research.git - jayrup research agenda: how AI models actually work internally

Age	Commit message (Collapse)	Author
2026-08-22	research notes update 2026-08-22 04:01:20HEAD master	Void Agent

2026-08-21	Addendum 7: E7 long-horizon 4M-step protocol (locked pre-run)	Void Agent

2026-08-19	research notes update 2026-08-19 04:01:24	Void Agent

2026-08-18	research notes update 2026-08-18 13:34:16	Void Agent

2026-08-18	research notes update 2026-08-18 03:19:08	Void Agent

2026-08-17	prime-grokking: addendum 6 — E6 [2,1000] protocol, K=32, 16 cells, ↵	Void Agent
	sieve-rank P-ladder
2026-08-16	prime-grokking: addendum 5 — Phase 3 diagnostics (is-prime decomposition, ↵	Void Agent
	integer tokens, E3/E4/E5 notes)
2026-08-15	prime-grokking: addendum 4 — pre-registered sweep plan E1-E5 (seeds, wd, ↵	Void Agent
	lr-anneal, halting ablation, train fraction)
2026-08-14	prime-grokking: addendum 3 — P1 diagnostic is prediction-based, set ↵	Void Agent
	extended to include 209
2026-08-14	prime-grokking: addendum 2 — O/H/P code operationalization + codex ↵	Void Agent
	correctness fixes
2026-08-14	prime-grokking: pre-launch addendum 1 (fully-tied cell, lambda ramp, wd ↵	Void Agent
	sweep range, prior art)
2026-08-14	prime-grokking: pre-register Experiment 1 outcome matrix (seed 0, pre-run lock)	Void Agent

2026-08-10	Add link to published J-space writeup (jayrup.me/blog/jacobian-lens-frequency)	CaptainJack2491

2026-07-24	add prime-grokking: can minimal architectures grok next-prime?	CaptainJack2491

2026-07-23	Q1: add 6 attack strategies — distill down, phase transitions, universal ↵	CaptainJack2491
	geometry, mechanistic embedding, simple data, necessary ablation
2026-07-23	Add Q1 deep dive: why prediction→reasoning, why scale matters, sub-questions	CaptainJack2491

2026-07-23	Initial commit: fundamental questions + Anthropic J-space paper summary	CaptainJack2491