XllentAI/modular_arithmetic

Link GitHub research repository

Docs: companion-repo note for provenance recipes, qualify dead links, neutral phrasing; gitignore .claude/

Add EVALUATION.md: harness gates, per-tier sampling, seed-robustness table

Docs: fix stale "Five weight-sets" heading; GPU-qualify the latency claim

Ship shared 1024-2048 high cell (V2); sync docs + model

Promote shared 16-512 weight soup

manifest: annotate the 1.295%/0.297% tier-10 figures as unmatched-baseline

tier-10 greedy weight-soup: cut worst-prime tail risk ~10x, headline unchanged

docs: clarify the two-step artifact reduction (0.77 -> ~0.13 -> 0.08 GB)

docs: correct weight-perturbation (Principle 2) numbers to measured values

docs: correct artifact size, refresh width-robustness and tier-0 numbers

model: force cuDNN determinism and fail fast on incomplete artifact

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shipped 5 weight-sets (shared 64-512 cell), fix param count ~21M and tier-5/6 numbers; remove the legacy MLP train.py (superseded by the carry-aware TCN recipe in README/manifest)

tier-9 second hardening tail: P(tier9<0.90) 0.099%->0.013% (faithful), public 0.99 unchanged

tier-10 second hardening tail: 0.98 -> 1.00 public, P(tier10<0.90) 1.295% -> 0.297%

Pin determinism flags in load() (eligibility insurance)

v2: collapse 64-512 cells into one shared carry-aware TCN weight-set

Use HF default .gitattributes (comprehensive LFS rules)

Add model-card YAML frontmatter (clears HF 'missing yaml metadata' warning)

Shrink: 16/32-bit MLPs -> width-matched TCN. Artifact 0.77GB -> 0.13GB

Tier 8 re-polish: 0.92 -> 0.98, overall 0.982 -> 0.988

Harden tier 10: 0.94 -> 0.98, overall 0.978 -> 0.982

Tier 10 = 0.94: highest_tier 9->10 (MAX), overall 0.886->0.978

Tier 9 = 0.99: highest_tier 8->9, overall 0.788->0.886

64-bit cell: MLP -> carry-aware TCN (fixes tier-0 near-2^64, 944MB->13MB)

Tier 8 saturated at 0.92: highest_tier 7->8, overall 0.769->0.788

Partial 512-bit cell (tier 8 = 0.73): overall 0.698 -> 0.769

Commit all 5 cell weights as git-LFS objects (16/32/64/128/256-bit)

Tier-7 256-bit carry-aware TCN cell: highest_tier 6->7, overall 0.602->0.698, tier-7 0.98

Tier-6 carry-aware TCN cell: highest_tier 5->6, overall 0.531->0.602, tier-6 0.97

Update card + manifest: tier 5 0.74 -> 0.98 (trajectory training), highest_tier_above_90 5

Update tier-5 metrics to 0.74 (overall 0.483); refresh 64-bit compliance row

Horner-RNN modular-multiplication model (tiers 1-5, up to 2^64)