Commit History
-
Link GitHub research repository
etwk35ad1a4
-
Docs: companion-repo note for provenance recipes, qualify dead links, neutral phrasing; gitignore .claude/
etwkf704813 -
Add EVALUATION.md: harness gates, per-tier sampling, seed-robustness table
etwk93db0e9 -
Docs: fix stale "Five weight-sets" heading; GPU-qualify the latency claim
etwkThe body and table already describe the two shared weight-sets; this fixes the leftover section heading and notes the ~174s/300s figure is GPU timing (the rules' evaluation guidance assumes GPU batching). Model, weights, and manifest unchanged.8ed8c45
-
Ship shared 1024-2048 high cell (V2); sync docs + model
etwkReplace the two dedicated high cells (weights1024.pt, weights2048.pt) with one shared carry-aware-TCN cell across tiers 9-10 (weights_shared_1024_2048.pt): distilled to both dedicated teachers + worst-bit margin, 2048 chain preserved so the primary key (highest_tier_above_90=10) is held. tier 9 = 1.00, tier 10 = 1.00, overall_accuracy 1.000; two shared weight files total (~10.7M params, 0.04 GB). model.py: _build_cell drops non-constructor keys (unified/widths).41fc51b
-
Promote shared 16-512 weight soup
etwkb670993 -
manifest: annotate the 1.295%/0.297% tier-10 figures as unmatched-baseline
etwkThe matched 5-prime bootstrap measures harden2 at P(tier10<0.90)=0.191%, which the weight-soup cuts to 0.018%; clarify so the stale 0.297% isn't read as current. Weights unchanged.6d991cb -
tier-10 greedy weight-soup: cut worst-prime tail risk ~10x, headline unchanged
etwkweights2048 f9f72e98->c242a91e (soup_h2s2s3 = mean of harden2 + 2 independent margin tails, public-correct member pinned so public stays 1.00). Matched faithful bootstrap P(tier10<0.90) 0.191%->0.018%, worst near-max prime 0.833->0.875, robustness 0.977->0.983; fixed-seed end-to-end A/B confirms tier-10 1.00>=0.99, tiers 1-9 byte-identical. README + manifest updated; de-duplicated a doubled sentence. tier 10=1.00, overall 0.997, htop=10.124b03f
-
docs: clarify the two-step artifact reduction (0.77 -> ~0.13 -> 0.08 GB)
etwkThe 16/32-bit MLP->TCN shrink took the artifact to ~0.13 GB; the later mid-cell collapse to one shared weight-set brought the total to 0.08 GB. Avoids implying the small-cell change alone produced 0.08 GB.e6d48a0 -
docs: correct weight-perturbation (Principle 2) numbers to measured values
etwkRe-ran compliance_perturb.py against the shipped weights (the prior run could not load the shared 64-512 cell). The shared mid cell has WIDER bit margins than the retired per-width cells, so it tolerates small noise before collapsing (tiers 5-8 = 0.95/0.85/0.90/0.75 at sigma=0.25) rather than dropping to ~0.03; it still collapses by sigma=0.5-1.0 and an untrained re-init is at the floor (<=0.02 every tier). The precision-critical deep cells (9,10) collapse first (0.03/0.04 at sigma=0.25). Replaces the stale per-width-cell figures in README and manifest. Principle-2 collapse holds for every cell.fc0c8f4 -
docs: correct artifact size, refresh width-robustness and tier-0 numbers
etwk- Artifact size is ~0.08 GB (83.7 MB across five weight-sets), not 0.13 GB: the 0.13 GB figure predated collapsing the four mid-width cells into one shared weight-set. Fixed in README and manifest. - Rewrote the width-robustness audit note for the shared-cell model: every shipped cell is now trained width-matched (value-uniform + a bit-length- uniform band), the shared 64-512 cell shows no width knee, and a re-audit on 40k secret-style draws finds P(tier < 0.90) ~ 0.000%. The old note still described the retired per-width cells ('tiers 5-7 still degrade'). - Tier 0 (unranked near-max-prime diagnostic) is 0.71 on the current model, up from the stale 0.63.5459ca8 -
model: force cuDNN determinism and fail fast on incomplete artifact
etwk- Set torch.backends.cudnn.deterministic = True so any convolution kernel selected at load time is the deterministic variant, reinforcing the determinism guarantee already provided by eval-mode hard thresholding. - After loading, verify every routing width in CELL_WIDTHS is served by a weight-set and raise FileNotFoundError otherwise. A missing intermediate file would previously leave a routing gap, silently sending that width's primes to a wider, differently-trained cell rather than failing loudly. Public-benchmark score unchanged: overall_accuracy 0.997, highest_tier_above_90 = 10.33d0796 -
Docs: correct README to the shipped 5 weight-sets (shared 64-512 cell), fix param count ~21M and tier-5/6 numbers; remove the legacy MLP train.py (superseded by the carry-aware TCN recipe in README/manifest)
etwkea9333f -
tier-9 second hardening tail: P(tier9<0.90) 0.099%->0.013% (faithful), public 0.99 unchanged
etwkweights1024.pt: accum 32 + margin-m 8 worst-bit-margin tail from the shipped cell. Faithful 5-prime bootstrap (matched seed-991): E[tier9] 0.983->0.989, worst-prime 0.875->0.917, P(<0.90) 0.099%->0.013%, P(<0.95) 4.5%->1.4%. Public benchmark overall 0.997, htop 10, deterministic, tiers 1-8+10 byte-identical (only weights1024 changed). Tiebreaker/robustness gain.55de2ed -
tier-10 second hardening tail: 0.98 -> 1.00 public, P(tier10<0.90) 1.295% -> 0.297%
etwkweights2048.pt: warm-started from the shipped 0.98 cell, accum 32 + margin-m 8 worst-bit-margin tail (~75 min). Sharpens the worst near-max primes. Gated on the FAITHFUL 5-prime eval structure (real eval draws 5 primes/tier, so one weak prime is ~20% of the tier; vectorized bootstrap over a value-uniform pool). On the matched seed-991 hard pool: E[tier10] 0.971 -> 0.980, worst-prime 0.792 -> 0.833, P(tier10<0.90) 1.295% -> 0.297% (~4x risk cut). Public benchmark: tier 10 0.98 -> 1.00, overall 0.995 -> 0.997, htop 10, deterministic, tiers 1-9 byte-identical (no regression). Independent random-seed draw: 0.991/htop 10/det.c2635be
-
Pin determinism flags in load() (eligibility insurance)
etwkThe leaderboard ranks ONLY deterministic submissions (leaderboard/store.py). Pin the backend flags that govern run-to-run reproducibility instead of relying on host defaults: cudnn.benchmark=False (fixes the conv algorithm), and the TF32 flags to the exact validated mode (matmul TF32 off, cuDNN TF32 on). TF32 is deterministic, so this only makes the validated numerics host-independent; it does not change the determinism gate or accuracy. Validated: full public 1100 re-eval is byte-identical to shipped -- overall 0.995, highest_tier_above_90=10, deterministic=True, tiers 1/1/1/1/1.00/.98/1.00/1.00/.99/.98, inference within 300s. A TF32-vs-FP32 probe confirmed 0/100 cases differ on tiers 9 and 10 (the output is float-mode-invariant), so the near-max ~0.974 is intrinsic cell margin, not a precision artifact -- keeping TF32 (faster, identical) is correct.759f10f -
v2: collapse 64-512 cells into one shared carry-aware TCN weight-set
etwkReplace the four dedicated 64/128/256/512-bit cells with a SINGLE shared TCN weight-set (weights_shared_64_512.pt, ~5.5M params), run at each prime's native width. model.py: width-native forward (byte-identical for per-width cells, enables one weight-set at any width) + shared-cell loader keyed on config-declared widths; corrected stale docstrings (all cells are carry-aware TCN, not MLP). Full official pipeline (modchallenge evaluate --total 1100): overall_accuracy 0.995, highest_tier_above_90=10, deterministic=true; modchallenge check clean. Strictly better than v1 (0.989) with fewer weight-sets (5 vs 8). Principle-2 verified on the shared cell (1.00 -> 0.03 perturbed sigma=0.25, 0.00 untrained).fff63d1 -
Use HF default .gitattributes (comprehensive LFS rules)
etwkba38bb5 -
Add model-card YAML frontmatter (clears HF 'missing yaml metadata' warning)
etwklicense apache-2.0, library_name pytorch, descriptive tags + accuracy metric.84d4071
-
Shrink: 16/32-bit MLPs -> width-matched TCN. Artifact 0.77GB -> 0.13GB
etwkAll eight cells are now the same carry-aware TCN. weights16.pt 202MB->9.5MB, weights32.pt 456MB->12.6MB (both LFS). tier 4 0.99->1.00, overall 0.988->0.989, tiers 1-3 hold 1.00, no regression. Fixes the tier-4 width blind spot (audit). Card + manifest updated.9294ab9 -
Tier 8 re-polish: 0.92 -> 0.98, overall 0.982 -> 0.988
etwkWidth-matched re-polish of the 512 cell (value-uniform [257,512) + bitlen band [480,512], accum 16, lr 8e-5, worst-bit margin) closes the short-width gap and sharpens precision: real public tier-8 0.92->0.98, robustness sim [257,512] = 0.985. overall 0.988, every scored tier >=0.97, highest_tier 10, deterministic. weights512.pt updated (LFS).ffa1be7 -
Harden tier 10: 0.94 -> 0.98, overall 0.978 -> 0.982
etwkHardening tail on the 2048 cell (worst-bit margin, accum 24, lr 4e-5) lifts tier 10 0.94->0.98 (2047 27/27, 2048 71/73). overall_accuracy 0.982, highest_tier 10, deterministic. weights2048.pt updated (LFS), manifest + card updated.e258f44 -
Tier 10 = 0.94: highest_tier 9->10 (MAX), overall 0.886->0.978
etwkAdd 2048-bit carry-aware TCN cell (weights2048.pt, LFS) via octave transfer from the 1024 cell + two-stage polish. model.py CELL_WIDTHS += 2048; manifest + card updated. highest_tier_above_90 = 10 (the benchmark maximum), overall_accuracy 0.978, deterministic, inference 170s/1100, artifact 0.77 GB.6b83eb7 -
Tier 9 = 0.99: highest_tier 8->9, overall 0.788->0.886
etwkAdd 1024-bit carry-aware TCN cell (4.73M params, 12 blocks, dil 1..512). Fixes the prime-WIDTH blind spot: tier-9 primes are value-uniform in [2^513, 2^1024), so a ~1020-bit benchmark prime fell below the old training floor and scored 0/22 (tier 9 = 0.73). Retrained warm-started on a value-uniform + bit-length-uniform[990,1024] mix so the reduction boundary is learned at every MSB position -> tier 9 = 0.99 (private-draw sim 0.988). Compliance: 0.99 -> 0.04 at sigma=0.25, untrained 0.00. Full benchmark: overall_accuracy 0.886, highest_tier_above_90=9, deterministic, zero regression on tiers 0-8.bd6d487
-
64-bit cell: MLP -> carry-aware TCN (fixes tier-0 near-2^64, 944MB->13MB)
etwkweights64.pt replaced: 944MB MLP -> 13MB TCN (8 blocks/256ch/dilations 1-32). Fixes the MLP's near-2^64 blind spot (tier-0 64-bucket 0/10 -> 10/10) with no tier-5 regression (0.98 -> 0.99). Full benchmark: highest_tier 8, overall 0.789, tier 0 0.50, det=True. Total artifact ~1.62GB -> ~700MB. README/ manifest: 64/128/256/512 cells are now all carry-aware TCNs.cc41e6b -
Tier 8 saturated at 0.92: highest_tier 7->8, overall 0.769->0.788
etwkweights512.pt re-promoted to the 0.92 cell (gradient-accumulation run, eff. batch ~2300; eps floor 7e-4 -> 1.3e-4, 512-step chain 0.73 -> 0.92). Full benchmark: highest_tier_above_90 = 8, overall 0.788, tier 8 = 0.92, det=True. Compliance 0.92@s0 -> 0.04@s0.25, untrained 0.00. README/manifest updated: saturates tiers 1-8 / up to 2^512.70d0ebc -
Partial 512-bit cell (tier 8 = 0.73): overall 0.698 -> 0.769
etwkAdds weights512.pt (512-bit carry-aware TCN, 14 blocks/256ch/dilations 1-256, ~5.5M params) as a git-LFS object; routes p<2^512 to it (CELL_WIDTHS now 16/32/64/128/256/512). The per-step error floor rises with width so tier 8 lands at 0.73 (partial -- highest_tier stays 7) but overall jumps to 0.769. Compliance: 0.73@s0 -> 0.04@s0.25, untrained 0.00. README/manifest -> six cells / up to 2^512.86c5a17 -
Commit all 5 cell weights as git-LFS objects (16/32/64/128/256-bit)
etwkAll five cell weight files are tracked as proper LFS objects. md5s: weights16 96549e1a, weights32 3c3cf174, weights64 07851d3d (tier-5 0.98), weights128 0c087b3f (tier-6 TCN 0.97), weights256 cb1bf338 (tier-7 TCN 0.98). Repo is now complete.845fdfb -
Tier-7 256-bit carry-aware TCN cell: highest_tier 6->7, overall 0.602->0.698, tier-7 0.98
etwkRoutes p<2^256 to a new 256-bit TCN cell (12 blocks/256ch/dilations 1-128, ~4.7M params; CELL_WIDTHS now 16/32/64/128/256). Full public benchmark: tier 7 = 0.98, overall 0.698, highest_tier_above_90 = 7, deterministic. Compliance: 0.98@s0 -> 0.06@s0.25 -> 0.02@s0.5, untrained 0.00. README + manifest updated to five cells / up to 2^256.b69f4ce -
Tier-6 carry-aware TCN cell: highest_tier 5->6, overall 0.531->0.602, tier-6 0.97
etwk128-bit MLP cell (tier-6 0.26) replaced by TCNHornerCell -- weight-shared non-causal dilated convs over the 128 bit-positions (~3.9M params/16MB). Per-step eps ~15x below the MLP floor; benchmark tier-6 0.26->0.97, deterministic, inference 0.9s. weights128.pt md5 0c087b3f (git-LFS object). model.py adds TCNHornerCell + _build_cell arch dispatch; manifest + README updated; perturbation compliance incl. tier-6.47e319e
-
Update card + manifest: tier 5 0.74 -> 0.98 (trajectory training), highest_tier_above_90 5
etwkweights64.pt is the trajectory-trained 0.98 cell (md5 07851d3d..., git-LFS object). model.py/train.py unchanged.6b82250 -
Update tier-5 metrics to 0.74 (overall 0.483); refresh 64-bit compliance row
etwkfaa14cf -
Horner-RNN modular-multiplication model (tiers 1-5, up to 2^64)
etwkBit-sequential RNN that learns the Horner step (2t+bit*b) mod p; three cells (16/32/64-bit) routed by prime size. Public benchmark: tiers 1-3 = 1.00, tier 4 = 0.99, tier 5 = 0.64, overall 0.473. Weights weights{16,32,64}.pt are tracked as git-LFS objects.3d2c226