
enforcement: not_enforced
sites:
- {file: research/operators/glm_kda_fixedpoint/README.md, symbol: 'Public initial-state provenance remains external.'}
- id: IS-25
category: input_state
scope: research
element: "qwen_conv_history incoming history opens the root the verifier holds (genesis or the root its previous accept returned); on accept the verifier returns the next root H(F, segment + 1, outgoing opening)"
relation: "tail_root = SHA-256(fp-conv-tail-v1, F, segment index, 3, three words); verify_bound compares it with the held root on public values, OUTSIDE the arithmetic proof (as PR #98's initial root)"
enforcement: verifier_validation
sites:
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'if(tail_root(st.f,st.segment,st.in_open)!=held) return std::nullopt;'}
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'return tail_root(st.f,st.segment+1,st.out_open);'}
- {file: research/operators/qwen_conv_history/README.md, symbol: 'This check is **outside** the arithmetic proof'}
note: "binds only if the verifier holds the right root; 5 of the 14 registered hostile sub-cases of issue #35 are rejected by this comparison, not by the proof (experiments/20261111-35-conv-history-binding/RESULTS.md, column 'rejected by'). Genesis is trusted (IS-27)"
- id: IS-26
category: input_state
scope: research
element: "qwen_conv_history in-proof pins: the three committed history cells equal the public incoming opening, the last three committed inputs X[1..3] equal the public outgoing opening; segment index and both roots are in the configuration text, so in the transcript"
relation: "six one-term affine relations with verifier-owned targets (bound profile only, 579 = 573 + 6); profile string history=bound-v1 segment= in-root= out-root= appended to the shape text"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'c.eq({c.term(x[j])},st.in_open[j]);'}
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'c.eq({c.term(x[4+j])},st.out_open[j]);'}
- {file: research/operators/qwen_conv_history/conv_history.cpp, symbol: 'p<<" history=bound-v1 segment="<<st.segment<<" in-root="'}
note: "registered suite (issue #35, seeds 1..10, release and UBSan): 4 complete malicious proofs per seed, every product gate holding and at least one pin violated, 0 of 80 accepted; sampled hostile classes, not exhaustive. One channel, T = 4, F = 12 only; not integrated with qwen_ops, qwen_handoff or cp_verify"
- id: IS-27
category: input_state
scope: research
element: "qwen_conv_history: correctness of the genesis root (what the first history is) and hiding of history and tail words"
relation: "none; the first root is whatever the verifier was given, and history and tail words are public openings (no in-circuit hash, no hidden commitment)"
enforcement: not_enforced
sites:
- {file: research/operators/qwen_conv_history/README.md, symbol: '**Genesis trust.**'}
- {file: research/operators/qwen_conv_history/README.md, symbol: '**No hiding.**'}
- id: IS-29
category: input_state
scope: research
element: "qwen_hidden_root window boundary: the first token's state cells, the incoming history and a private salt open the verifier-held incoming commitment, and the last token's state cells, the outgoing history and a second salt open the claimed outgoing commitment; no boundary state is public"
relation: "two in-circuit sponge evaluations (width 12, x^7, 8+22 rounds; 4 product columns per S-box, linear layers folded into affine terms) with 4 one-term pins each to the commitment words in the statement; public initial and final state pins of the lowering are removed; capacity is SHA-256 of the context (F, D, C, N, request, layer, policy, boundary index, model and weights labels), computed by the verifier"
enforcement: proof_constraint
sites:
- {file: research/operators/qwen_hidden_root/window.cpp, symbol: 'for (int i = 0; i < 4; i++) { cc.pin(ci[i], s.cin[i]); w.in_hash[i] = ci[i].v; }'}
- {file: research/operators/qwen_hidden_root/window.cpp, symbol: 'for (int i = 0; i < 4; i++) { cc.pin(co[i], s.cout[i]); w.out_hash[i] = co[i].v; }'}
- {file: research/operators/qwen_hidden_root/window.cpp, symbol: 'if (initial || final) continue;'}
note: "registered suite (issue #93, seeds 201..205, D8/F12/C2/N3): 9 complete hostile window proofs per seed and 3 spliced chains rejected by the proof verifier, 0 of 60 accepted; chain truncation is rejected by the verifier-held window count outside the arithmetic. Unvetted sponge, toy coins, not zero-knowledge (IS-30)"
- id: IS-30
category: input_state
scope: research
element: "qwen_hidden_root: collision and binding resistance of the sponge parameters, hiding of the state inside the proof, origin of the initial root R0, and chain length inside the arithmetic"
relation: "none; the sponge is a Poseidon-shaped stand-in without cryptanalysis, the proof system is not zero-knowledge, R0 is whatever the verifier was given, and the window count is checked by the chain verifier, not the proof"
enforcement: not_enforced
sites:
- {file: research/operators/qwen_hidden_root/README.md, symbol: 'The sponge parameters are unvetted'}
- {file: research/operators/qwen_hidden_root/README.md, symbol: 'Not zero-knowledge: "hidden" means the verifier is not handed the state'}
# ---------------------------------------------------------------- output positions
- id: OUT-01
category: output
scope: shipped
element: "checked positions r .. min(e, T - 1) - 1: under exact match each checked token is the top choice (gap 0, ties to the lowest id)"
relation: "margin digits HD0..HD2 at every real place; the head's policy lines"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_output_head_check.inc, symbol: 'the three digits of a margin D'}
- {file: verifier/src/cp_verify.cpp, symbol: 'a token that ties for the largest logit with a lower id has the gap 1, not 0'}
- id: OUT-02
category: output
scope: shipped
element: "clipped gap: mean of clipped costs at most theta, at most cap costs above C"
relation: "rhoC, phiC lines of the head; aboveClip bound"
enforcement: proof_constraint
sites:
- {file: verifier/src/vf_output_head_check.inc, symbol: 'F.lin(gapAt("rhoC", 0, R_)'}
- {file: verifier/src/vf_common.h, symbol: 'static constexpr u64 aboveClip('}
note: "cap tokens may be any tokens at any positions (docs/protocol.md section 2); WITHIN LIMITS is a weaker claim than YES"
- id: OUT-03
category: output
scope: shipped
element: "next token y after the window when e = T is the top choice of its position (under every policy)"
relation: "y enters the chain at the output head as the REPLY scalar; the head's margin at position T - 1 is 0"
enforcement: proof_constraint
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'y must be the argmax of its position'}
note: "split after review of PR #97: the comparison with the caller's y is OUT-09"
- id: OUT-09
category: output
scope: shipped
element: "the y in the proof is the y the caller names (the caller's y file)"
relation: "check_proof_if_tested compares the proof's unmasked reply with the caller's y; NO with 'the proof is about another reply'"
enforcement: verifier_validation
sites:
- {file: blocks/frame/check.h, symbol: 'why = "the proof is about another reply"'}
- {file: verifier/src/cp_verify.cpp, symbol: '"the proof is about another reply") ? "reply-differs"'}
note: "application validation on a value the caller states, as IS-01 and OUT-05: it binds only the y the caller brings"
- id: OUT-04
category: output
scope: shipped
element: "tokens 0 .. r (the prompt and any earlier turns)"
relation: "not checked by design"
enforcement: not_enforced
sites:
- {file: docs/limits.md, symbol: 'cover tokens 0 to r'}
- id: OUT-08
category: output
scope: shipped
element: "tokens e + 1 .. T - 1 after the checked range, when e < T"
relation: "not checked: only positions r .. e - 1 are marked (hOn); a truncated reply verifies as a shorter reply"
enforcement: not_enforced
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'for (u64 t = rowR; t < rowE; t++) {'}
- {file: docs/limits.md, symbol: 'A truncated reply verifies'}
note: "added after review of PR #97; OUT-05 (client range comparison) and OUT-06 (end token) are what stop a client from taking this as the whole reply"
- id: OUT-05
category: output
scope: shipped
element: "the checked range is the one the client asked for (no coverage truncation)"
relation: "cp_verify prints covered:/r= e= for the caller to compare; integrations/check_proof.py compares --range with the request file"
enforcement: verifier_validation
sites:
- {file: integrations/check_proof.py, symbol: "the request file's checked range is %d:%d, but you expect %d:%d"}
note: "negative test N3: cp_verify alone accepts the narrower proof (YES); only the client's range comparison rejects it"
- id: OUT-06
category: output
scope: shipped
element: "reply complete: the last checked token is an end token and no earlier reply token is one"
relation: "integrations/check_proof.py --end-token; cp_verify has no notion of end tokens"
enforcement: verifier_validation
sites:
- {file: integrations/check_proof.py, symbol: 'def check_end_token(self, allowed):'}
note: "only in the integration, and skipped with --end-token none; a bare cp_verify YES covers a truncated reply; substitution suite (experiments/20261111-6-substitution-suite, PR #108): check_proof.py with --end-token 5 on a reply ending in 199 rejected, 5 of 5 (C-TOK4)"
- id: OUT-07
category: output
scope: shipped
element: "which claim kind an accept states (exact-match, policy, next-token)"
relation: "result-v1 claim= field"
enforcement: verifier_validation
sites:
- {file: verifier/src/cp_verify.cpp, symbol: 'o.claim = r + 1 > last ? "next-token" : o.otherKind ? "policy" : "exact-match";'}
# Research directories under research/operators/ that no row covers, with the reason. The checker fails
# if a directory there is neither cited by a row nor listed here, so a new prototype cannot land unseen.
excluded:
- path: research/operators/evidence
reason: "receipts of the first top-k prototype (OP-R7), no relation of its own"
- path: research/operators/public_binding
reason: "public-coordinate product proof, a substrate of affine_rows (OP-R1); its own coordinate relation is not inventoried row by row"
- path: research/operators/shared_wires
reason: "shared-wire equality proof, a substrate of affine_rows (OP-R1) and the GLM slices; not inventoried row by row"
- path: research/operators/qwen_deltanet