CC engine: orbit-compatible even-weight overlap tie-break (GH-162, opt-in; do not merge before PI says merge) by guluchen · Pull Request #180 · guluchen/DistQLDPC

guluchen · 2026-10-29T14:02:11Z

Closes nothing yet; issue #162, hub #15. Do not merge before the PI says merge.

The report below is the GH-162 REPORT. My tool setup blocked writing optimization/experiments/GH-162/REPORT.md; all raw data, scripts and tables it cites are committed under optimization/experiments/GH-162/.

GH-162: orbit-compatible even-weight overlap tie-break (`tiebreak`)

Baseline: main 6083e662. Its CC sources are identical to 346d1c68 (PR Make -cc-aut=on the CC default (GH-152; do not merge before PI says merge) #155 -cc-aut=on; PR Rename six tier5 TN stems with GH-152 CC lower bounds #163 is data only). The main arm runs frozen server-main-346d1c6 (sha256 6c1dfb3c…).
Candidate: 97c8629 (frozen server-gh162-97c8629, sha256 3ea51003…).
Server: yfclab2 cores 0-23 only (cores 4, 14 and 21 were left out while other users ran there), one process per pinned core, 4 GB cap. Nothing was timed on the Mac.
Credit: the first(g) tie-break idea is from Qubitserf 04b9349 (S. Cercelescu, MIT). It was re-derived from the mathematics; no Qubitserf code was copied. No external solver was used.

Summary

Item	Result
What	New `-cc-rules` token `tiebreak` (default off; acts with `overlap`). While seed r_i is searched, each even stabilizer row g is counted on g minus first_i(g), with limit \|g\|/2 - 1. first_i uses a per-seed order: r_i, then the earlier orbits' qubits, then descending label.
Soundness	Per-seed lex-min argument (MODIFICATIONS.md GH-162; `cc_half.h`). excl, pack, parity, leaf, mrv and overlap keep any fixed weight-w target reachable. Only the seed rule constrains the target, and the lex-least element in an order that starts with the seed is a target of that seed.
Default path	Byte-identical on tier0+1 `-v`, 8 configs x 15 codes: 118/120. The 2 exceptions are MaxSAT `-ub-pre` progress lines (time-sliced finder, as in GH-152). All `-ub-pre=off` and CC configs are identical, including `-cc-aut=off`.
Tier 0	PASS. `make test-cc`: new Part T 76 800 ladders, 0 bad; tiebreak driver configs in Parts B/H, 0 bad. Part T with 3 000 instances: 768 000 ladders, 0 bad. Mutants 17 and 18 caught. `test-cc-flags` 417/417. Answers 184/184. 108 witnesses GF(2)-checked.
Tier 1 / 2 / 3 / 4	0.959 / 0.928 / 0.992 / 0.927, max cell 1.000. PASS. Tier 1 is at the process floor with identical node counts.
Tier 5 (6 h)	BB_360 0.40x time. BB_756/864/1080 reach lb 30 in 0.37 / 0.36 / 0.24x of main's time. BB_1080_4_30to48: lb 32 at 19 378 s (main: 30). REPRODUCED by GH-183's Tier 5 runs (different branching order, same codebase): lb 32 at 7 127 s and 6 739 s. The 8 h relabelled-seed runs here stopped at lb 30. TN_496 0.75x. LP 0.98-1.03 (idle).

Scientific semantics changed: no (opt-in pruning rule with a proof). PI action required: yes.

1. Proposals

Selected: orbit-compatible tie-break.
Tie-break only without orbits. Loses BB/GB/LP.
Per-seed orders. The issue dismissed this, but that was wrong: the argument works with a different order for each seed, provided each order starts with its seed. The candidate implements per-seed orders, which contain proposal 1 as a special case.

2. Why it is sound (short)

Seeds r_1, r_2, … have blocks O_1, O_2, … (orbits or single qubits). O_1..O_{i-1} are forbidden while r_i runs.

(A) Let j be the smallest index for which some weight-w element contains r_j and avoids O_<j. Then every weight-w element avoids O_<j; otherwise an automorphism would move it to a smaller index.
(B) Fix any order pi_j that starts with r_j, and let e* be the pi_j-lex-least weight-w element. Then e* contains r_j and avoids O_<j, so it is a target of seed r_j.
(C) If |e∩g| = |g|/2 and first_j(g) ∉ e, then e*+g is in F, has weight w and is lex-smaller. Hence |e* ∩ (g∖first_j(g))| ≤ |g|/2 − 1.
Conclusion. The tiebreak forbids avoid e*, and the other rules keep any fixed target. So seed r_j reaches e*, or an earlier genuine element. Other seeds can only miss elements, never invent one.
Mutant 18. An order without the seed first breaks (B). Mutant 18 uses such an order and fails only on instances with orbit seeds.
[Mac agent] CC engine: fractional-charge / edge-cover node lower bounds #157. charge/ecover (PR CC engine: charge node lower bound as a default CC rule (GH-157) #161) are target-independent bounds over free qubits like pack, so they are compatible. The combination was not run.

3. Implementation

cc_half.{h,cc}:
- Per-row exempt qubit (kex_, kexbase_, kexfb_).
- ov_add_tb / ov_del_tb, used only with the token; the old hot loop is unchanged when it is off.
- tb_seed / tb_block_done / tb_set switch the exemptions between seeds while S is empty. The |g| = 2 forbids move with the exemption.
- Test-only mutants 17 (naive tie cut through the tiebreak path) and 18 (seed not first). Mutants 15 and 16 stay reserved for [Mac agent] CC engine: fractional-charge / edge-cover node lower bounds #157.
cc_driver.cc: the token, plus -v lines that appear only with the token.
distqldpc.cc: -cc-mutant accepts 17 and 18; help text.
Tests: Part T, tiebreak configs in Parts B and H, test_cc_flags section 7g.
MODIFICATIONS.md (proof), NOTICE (credit), README.

4. S0: nodes at the same level, and the choice of order

Settings: -ub-search=off, seed 0. Ratio vs main at the deepest level every arm refuted (pre/).

Code (L)	asc / blk	blkrev	desc (chosen)	rnd
BB_1080 (24)	0.494	0.443	0.306	0.460
BB_756 (24)	0.547	0.367	0.442	0.491
BB_864 (24)	0.520	0.436	0.450	0.528
BB_360 (22)	0.469	0.350	0.350	0.423
BB_288 (16)	0.731	0.629	0.629	0.647
BB_144_14_14 (13)	0.850	0.861	0.893	0.870
TN_496 (19)	0.777	0.774	0.749	0.780
TN_250 (14)	0.896	0.893	0.884	0.915
GB_144_12_12 (11)	0.968	0.938	0.946	0.945
LP_1428 / LP_714	~1.00	~1.00	0.99-1.00	~0.99

The geomean over the eight BB/GB/TN rows is 0.551 for desc and 0.561 for blkrev. desc is simpler, so it was chosen.
Time per node is 3-10 % higher on BB (extra forbids).
LP rows (weight 8) rarely tie at a minimum element, and the weight-9 TN codes have only odd rows (rule idle).

5. Tier 0

Identity. See the summary table (S0/identity_tier01.txt).
make test-cc, default (S0/test_cc_default.txt): PASS.
- Part A: 115 200 ladders. Part B: 3 800 runs. [Mac agent] CC engine: Tanner-graph automorphism discovery for orbit seeds and the XZ-dual map (open TN codes) #152 Part G: 115 200 ladders. Part H: 1 950 runs.
- Part T: 300 instances, 76 800 ladders over the 64 rule subsets with overlap+tiebreak x 4 relabelling seeds; brute-force oracle, 0 bad.
- In 49 Part T instances the cut removes a minimum element of F; 8 of them have orbits.
- Mutants 1-14 are still detected.
Part T with 1 000 / 3 000 instances: 256 000 / 768 000 ladders, 0 bad. The cut binds on 138 / 439 instances (24 / 86 with orbits).
Mutant 17 (naive tie cut, no first(g) exemption): wrong on 5 152/76 800, 17 088/256 000 and 53 576/768 000 ladders.
Mutant 18 (orbit-incompatible order, seed not first): wrong on 160/76 800, 1 088/256 000 and 3 272/768 000 ladders, all on instances with orbit seeds.
test-cc-flags: 417/417.
Answers: 23 tier0-3 codes x 8 tiebreak variants (seeds 0-5, dual off, -cc-aut=off, -no-symbreak, mrv, finder and basis off, reduced rule sets): 184/184 = stem.
Independent GF(2) witness check (runner/witness_check.py): 108 witnesses, 0 bad. A planted bad witness is rejected.

6. Tiers 1-4 (5 seeds; main and tb adjacent in the queue)

Tier	Code	main median s (range)	tb median s (range)	ratio	nodes tb/main
1	LP_340_56_8	0.024 (0.024-0.026)	0.023 (0.022-0.024)	0.958	1.000
1	LP_442_68_10	0.025 (0.024-0.027)	0.024 (0.023-0.026)	0.960	1.000
2	BB_144_14_14	1.063 (1.033-1.143)	0.992 (0.971-1.082)	0.933	0.888
2	BB_288_12_18	0.278 (0.276-0.289)	0.205 (0.197-0.215)	0.737	0.643
2	GB_144_12_12	0.084 (0.074-0.085)	0.083 (0.074-0.084)	0.988	0.936
2	LP_544_80_12	0.038 (0.037-0.046)	0.038 (0.037-0.045)	1.000	0.994
2	TN_108_2_12	2.002 (0.841-2.765)	2.000 (0.838-2.748)	0.999	1.000
2	TN_200_10_10	0.164 (0.116-0.176)	0.154 (0.106-0.165)	0.939	0.959
3	LP_714_100_16	0.371 (0.265-0.419)	0.367 (0.271-0.414)	0.989	0.981
3	TN_144_2_13	5.536 (1.939-6.704)	5.507 (1.948-6.694)	0.995	1.000
4	TN_250_10_15	56.18 (31.69-58.19)	52.08 (29.05-54.51)	0.927	0.901

Geomean per tier: 0.959 / 0.928 / 0.992 / 0.927. No cell exceeds 1.25. All 110 answers equal the stems.
Tier 1's 0.96 is noise and binary layout, since the node counts are identical.

7. Tier 5 (6 h, default seed, main and tb concurrent; the 11 codes with even stabilizer rows)

Code	main	tb	time to main's final value (ratio)	largest level done by both: nodes, time
BB_360_12_24	d=24 @ 169 s	d=24 @ 67 s	0.399	L22 633.1M -> 221.4M (0.350), 0.390
BB_756_16_30to34	lb 30	lb 30	15928 -> 5893 s (0.370)	L28 49.07G -> 16.37G (0.334), 0.362
BB_864_4_30to40	lb 30	lb 30	11793 -> 4294 s (0.364)	L28 34.70G -> 11.89G (0.343), 0.356
BB_1080_4_30to48	lb 30	lb 32 @ 19378 s	13518 -> 3173 s (0.235)	L28 39.76G -> 8.51G (0.214), 0.227
LP_1020/1054/1122/1428/1768/2610	solved	solved	0.98-1.03	nodes 0.99-1.00
TN_496_2_23to30	lb 23	lb 23	12823 -> 9579 s (0.747)	L22 26.54G -> 18.73G (0.706), 0.738

The geomean over codes of the largest common level is 0.598 for nodes and 0.616 for time. 22/22 runs had 0 science problems.

New lower bound (PLAUSIBLE): BB_1080_4_30to48 has no logical of weight ≤ 31, so d ≥ 32 (X half under the verified dual map).

Cross-check (S1/xcheck_results.jsonl, S1/xcheck_logs.tar.gz): the same binary and rule with relabelling seeds 1, 2 and 3 (independent search orders), 28 800 s each, cores 1-3, run concurrently. Verdict of these three runs alone: not reproduced (level 30 unfinished). Overall verdict: REPRODUCED; see the [Mac agent] CC engine: MRV branching and BFS row order on top of tiebreak #183 check below.

seed	lb 28 at	lb 30 (level 28 EMPTY) at	level 28 nodes	level 30	final lb
0 (Tier 5 run)	519 s	3 173 s	8.51G	EMPTY at 19 378 s (16 205 s in the level)	32
1	831 s	5 662 s	14.90G	unfinished at 28 800 s (23 138 s in the level)	30
2	799 s	5 438 s	14.40G	unfinished at 28 800 s (23 362 s in the level)	30
3	840 s	5 685 s	14.99G	unfinished at 28 800 s (23 115 s in the level)	30

The relabelled orders need about 1.7x the level-28 nodes of seed 0. At that ratio, level 30 would need about 28 000 s, so it ends near 33 000-34 000 s, after the 8 h cap. No seed contradicts lb 32 (no level-30 FOUND, every finder upper bound >= 56). - Reproduction from [Mac agent] CC engine: MRV branching and BFS row order on top of tiebreak #183 (PR CC engine: MRV branching and seed-anchored reversed-BFS row order (GH-183, opt-in; do not merge before PI says merge) #204), same standard as [Mac agent] CC engine: fractional-charge / edge-cover node lower bounds #157: same codebase, different search order.

Runs: [Mac agent] CC engine: MRV branching and BFS row order on top of tiebreak #183 Tier 5, binary server-gh183-d26070d, -cc-rules=excl,pack,orbf,parity,leaf,overlap,tiebreak,mrv -cc-roworder=bfs, plus a second run with -cc-tbbase=asc. Logs: yfclab2:~/mac-agent-20261009/gh183-runs/tier5/logs/t5.BB_1080_4_30to48.tbmrvbfs{,a}.log.
Every level 2..30 ended EMPTY. Odd levels are skipped by the parity rule, and the X half is searched under the verified dual map.
There is no FOUND, INCONSISTENT, TRAP or error line in either log.
lb 32 was reached at 7 127 s and 6 739 s.

The node counts differ at every level from level 6 on, so these are different trees:

level	seed 0 here	[Mac agent] CC engine: MRV branching and BFS row order on top of tiebreak #183 mrv+bfs	[Mac agent] CC engine: MRV branching and BFS row order on top of tiebreak #183 mrv+bfs, base asc
6	45	38	38
10	1 243	966	1 044
20	7 272 401	3 605 794	3 508 140
26	1 437 184 756	555 988 543	529 925 377
28	8 512 752 562	3 032 679 237	2 875 368 469
30	50 869 528 776 (16 205 s)	16 674 839 512 (5 865 s)	15 728 534 392 (5 539 s)

Caveat: all four runs use the tiebreak rule and the same CC codebase. The reproduction is independent of the search order, but not of the rule's proof or implementation.
The stem has not been renamed (PI decision).

No other open code gained a bound in 6 h. BB_756/864 would need about 2x more time for level 30.

8. Recommendation and PI decisions

Recommendation: ADOPT CANDIDATE. Tiers 0-5 passed, with a clear node reduction on every code with tie-prone even rows (BB 2.9-4.7x at the deepest levels; TN_496/250 1.1-1.4x) and no loss elsewhere.
Decision 1: make tiebreak part of the default CC rules. Not done here, since it needs PI approval; a follow-up adoption PR would flip the default.
Decision 2: BB_1080 lb 32 is REPRODUCED ([Mac agent] CC engine: MRV branching and BFS row order on top of tiebreak #183 runs, different search order, same codebase). Whether to rename the stem BB_1080_4_30to48 to BB_1080_4_32to48 is the PI's decision.
Decision 3: merge this PR (opt-in) or keep it open.

Files

optimization/experiments/GH-162/:

runner/: queue with per-job env/bin (cores 0-23), job maker, identity, GF(2) witness check, panel/tier5 analysis.
pre/: order exploration.
S0/: Tier 0.
S1/: panel and Tier 5.

Server data: yfclab2:~/mac-agent-20261009/gh162-runs/.

🤖 Generated with Claude Code

…eak), opt-in New -cc-rules token 'tiebreak' (default off, acts with overlap): an even stabilizer row g is counted on g minus first(g) with limit |g|/2 - 1, first(g) taken in a per-seed order (the seed, then the earlier orbits, then descending label). Sound by the per-seed lex-min argument (MODIFICATIONS.md, GH-162). Test-only mutants 17 (naive tie cut) and 18 (seed not first). make test-cc Part T and tiebreak driver configurations in Parts B and H; test_cc_flags section 7g. The first(g) tie-break idea is from Qubitserf 04b9349 (S. Cercelescu, MIT); re-derived, no code copied. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…, runners); Part T progress env Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

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

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

…el 30 unfinished; lb 32 stays PLAUSIBLE Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Conflicts resolved by keeping both sides: - Rules: charge (default on) and tiebreak (opt-in) both kept; parse_rules, -cc-rules error text, help text, README, MODIFICATIONS file table. - cc_half.cc: tiebreak's earlier seed construction kept; GH-157 charge scratch added after the K rows. - -cc-mutant accepts 0..15, 17, 18 (16 stays reserved/rejected). - tests/test_cc_engine.cc: the GH-162 mask bit moves from 128 (now charge) to RM_TB = 256. Driver configs "tiebreak", "tiebreak-dual-off- seed1" and "aut-tiebreak" now use the new default (RM_DEF) + tiebreak; "tiebreak-mrv-seed2" adds charge; new "tiebreak-nocharge". Part T runs charge together with tiebreak on seeds 1 and 3. - tests/test_cc_flags.sh: default rules + tiebreak accept / -v rules line. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…unds PI decision 2026-10-31. Same matrices (git mv of Hx/Hz/Gx/Gz; only the Gx/Gz header comment changes). Upper bounds unchanged. TN_360_4_17to24 -> TN_360_4_23to24 (GH-157) TN_396_4_17to26 -> TN_396_4_23to26 (GH-157) TN_504_2_17to34 -> TN_504_2_24to34 (GH-157) TN_648_10_9to32 -> TN_648_10_22to32 (GH-157) BB_1080_4_30to48 -> BB_1080_4_32to48 (GH-162, reproduced by GH-183) BB_756_16_30to34 -> BB_756_16_32to34 (GH-183) BB_864_4_30to40 -> BB_864_4_32to40 (GH-183) TN_496_2_23to30 -> TN_496_2_27to30 (GH-181) Lower bounds: DistQLDPC's optional CC engine with opt-in rules that are not on main (charge, PR #161; tiebreak with mrv / -cc-roworder=bfs, PRs #180 / #204; sparse,sparseov,sparsemrv, PR #203), each reproduced with a different search order or binary of the same engine (not an independent second method). Updated: data/matrices/README.md (tier5 list, rename history, open items), NOTICE, README naming example, name columns of optimization/baselines/SERVER.md and main_server_reference.json. The upper-bound witnesses for these eight stems are copied under the new stems to optimization/experiments/GH-146/witnesses/renamed-2026-10-31/ (VERIFY.txt: 8/8 VALID). Historical experiment records are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…unds PI decision 2026-10-31. Same matrices (git mv of Hx/Hz/Gx/Gz; only the Gx/Gz header comment changes). Upper bounds unchanged. TN_360_4_17to24 -> TN_360_4_23to24 (GH-157) TN_396_4_17to26 -> TN_396_4_23to26 (GH-157) TN_504_2_17to34 -> TN_504_2_24to34 (GH-157) TN_648_10_9to32 -> TN_648_10_22to32 (GH-157) BB_1080_4_30to48 -> BB_1080_4_32to48 (GH-162, reproduced by GH-183) BB_756_16_30to34 -> BB_756_16_32to34 (GH-183) BB_864_4_30to40 -> BB_864_4_32to40 (GH-183) TN_496_2_23to30 -> TN_496_2_27to30 (GH-181) Lower bounds: DistQLDPC's optional CC engine with rules that were opt-in when the bounds were found (charge, default on main since PR #161; tiebreak with mrv / -cc-roworder=bfs, PRs #180 / #204, and sparse,sparseov,sparsemrv, PR #203, not on main), each reproduced with a different search order or binary of the same engine (not an independent second method). Updated: data/matrices/README.md (tier5 list, rename history, open items), NOTICE, README naming example, name columns of optimization/baselines/SERVER.md and main_server_reference.json. The upper-bound witnesses for these eight stems are copied under the new stems to optimization/experiments/GH-146/witnesses/renamed-2026-10-31/ (VERIFY.txt: 8/8 VALID). Historical experiment records are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Data/doc-only update from main; src/ and tests/ unchanged by it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

yfc and others added 4 commits October 29, 2026 15:56

GH-162: experiment records (order exploration, Tier 0, Tier 1-4 panel…

208213c

…, runners); Part T progress env Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

GH-162: Part T 3000-instance run (768 000 ladders, 0 bad; mutants 17/…

480eeca

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

GH-162: Tier 5 results (11 codes, main vs tiebreak, 6 h)

3475ff5

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

GH-162: BB_1080 cross-check (relabel seeds 1-3, 8 h): lb 30 each, lev…

07a12d2

…el 30 unfinished; lb 32 stays PLAUSIBLE Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This was referenced Oct 30, 2026

[Codex] Index pinned library branches without double-counting shared inputs #201

Open

CC engine: per-half probe between tiebreak and MRV + reversed-BFS row order (-cc-roworder=auto, GH-206; do not merge before PI says merge) #215

Merged

guluchen mentioned this pull request Oct 30, 2026

Rename eight tier5 stems with GH-157/GH-162/GH-183/GH-181 CC lower bounds #216

Merged

Merge main (39ba87e, stem renames #216) into GH-162 tiebreak

32dc9d4

Data/doc-only update from main; src/ and tests/ unchanged by it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

guluchen merged commit a516ef0 into main Oct 30, 2026
3 checks passed

Navigation Menu

Repository navigation

CC engine: orbit-compatible even-weight overlap tie-break (GH-162, opt-in; do not merge before PI says merge) - #180

CC engine: orbit-compatible even-weight overlap tie-break (GH-162, opt-in; do not merge before PI says merge)#180
guluchen merged 7 commits into
mainfrom
experiment/gh-162-overlap-tiebreak

guluchen commented Oct 29, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

guluchen commented Oct 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

GH-162: orbit-compatible even-weight overlap tie-break (tiebreak)

Summary

1. Proposals

2. Why it is sound (short)

3. Implementation

4. S0: nodes at the same level, and the choice of order

5. Tier 0

6. Tiers 1-4 (5 seeds; main and tb adjacent in the queue)

7. Tier 5 (6 h, default seed, main and tb concurrent; the 11 codes with even stabilizer rows)

8. Recommendation and PI decisions

Files

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

guluchen commented Oct 29, 2026 •

edited

Loading

GH-162: orbit-compatible even-weight overlap tie-break (`tiebreak`)