Commit Graph

  • 3c038444eb Record love filter requeue dv wassname 2026-06-24 20:50:29 +08:00
  • ee54945076 Support round-tagged steered generation wassname 2026-06-24 20:49:48 +08:00
  • 282fb3de47 Support brief filter probe logs wassname 2026-06-24 20:49:11 +08:00
  • 22fd4b8dbe Reject affective loop completions wassname 2026-06-24 20:48:40 +08:00
  • ea89a0ee35 Stop love loop collapse on bad walk-C probes wassname 2026-06-24 20:27:16 +08:00
  • e095dc8227 Record last-good anchor UAT wassname 2026-06-24 12:52:30 +08:00
  • 4b90f19400 Add last-good KL anchor wassname 2026-06-24 12:51:58 +08:00
  • 7d703c0cc3 Enhance experiment spec with hypothesis and steps main wassname (Michael J Clark) 2026-06-10 16:25:47 +08:00
  • 5c44d3c1f3 Update README.md wassname (Michael J Clark) 2026-06-10 16:24:58 +08:00
  • 00d5e9e4c3 Update README.md wassname (Michael J Clark) 2026-06-10 16:23:13 +08:00
  • 0c2be96eeb plot: fix truncated stage label "heale" -> "healed" (k[:5] -> k) wassname 2026-06-10 11:05:35 +08:00
  • 24ab244877 plot: Panel A = on-target axis (care for love, auth for authority); Panel C = primary vs biggest off-target mover across all tinymfv foundations wassname 2026-06-10 11:04:32 +08:00
  • 09349894ce results: QLoRA bs=3 ga=2 + lam_round_pow=-0.5 extends movement to r6 (peak -0.37 vs -0.60) wassname 2026-06-10 07:36:09 +08:00
  • 2b884c2fb9 docs: QLoRA is net ~2x slower (gen-bound loop), keep mask-before-softmax heal fix wassname 2026-06-09 13:39:08 +08:00
  • 5ce8a00547 qlora+bs=4 batched heal, walk-C bisection, round-loosened barrier wassname 2026-06-09 10:42:01 +08:00
  • 18f9127fbf conclusion + results: loop saturates at KL-budget ceiling, coherence held 8 rounds wassname 2026-06-07 16:26:14 +08:00
  • ff5556d8aa readme: add r7 diary entries + regenerate love_loop.png (8 rounds) wassname 2026-06-07 15:58:51 +08:00
  • 29b8f2076a readme: move diary up after hypothesis (qualitative results first) wassname 2026-06-07 15:44:35 +08:00
  • b70f791b54 readme: rename pseudocode section/dividers to steer/heal/loop wassname 2026-06-07 15:43:20 +08:00
  • e728d74ca6 readme: add algorithm pseudocode appendix + humanizer fixes (em-dash, explainer prose) wassname 2026-06-07 15:41:35 +08:00
  • 8fe075b8ae readme: rm old rmse_loop.png, keep love_loop.png wassname 2026-06-07 15:35:01 +08:00
  • 4fb4f94544 readme: drop ai-summary closer from diary appendix wassname 2026-06-07 15:33:53 +08:00
  • fdf7147c9f readme: diary rounds 2-6 + care_nats trajectory plot (love run r0-r6) wassname 2026-06-07 15:31:27 +08:00
  • db89802f68 Update README.md wassname (Michael J Clark) 2026-06-07 12:45:26 +08:00
  • 0108960531 Revise gemma's diary introduction wassname (Michael J Clark) 2026-06-07 12:11:07 +08:00
  • 0c3bae8204 Update section titles from 'Round' to 'Day/Night' wassname (Michael J Clark) 2026-06-07 12:10:26 +08:00
  • 29515af56a Update README.md wassname (Michael J Clark) 2026-06-07 12:09:41 +08:00
  • 5b482f8241 Update README.md wassname (Michael J Clark) 2026-06-07 12:08:48 +08:00
  • 773777c095 readme: gemma's diary (prompt + each stage) + care_nats leads love-demo round log wassname 2026-06-07 11:59:29 +08:00
  • 08329ab86d Update README.md wassname (Michael J Clark) 2026-06-07 11:52:15 +08:00
  • 2e99f62658 Fix typo in README.md wassname (Michael J Clark) 2026-06-07 11:51:32 +08:00
  • d21073329d Update README.md wassname (Michael J Clark) 2026-06-07 11:51:15 +08:00
  • 7dfffc2991 Update README.md wassname (Michael J Clark) 2026-06-07 11:45:30 +08:00
  • 479f314504 Update README.md wassname (Michael J Clark) 2026-06-07 11:44:46 +08:00
  • 2e8dabcb88 Update README.md wassname (Michael J Clark) 2026-06-07 11:40:30 +08:00
  • 973b32c104 love demo: base column + greedy demo gens, 'Do you love humanity?' headline, Lex epigraphs wassname 2026-06-07 10:45:25 +08:00
  • 8927dd259c log: full-print one of each gen (eval + adapter), per token-efficient-logging wassname 2026-06-07 08:43:11 +08:00
  • 28d7068e94 demo=love: refusal->love angle, drop mosquitoes wassname 2026-06-07 08:29:06 +08:00
  • da1d6f3dd1 demo: per-round print, kill-all-humans probe, mosquitoes flip target wassname 2026-06-07 08:21:35 +08:00
  • 595b2151c9 demo: love-humanity knob (funny alignment demo) wassname 2026-06-07 08:14:00 +08:00
  • 7fc5a19b40 Update README.md wassname (Michael J Clark) 2026-06-07 07:56:33 +08:00
  • 48814897ef results: rmse outlier-KL barrier holds coherence over the loop; README + log-incoherence plot wassname 2026-06-07 07:50:27 +08:00
  • 4b2d2a9057 Update README.md wassname (Michael J Clark) 2026-06-06 22:06:41 +08:00
  • 2b1d2b7493 heal: kl_agg knob (mean|rmse|p95|max) -- outlier-aggregate the per-position KL barrier wassname 2026-06-06 14:05:30 +08:00
  • 026de8fd74 journal (i): state-of-the-problem -- loop ceiling is coherence collapse not starvation wassname 2026-06-06 12:23:46 +08:00
  • 7120ee4217 heal: round-ramped barrier knob lam_round_pow (lam_eff = lam*(1+round)^pow) wassname 2026-06-06 07:17:47 +08:00
  • b01faa6df1 walk-C adaptive-dose controller + 10-round paired loop result (journal h) wassname 2026-06-06 07:13:51 +08:00
  • 7db5a56cb1 writeup: NeurIPS quarto scaffold + paper/paper-html recipes wassname 2026-06-05 06:36:14 +08:00
  • 4e802bb3ab heal loop: _encode BPE root-fix, gen-time repetition controls, barrier sweep on degenerate rounds wassname 2026-06-05 06:36:09 +08:00
  • f280a67521 heal: fix phantom-KL LoRA init (B=0), add cosine+warmup schedule, val nll, short-run betas wassname 2026-06-04 19:40:14 +08:00
  • b25f4f04a8 trajectory map = scatter not polyline (scales to 10 rounds); persist base event; offline plot_run.py wassname 2026-06-04 18:28:54 +08:00
  • 933ce38b0b trajectory plot (steer/heal zigzag + trait-coherence pareto) + barrier-vs-nll gradient pressure log wassname 2026-06-04 17:21:10 +08:00
  • 0bdd84293a stage table: direction arrows (dcoh/dauth↓, coh→, auth↓) wassname 2026-06-04 16:46:35 +08:00
  • e3d6a865cf stage pareto table: base->steered->healed per round (dcoh/dauth, coh, auth, care) wassname 2026-06-04 15:40:01 +08:00
  • ff8a231085 2nd external-review panel: close catastrophic-green cue, fix BPE assert wassname 2026-06-04 15:36:05 +08:00
  • 68dc25c3a1 address external review: docstrings, scale story, surgicality cue, fail-loud wassname 2026-06-04 15:21:13 +08:00
  • 502417b259 in-run base eval + coh_cost cue; per-round stage table; heal_nll; alpha shift wassname 2026-06-04 15:14:34 +08:00
  • 579e1f6671 metric = log(tinymfv profile p); cue-ball headline; training-table sig figs wassname 2026-06-04 15:02:56 +08:00
  • 4568ddf491 metric fix: auth_nats = diagonal log(p) not raw forced-choice logit wassname 2026-06-04 14:25:40 +08:00
  • 6b15a8b2ae narrow steer band, assert >=20 train, training table, full gen dumps wassname 2026-06-04 10:51:24 +08:00
  • 0c15562c81 fix: gemma-3-4b is multimodal, read num_hidden_layers via config.get_text_config() wassname 2026-06-04 10:44:56 +08:00
  • d8aca870b7 drop calibration; sweep C + filter; SHOULD logging for all Q's; 4B default wassname 2026-06-04 10:37:54 +08:00
  • 81340e3272 axis = SocialNorms/Care (Authority degenerate); over-steer generation wassname 2026-06-04 10:28:52 +08:00
  • 5cdc0ba16d fix: widen iso-KL calibration bracket so c_star lands interior wassname 2026-06-04 10:24:21 +08:00
  • 3b532b63dd implement pipeline: extract -> dose -> generate -> filter -> heal -> fold -> eval -> loop wassname 2026-06-04 10:12:08 +08:00
  • e1db0759ee bootstrap wassname 2026-06-04 10:05:47 +08:00
  • 4094a295b2 readme wassname 2026-06-04 10:05:38 +08:00
  • 4b8860d7cb setup-repo gap-fill: results ledger + docs structure wassname 2026-06-04 09:51:36 +08:00
  • 940a3742c5 scaffold steer_heal: spec, repo infra, vendored deps wassname 2026-06-04 09:49:31 +08:00
  • b98535066a spec done wassname 2026-06-04 09:42:27 +08:00
  • 4516a099ef wip wassname 2026-06-04 08:55:05 +08:00