prime-grokking.git - Prime Grokking — can minimal architectures grok next-prime? Pre-registered negative-class experiment (weight-tied RNN vs transformer).

Age	Commit message (Collapse)	Author
7 days	eval: strip torch.compile _orig_mod./module. prefixes when loading ↵	Void Agent
	checkpoints (salvages compiled-GPU runs)
10 days	run_sweep: 2 threads for --device cpu jobs (matches Addendum 7 protocol)	Void Agent

10 days	E7 jobs: 4M-step long-horizon, mixed-device (rnn->cpu, ↵	Void Agent
	transformer->cuda+compile)
11 days	fix(train): enforce epoch permutation without replacement, add max cache OOM ↵	CaptainJack2491
	guard, and handle empty split tensors
11 days	perf(train): VRAM-resident tensor caching, zero-copy GPU batch sampling, and ↵	CaptainJack2491
	adaptive eval schedule
11 days	fix(eval): correct sieve-rank indexing + test expectations	Void Agent
	- _compute_sieve_rank: find LARGEST k (not smallest) whose signature contains all error preds — fixes P5(4) vs P5(1) bug - _sieve_rank_signature: iterate all small_primes (not just p²≤n) to catch composites like 209=11*19 where 11²>209 - Test expectations updated for [101,200] range (209 outside it) - is_prime task: P1→P3 (no sieve-rank for classification tasks) 48/48 green.
11 days	fix(eval): P5/P6 sieve-rank ladder replaces hardcoded FLAGGED_SIEVE_PREDS	Void Agent
	- probe_report: dynamic lo/hi (no more hardcoded [101,200]) - main(): probe range derived from range_end ([range_end+1, range_end+1000]) - P6 = exact (no probe misses), P5(k) = errors match rank-k sieve signature - _sieve_rank_signature: computes composites with all factors > p_k - _compute_sieve_rank: identifies rank from model's error predictions - build_report.py: probe_fig uses sieve rank for coloring - Tests updated for P5/P6 codes (48/48 green)
13 days	E6 Phase 5: complete 16/16 cells (no O1, all P4, no sieve rank)	Void Agent
	Safety-net run: rsync'd the remaining 8 cells from ichi (wd-1.0 seeds {1,2}, wd-3.0, batch-128), classified the full grid against Addendum 6. Verdict unchanged and now complete: no O1 anywhere, P4 out-of-range in all 16 cells, sieve-rank ladder resolves to 'no sieve' (smallest in-range composite predictions 1001/1003/1010/1079/1099, all small-factor). In-range only: low-wd RNN O-PARTIAL (seed-0), wd-1.0 replication seed-stable O4, batch-128 transformer destabilizes late (train 1.0->0.55, val 0.83->0.45). Tracks e6 summary.csv.
14 days	E6 Phase 5: classify 8/16 finished cells (no O1, all P4, no sieve rank)	Void Agent
	- Apply Addendum 6 P-ladder on [1001,2000] probe (computed on CPU; eval.py still hardcodes [101,200]) - wd 0.01/0.1/0.3: O-PARTIAL + P4 (val EM 0.77-0.84 in-range); wd 1.0: O4 + P4 - Sieve rank: undefined in all cells (smallest composite pred has small factors; no P5(k)/P6) - 960->967 in all cells is training-split memorization, not sieve evidence - Verdict: data pressure did not change outcome class; walls algorithmic, not data-bound - 4 cells running (wd1.0 seeds 1,2), 4 queued (wd3.0, batch-128) -> 08:30 safety-net cron
2026-08-17	README: blog part 1+2 links, part 3 slot	Void Agent

2026-08-17	track per-phase summary CSVs (force-added under runs/ ignore)	Void Agent

2026-08-17	public release prep: full preregistration mirror (addenda 1-6), public ↵	Void Agent
	README, track per-phase summary CSVs
2026-08-17	E6 jobs: 16 cells [2,1000] K=32 cuda AMP (14 primary + 2 batch-128 ↵	Void Agent
	ablation); Addendum 6 locked 61b002f
2026-08-17	perf(cuda): add AMP fp16 autocast with GradScaler and torch.compile ↵	CaptainJack2491
	reduce-overhead mode
2026-08-17	feat(gpu): add CUDA support, chunked eval, bisect prime search, and scaling ↵	CaptainJack2491
	to 100k (48 tests green)
2026-08-17	fix D2: integers-mode probe is N/A by construction (OOV probe inputs), ↵	Void Agent
	functional regression test; D2 scored (int tokens collapse task: O2, 0-13% vs digits 17-87%); NOTES + report Phase 4 updated; 42 tests
2026-08-16	Phase 3 diagnostics results + analysis + report phase 4	Void Agent

2026-08-16	phase3: task_mode (is_prime), train_frac, lr_decay flags + tests (41 green); ↵	Void Agent
	Addendum 5 locked; 11-job batch
2026-08-15	E2 wd sweep results + analysis + report phase 3	Void Agent

2026-08-15	E2 wd sweep: 8 jobs (0.01/0.1/0.3/3.0 x both models, seed 0; 1.0 = existing ↵	Void Agent
	control)
2026-08-15	E1 results + analysis + report phase 2	Void Agent

2026-08-15	speedup: batched eval (46x, sieve-stub verified), run_sweep orchestrator ↵	Void Agent
	(2-way parallel, idempotent, summaries), E1 jobs; 36 tests
2026-08-14	tidy: drop dead --ckpt arg, unused imports/vars; verified report pipeline ↵	Void Agent
	ad-hoc + 34 suite green
2026-08-14	report pipeline: build_report.py (self-contained HTML), eval --runs-dir, ↵	Void Agent
	plot fix, halting per-input steps
2026-08-14	eval: P1 diagnostic prediction-based (flagged set +209), classifier ↵	Void Agent
	regression suite (10 tests, stub ground truth); 34 tests green
2026-08-14	fix codex BLOCKERs: global fixed layout invariance, min_steps off-by-one, ↵	Void Agent
	integers EOS alias, rerun guard, eval prereg-literal codes + dual-checkpoint honesty; +7 regression tests
2026-08-14	rnn: fully-tied cell (read-in+compute+decode), lambda ramp, ACT regression ↵	Void Agent
	test; design: prior art + review
2026-08-14	implement Experiment 1: data pipeline, tied-RNN w/ ACT halting, transformer ↵	Void Agent
	baseline, train/eval/plot, 16 tests
2026-08-14	scaffold: README, LICENSE, design docs (spec + prereg from research repo ↵	Void Agent
	00c696d)