CC engine: erasure bound (GH-195, opt-in token erasure) by guluchen · Pull Request #219 · guluchen/DistQLDPC

guluchen · 2026-11-03T16:34:53Z

Issue #195 (hub #15). Opt-in CC rule erasure (-cc-rules=...,erasure, needs pack) plus -cc-erasure-start=N (default 16384). Default output is byte-identical with the token off.

Base: main 134d756 (merged; solver as of 02879a9). Measured binary 8667bc0 (solver sources identical to the head). Records: optimization/experiments/GH-195/REPORT.md (main 02879a9 results under main02879/, the earlier PR #180-based runs kept as the old-base records).

Rule

At a node S whose packing scan found disjoint odd rows (rows of H, or GH-181 redundant rows) with free candidate sets R_1..R_p, let R be their union and extra = (w - |S|) - p. If extra <= 2 (and |R| <= 128), the node is cut unless some y in R and z (at most extra free qubits outside R) give H(S + y + z) = 0 and G(S + y + z) != 0. Decided by peeling + dense elimination of the region with exact back-substitution (phi H_R = 0 with 64 random lanes, F H_R = G_R with full-width labels), then single/pair lookups (pair table built lazily). Tests start at the level after the first level of the half with >= N nodes; with -cc-roworder=auto the probe runs the other configuration with the current erasure state.

Soundness (MODIFICATIONS.md, GH-195): every other rule keeps a fixed weight-w target e* with S in e* and e* disjoint from the forbidden set (excl, orbf, overlap, tiebreak Lemmas A-C, xorunit forbids). Each packed row is odd in syn(S) and even in syn(e*), so it holds a free qubit of e* \ S in its own R_i; at most extra qubits of e* \ S lie outside R, so e* solves the relaxation. Fingerprint collisions only stop a cut; labels are exact; gates are cost-only.

Credit: the erasure bound idea is from Qubitserf 04b9349 (S. Cercelescu, MIT); clean-room implementation (NOTICE, MODIFICATIONS.md).

Tier 0

make test-cc PASS: new Part U, 259 200 ladders over rule subsets with pack,erasure (charge, bfs order and [Mac agent] CC engine: redundant sparse checks and XOR unit propagation #181 tokens mixed in), 0 bad; mutants 25 (bound weakened by one) 1370/9600 and 26 (fixed support ignored) 2204/9600 ladders wrong. Parts B/H erasure driver configurations (incl. -cc-roworder=auto) 0 bad.
make test-cc-flags PASS (566 checks).
Answers: 253 runs (23 tier0-3 codes x 11 erasure variants) all equal to the stem; 145 witnesses GF(2)-verified, 0 bad.
Byte identity with the token off vs main 02879a9 (default, and with tiebreak -cc-roworder=auto): identical (one MaxSAT ub-pre timing line, reproduced base-vs-base).

Tiers 1-4 (5 seeds, medians; yfclab2 CPUs 120-127)

Arms: def = main 02879a9 CC default; defer = +erasure; tbra = +tiebreak, -cc-roworder=auto; tbraer = tbra + erasure.
(a) def -> defer: PASS in all tiers. (b) tbra -> tbraer: Tiers 2-4 PASS; Tier 1 geomean 1.000 because the rule never reaches its start level on LP_340 / LP_442 (0 tests, identical nodes).

(a) main default -> +erasure

Tier	Code	def wall median (range) s	defer wall median (range) s	wall ratio	nodes ratio (median)	search-time ratio (median)	defer erasure calls / prunes (median)	answers
1	LP_340_56_8	0.022 (0.021-0.022)	0.021 (0.020-0.022)	0.955	1.000	0.996	0 / 0	8
1	LP_442_68_10	0.025 (0.024-0.033)	0.026 (0.024-0.033)	1.040	1.000	1.020	0 / 0	10
2	BB_144_14_14	0.748 (0.728-0.913)	0.255 (0.254-0.316)	0.341	0.085	0.320	30075 / 30068	14
2	BB_288_12_18	0.469 (0.449-0.489)	0.267 (0.256-0.278)	0.569	0.162	0.537	31943 / 31943	18
2	GB_144_12_12	0.085 (0.075-0.093)	0.053 (0.042-0.086)	0.624	0.268	0.448	2181 / 2181	12
2	LP_544_80_12	0.047 (0.045-0.058)	0.046 (0.043-0.059)	0.979	1.000	1.016	0 / 0	12
2	TN_108_2_12	2.048 (0.918-2.682)	0.749 (0.405-0.869)	0.366	0.071	0.353	130018 / 129744	12
2	TN_200_10_10	0.193 (0.145-0.214)	0.134 (0.104-0.134)	0.694	0.363	0.604	10464 / 10463	10
3	LP_714_100_16	0.462 (0.304-0.504)	0.343 (0.240-0.391)	0.742	0.065	0.732	34151 / 34137	16
3	TN_144_2_13	5.564 (1.900-7.133)	1.770 (0.870-2.398)	0.318	0.056	0.334	251582 / 251580	13
4	TN_250_10_15	65.453 (35.272-66.635)	28.286 (18.284-28.487)	0.432	0.047	0.432	4317073 / 4317069	15

Tier	cells	geomean wall ratio	max cell	verdict
1	2	0.996	1.040	PASS
2	6	0.558	0.979	PASS
3	2	0.486	0.742	PASS
4	1	0.432	0.432	PASS

(b) main default + tiebreak + roworder auto -> + erasure

Tier	Code	tbra wall median (range) s	tbraer wall median (range) s	wall ratio	nodes ratio (median)	search-time ratio (median)	tbraer erasure calls / prunes (median)	answers
1	LP_340_56_8	0.021 (0.021-0.022)	0.021 (0.021-0.022)	1.000	1.000	1.043	0 / 0	8
1	LP_442_68_10	0.033 (0.025-0.034)	0.033 (0.032-0.034)	1.000	1.000	1.067	0 / 0	10
2	BB_144_14_14	0.628 (0.618-0.772)	0.255 (0.244-0.296)	0.406	0.127	0.435	28492 / 28485	14
2	BB_288_12_18	0.348 (0.328-0.367)	0.208 (0.197-0.227)	0.598	0.184	0.553	24237 / 24237	18
2	GB_144_12_12	0.090 (0.084-0.093)	0.054 (0.053-0.084)	0.600	0.272	0.461	2097 / 2097	12
2	LP_544_80_12	0.046 (0.045-0.048)	0.045 (0.045-0.056)	0.978	1.000	0.979	0 / 0	12
2	TN_108_2_12	1.494 (1.059-1.555)	0.758 (0.417-0.879)	0.507	0.125	0.601	130018 / 129744	12
2	TN_200_10_10	0.195 (0.142-0.213)	0.135 (0.101-0.136)	0.692	0.367	0.612	10245 / 10244	10
3	LP_714_100_16	0.461 (0.311-0.512)	0.350 (0.251-0.391)	0.759	0.164	0.727	29869 / 29855	16
3	TN_144_2_13	3.370 (2.292-3.753)	1.599 (0.991-1.842)	0.474	0.084	0.458	167253 / 167244	13
4	TN_250_10_15	28.708 (21.378-30.514)	13.879 (10.434-16.071)	0.483	0.037	0.507	2061766 / 2061755	15

Tier	cells	geomean wall ratio	max cell	verdict
1	2	1.000	1.000	FAIL
2	6	0.607	0.978	PASS
3	2	0.600	0.759	PASS
4	1	0.483	0.483	PASS

220 runs, science problems: []
OVERALL FAIL

Tier 5 (6 h)

Solved codes 0.44-0.83x (no solved code slower: PASS). Unsolved: TN codes 0.61-0.67x to the baseline's final lb; BB_756 / BB_1080 1.03-1.04x. No new lower bound (TN_496 lb 25 and TN_360 lb 23 are at or below the stems).

def -> defer

Code	def	defer	time to the def final value s (ratio)	largest level done by both: nodes (ratio), level s (ratio)
BB_360_12_24	d=24 @ 190 s	d=24 @ 110 s	190 -> 110 (0.581)	22 (X -> X): 633.1M -> 24.5M (0.039), 170 -> 98 (0.576)
TN_180_2_17	d=17 @ 156 s	d=17 @ 69 s	156 -> 69 (0.442)	16 (X+Z -> X+Z): 188.6M -> 10.3M (0.054), 107 -> 49 (0.454)
TN_216_4_18	d=18 @ 470 s	d=18 @ 236 s	470 -> 236 (0.501)	17 (X+Z -> X+Z): 585.0M -> 34.7M (0.059), 330 -> 164 (0.499)
TN_252_2_21	d=21 @ 8568 s	d=21 @ 4514 s	8568 -> 4514 (0.527)	20 (X+Z -> X+Z): 9076.2M -> 497.8M (0.055), 6181 -> 3243 (0.525)
TN_360_4_23to24	TIMEOUT lb 22	TIMEOUT lb 23	7521 -> 5016 (0.667)	22 (X -> X+Z): 11413.2M -> 1518.3M (0.133), 8096 -> 11030 (1.362)

geomean time to the baseline final value 0.538 (5 codes)

tbra -> tbraer

Code	tbra	tbraer	time to the tbra final value s (ratio)	largest level done by both: nodes (ratio), level s (ratio)
BB_360_12_24	d=24 @ 32 s	d=24 @ 26 s	32 -> 26 (0.827)	22 (X -> X): 83.4M -> 5.3M (0.063), 27 -> 22 (0.834)
BB_756_16_32to34	TIMEOUT lb 32	TIMEOUT lb 32	12616 -> 12979 (1.029)	30 (X -> X): 25055.6M -> 1853.8M (0.074), 10478 -> 10760 (1.027)
BB_1080_4_32to48	TIMEOUT lb 32	TIMEOUT lb 32	8499 -> 8833 (1.039)	30 (X -> X): 16674.8M -> 1297.4M (0.078), 6986 -> 7270 (1.041)
TN_180_2_17	d=17 @ 151 s	d=17 @ 69 s	151 -> 69 (0.461)	16 (X+Z -> X+Z): 188.6M -> 10.3M (0.054), 108 -> 49 (0.450)
TN_216_4_18	d=18 @ 395 s	d=18 @ 210 s	395 -> 210 (0.532)	17 (X+Z -> X+Z): 437.0M -> 27.5M (0.063), 273 -> 145 (0.530)
TN_252_2_21	d=21 @ 8587 s	d=21 @ 4520 s	8587 -> 4520 (0.526)	20 (X+Z -> X+Z): 9076.2M -> 497.8M (0.055), 6193 -> 3237 (0.523)
TN_288_2_22to24	TIMEOUT lb 22	TIMEOUT lb 22	16183 -> 9795 (0.605)	21 (X+Z -> X+Z): 16726.9M -> 998.1M (0.060), 11354 -> 6918 (0.609)
TN_496_2_27to30	TIMEOUT lb 24	TIMEOUT lb 25	10481 -> 6942 (0.662)	23 (X -> X): 17317.6M -> 667.4M (0.039), 6912 -> 4512 (0.653)

geomean time to the baseline final value 0.680 (8 codes)

Bound events (lb value@seconds, from 18)

Code	def	defer	tbra	tbraer
BB_360_12_24	18@0 20@2 22@20 24@190	18@0 20@1 22@13 24@110	18@0 20@1 22@5 24@32	18@0 20@1 22@4 24@26
BB_756_16_32to34	-	-	18@0 20@1 22@3 24@12 26@65 28@366 30@2138 32@12616 32@21600	18@0 20@1 22@3 24@13 26@71 28@389 30@2219 32@12979 32@21600
BB_1080_4_32to48	-	-	18@0 20@1 22@3 24@10 26@49 28@270 30@1512 32@8499 32@21600	18@0 20@1 22@2 24@10 26@53 28@281 30@1564 32@8833 32@21600
TN_180_2_17
TN_216_4_18	18@470	18@236	18@395	18@210
TN_252_2_21	18@192 19@676 20@2387 21@8568	18@110 19@374 20@1271 21@4514	18@194 19@674 20@2394 21@8587	18@110 19@375 20@1283 21@4520
TN_288_2_22to24	-	-	18@127 19@420 20@1420 21@4829 22@16183 22@21600	18@83 19@263 20@863 21@2877 22@9795 22@21600
TN_360_4_23to24	18@69 19@216 20@695 21@2269 22@7521 22@21600	18@51 19@154 20@485 21@1548 22@5016 23@16046 23@21600	-	-
TN_496_2_27to30	-	-	18@17 19@48 20@138 21@403 22@1221 23@3569 24@10481 24@21600	18@13 19@37 20@101 21@287 22@852 23@2430 24@6942 25@20291 25@21600

science problems: []
TIER5 PASS (no solved code more than 1.2x slower)

🤖 Generated with Claude Code

…, opt-in -cc-rules token erasure (default off; needs pack): at a node whose disjoint odd-row packing leaves extra = (w - |S|) - p <= 2 added qubits uncharged, the node is cut when no completion S + y + z (y in the packed rows' free candidates, z at most extra further free qubits) is syndrome-free and detected. Decided by peeling + dense elimination of the region with exact back-substitution (phi H_R = 0, F H_R = G_R); fingerprint collisions only make the rule give up. Idea from Qubitserf 04b9349 (S. Cercelescu, MIT), re-derived for our search state; clean-room code. Test Part R (own RNG stream), mutants 22/23, flag checks; runner scripts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…erf 04b9349, S. Cercelescu, MIT) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ts by scan) Tier 1 panel: LP_442 paid ~10 ms for the eager pair table (wall 1.44x on a 25 ms run). Node counts unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…c-erasure-start=N, default 16384) A test costs microseconds; on the Tier 1 LP codes (6-30k nodes in total) it cost 1-3 ms on 25 ms runs. Deterministic (node counts), sound for any per-node choice. Driver tests use N = 0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…y; gate exploration; Tier 1-4 panel) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…rse rules, stem renames) into GH-195 The erasure rule is ported onto the new search: packed redundant rows (GH-181 sparse) contribute their free candidates to the region, the test runs after the charge check, mutants renumbered 25/26 (22-24 are GH-181's), test Part U (argv[8]) adds charge, bfs and GH-181 tokens, driver configurations with -cc-roworder=auto. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… current erasure state The probe compared A with erasure against B before B's own start level (B had no erasure tests yet), which steered BB_360 to the index order. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…entity; Tier 1-4 panel) and the old-base Tier 5 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…port Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…-195 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # MODIFICATIONS.md # README.md # src/cc/cc_half.h # src/core/distqldpc.cc # tests/test_cc_flags.sh

yfc and others added 11 commits October 30, 2026 03:29

GH-195: docs (MODIFICATIONS proof, NOTICE and README credit to Qubits…

e712e4f

…erf 04b9349, S. Cercelescu, MIT) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

GH-195: build the erasure pair table lazily (after n/2 pair-stage tes…

dfa26ed

…ts by scan) Tier 1 panel: LP_442 paid ~10 ms for the eager pair table (wall 1.44x on a 25 ms run). Node counts unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

GH-195: experiment records (Tier 0 tests, answers, witnesses, identit…

d616ec8

…y; gate exploration; Tier 1-4 panel) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

GH-195: MODIFICATIONS: erasure state in the -cc-roworder=auto probe

60a86f3

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

GH-195: records on main 02879a9 (Tier 0 tests, answers, witnesses, id…

731c52f

…entity; Tier 1-4 panel) and the old-base Tier 5 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

GH-195: Tier 5 on main 02879a9 (solved codes 0.44-0.83x, PASS) and re…

041ee00

…port Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Merge main (134d756: collapsed AMC library, no solver change) into GH…

2da1ffc

…-195 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

guluchen mentioned this pull request Nov 4, 2026

CC: erasure bound (opt-in token) #195

Open

guluchen merged commit 86d1d59 into main Nov 4, 2026
3 checks passed

Navigation Menu

Repository navigation

CC engine: erasure bound (GH-195, opt-in token erasure) - #219

CC engine: erasure bound (GH-195, opt-in token erasure)#219
guluchen merged 11 commits into
mainfrom
experiment/gh-195-cc-erasure

guluchen commented Nov 3, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

guluchen commented Nov 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Rule

Tier 0

Tiers 1-4 (5 seeds, medians; yfclab2 CPUs 120-127)

(a) main default -> +erasure

(b) main default + tiebreak + roworder auto -> + erasure

Tier 5 (6 h)

def -> defer

tbra -> tbraer

Bound events (lb value@seconds, from 18)

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

guluchen commented Nov 3, 2026 •

edited

Loading