Commit Graph
75 Commits
Author SHA1 Message Date
NilsandClaude Fable 5 d81afef488 items 16 + 15-closure scored: cross-task transfer toxic (task-locality confirmed 3-point); noise-s0 50%+ was a lucky seed
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 20:57:44 +02:00
NilsandClaude Fable 5 1e6473e1aa parcae seed 1: fidelity refutation replicates (easy ~71-73% both seeds)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 20:56:28 +02:00
NilsandClaude Fable 5 0f386ea29d fig_kcurves.py: restore original output filename
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 18:34:18 +02:00
NilsandClaude Fable 5 b6d54c2726 LESSONS: thermal hard-freeze pattern and diagnosis by exclusion
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 18:30:01 +02:00
NilsandClaude Fable 5 b17d7faea0 rename kcurves figure -> fig_kcurves2.png (cache-bust)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 18:00:07 +02:00
NilsandClaude Fable 5 873dc351a0 kcurves updated: noise-s0 and h2048 panels filled (both cross the distill line)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 17:56:04 +02:00
NilsandClaude Fable 5 cf9e38f858 h2048 outcome: per-depth exonerated (sharing >> time-variation at matched params); ceiling nudged upward
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 15:00:32 +02:00
NilsandClaude Fable 5 713fbdce76 randk + noise-s0 eval JSONs (landed pre-crash)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 14:15:26 +02:00
NilsandClaude Fable 5 f0e524cf68 noise-s0 exceeds prediction upward; seed arms queued and pre-registered
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 13:01:14 +02:00
NilsandClaude Fable 5 ebf0eb8f49 pre-register item 16: code->GSM8K cross-task transfer
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 12:54:58 +02:00
NilsandClaude Fable 5 b013ca27f1 kcurves: full design-space grid (10 panels, distill reference line, auto-fills pending arms)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 12:42:29 +02:00
NilsandClaude Fable 5 4455d4bb9d ladder figure: base vs FF vs loops vs distill (fig_loop_vs_ff)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 12:26:22 +02:00
NilsandClaude Fable 5 13643fa504 abstract rewritten: fold in regime sweep (two independent dials, B-tie isolation, ceiling survival, output-orbit)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 11:42:15 +02:00
NilsandClaude Fable 5 fb5fb13806 tied-alpha outcome: all predictions confirmed; free B causally isolated as fidelity culprit
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 04:13:11 +02:00
NilsandClaude Fable 5 0a5cd80dc1 halting probe: output-stable orbit, not state fixed point; no free ACT at state level; paper claims softened
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 03:29:26 +02:00
NilsandClaude Fable 5 8981cda31d phase diagram: add per-depth point (fidelity kept, gain depth-stranded)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 03:25:16 +02:00
NilsandClaude Fable 5 b53538e82c per-depth outcome scored (depth-stranded content, shared-tail rescue); kcurves updated with k=8
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 03:20:49 +02:00
NilsandClaude Fable 5 82cc1b85b8 depth-curve small multiples per regime (fig_kcurves)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 03:16:58 +02:00
NilsandClaude Fable 5 0e6d6adcb9 fidelity factorial arms (randk / noises0 / hidden-capacity) + pre-registration item 15
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 02:47:14 +02:00
NilsandClaude Fable 5 d39e02f4b9 parcae outcome scored: dynamics confirmed, fidelity refuted -- stability and fidelity are independent dials
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 02:38:51 +02:00
NilsandClaude Fable 5 16f4a057aa TiedAlphaAdapter: learned per-dim alpha with tied B (anchored by construction); pre-registration item 14
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 02:33:14 +02:00
NilsandClaude Fable 5 f370883e23 phase diagram: stability vs fidelity are independent dials (fig_phase)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 02:24:56 +02:00
NilsandClaude Fable 5 1076d57d96 rec arm outcome: rho->4.5 (norm-projected churn), prediction (a) confirmed through k=8; k16/32 cancelled
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 01:39:22 +02:00
NilsandClaude Fable 5 3477fb5da0 method: truncated-BPTT bias bound via rho(A) — contraction certifies tail-only gradients
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 01:08:58 +02:00
NilsandClaude Fable 5 d00f120f27 PerDepthAdapter (Bae-style per-iteration merges) + convergence-halting probe (free ACT); pre-registration item 13
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 01:07:01 +02:00
NilsandClaude Fable 5 f254b20b62 ParcaeAdapter: rho(A)<1 by construction (ZOH negative-diag), rho logging in rec arms, pre-registration item 12
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 00:53:34 +02:00
NilsandClaude Fable 5 d2da8044c3 RecurrentAdapter arm: Huginn-regime retrofit (learned A/B, noise h0, randomized depth) + pre-registration item 11
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 00:49:44 +02:00
NilsandClaude Fable 5 0fd93cb328 n=500 controls: paired net-effect significant (loop vs untrained merge 17-4, p=0.007); ladder and net accounting finalized
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 20:08:57 +02:00
NilsandClaude Fable 5 1d4cffd6ad integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 19:44:07 +02:00
NilsandClaude Fable 5 caef95f10f budget-CoT per-item: paired ties with latent arms confirmed (p>=0.69)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 17:29:56 +02:00
NilsandClaude Fable 5 2ce646d175 best-of-3 oracle/deployable split: realistic selector loses the overall edge (55.0/27.3 vs oracle 57.8/34.5); paper economics rewritten
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 17:16:00 +02:00
NilsandClaude Fable 5 e5fc453005 pin gemma-4 model revisions (upstream 12B template regression: thought channel -> empty generations)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 17:13:35 +02:00
NilsandClaude Fable 5 378bb36b8c node_setup: retry+timeout around model downloads (xet mid-transfer stalls)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 16:56:17 +02:00
NilsandClaude Fable 5 7821f304f8 LiveCodeBench transfer harness: raw-jsonl loader (script datasets dead), stdin judge, newest-window selection, plan labeling
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 16:48:12 +02:00
NilsandClaude Fable 5 dcd0b0b2dc HumanEval transfer: trained-vs-untrained-merge paired test (p=0.80, n.s.) -- transfer gain is the merge itself, not trained content
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 16:30:36 +02:00
NilsandClaude Fable 5 d1bafc2d4d revision per review: seed means in headlines, net accounting vs untrained merge, bucket definition up front, GSM12B n.s. (p=0.86), law->regularity, norm spec, compute appendix, public repo URL
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 16:05:40 +02:00
NilsandClaude Fable 5 01fd272028 panel: chat-template MC scoring (raw scoring is chance on -it model), parquet dataset mirrors, incremental dump
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 16:00:17 +02:00
NilsandClaude Fable 5 87308cfb0b score outcomes against pre-registered protocol items 1-10
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 15:48:36 +02:00
NilsandClaude Fable 5 55665ce251 CPU endgame: stats pass (Wilson/McNemar/pooled hard), final figures, PAPER.md rewrite around amortizable-content thesis
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 15:46:58 +02:00
NilsandClaude Fable 5 6a954bd3e4 MC capability panel: ARC/WinoGrande/HellaSwag/MMLU with prompt-loop
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 13:58:57 +02:00
NilsandClaude Fable 5 b461970c9a warm-start flag (stacking arm), adaptive flags for Blocksworld
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 13:02:46 +02:00
NilsandClaude Fable 5 aa3e11fb20 gpuq worker: tee job output to pane for live viewing
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 13:00:38 +02:00
NilsandClaude Fable 5 f2cbce5a03 related_work: log 12b_adaptive outcome as confirmation of the anchoring account
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:57:08 +02:00
NilsandClaude Fable 5 e152a91c05 parse_known_args: module-level parser must tolerate importers' argv
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:50:45 +02:00
NilsandClaude Fable 5 cd8e94dda5 gpuq: idle-time class dispatcher (single-writer, race-free) + docs
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:47:23 +02:00
NilsandClaude Fable 5 9329f22a63 gpuq: class-based scheduling at submit time (pending count + live util)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:41:34 +02:00
NilsandClaude Fable 5 c718c636a5 GPUQ: as-deployed cheat sheet + job conventions
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:40:39 +02:00
NilsandClaude Fable 5 d2d372c909 feedforward flag for HumanEval eval
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:34:57 +02:00
NilsandClaude Fable 5 12ec01becc adaptive flag for GSM eval
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:33:03 +02:00
NilsandClaude Fable 5 95be35d20f adaptive flag for unified trainer
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:32:36 +02:00
NilsandClaude Fable 5 32728e7b07 bootstrap via public https clone (repo now public)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 12:25:07 +02:00
NilsandClaude Fable 5 471cf336d0 gpuq: document the scheduler (design, contract, ops, failure modes, alternatives)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 11:54:38 +02:00
NilsandClaude Fable 5 0252641701 gpuq: per-node supervisor (worker resurrection + bucket heartbeat)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 11:50:10 +02:00
NilsandClaude Fable 5 5efe5b790a gpuq: bucket-backed per-GPU job queue (worker loop + submit/status helper)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 11:48:03 +02:00
NilsandClaude Fable 5 1460cf241f GSM plan-distillation control (width/depth law falsification test)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 11:22:46 +02:00
NilsandClaude Fable 5 8fe8e34b47 rung-2 pilot: loop-only entrance-faded band LoRA, warm-start, inline eval; BW hard-only fallback
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 11:00:42 +02:00
NilsandClaude Fable 5 507c170744 train_merge_bw: fall back to hard-only curriculum when easy pool is empty (BW direct pass ~1%)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:59:32 +02:00
NilsandClaude Fable 5 1ad019d76d add --lr flag (wide-adapter LR rescale diagnostics at 12B)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:46:46 +02:00
NilsandClaude Fable 5 53a8c9b609 adaptive-alpha merge (Lys-inspired, trained) + truncated-BPTT deep-k training
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:29:57 +02:00
NilsandClaude Fable 5 cd0ed80ea6 related_work: track the two arXiv PDFs (retrofitted recurrence, inner loop inference)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:28:34 +02:00
NilsandClaude Fable 5 c08c8d503d related_work: archive the 2026 workspace paper (HTML + 18 figures + bib) for offline reference
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:26:54 +02:00
NilsandClaude Fable 5 ba3bacc252 unification paragraph: anchor coefficient as the stability dial across Lys/ours/McLeish
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:25:04 +02:00
NilsandClaude Fable 5 8e7bd2019e integrate McLeish et al. + Lys et al.: reposition related work, rung-2 Muon note
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:24:17 +02:00
NilsandClaude Fable 5 7ac365271a related_work: unify anchoring across McLeish/Lys/ours; link 12B alpha miscalibration to their stability results
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:24:12 +02:00
NilsandClaude Fable 5 89df135631 related_work: relevance notes for 2511.07384 (retrofitted recurrence) and 2602.14759 (inner loop inference)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:22:31 +02:00
NilsandClaude Fable 5 f48011c470 decouple --pause from feedforward: enable loop+pause combo arm
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:09:18 +02:00
NilsandClaude Fable 5 5c4b76ba14 bootstrap via bucket git-bundle (no ssh keys on rented nodes)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:06:41 +02:00
NilsandClaude Fable 5 01ac15d3e1 node kit: alpha flag through train/eval, carry env+width fixes, bootstrap script
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 10:05:30 +02:00
NilsandClaude Fable 5 674b814b65 pre-register L9 discriminator arm
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 09:57:04 +02:00
NilsandClaude Fable 5 263a07eab9 three-paper split: plans, claim/evidence manifests, work plans
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 09:52:25 +02:00
NilsandClaude Fable 5 54f2b3c6cc MultiPL-E Rust harness: compile-run verifier, STaR labeling, transfer eval
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 09:45:46 +02:00
NilsandClaude Fable 5 6c39c38eb9 track protocol pre-registrations and interim report in git
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:54:52 +02:00
NilsandClaude Fable 5 42304aacf8 fix: load task data files only for selected tasks
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 01:28:12 +02:00
NilsandClaude Fable 5 69ce16ab53 vast.ai runbook + per-tokenizer eos handling in loop_common
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 01:14:57 +02:00
NilsandClaude Fable 5 ef9c08966c J-lens workspace reproduction + loop retrofit: lens, band looping, adapters, controls, multi-task evals
Reproduction of the 2026 workspace/J-lens paper on gemma-4 (E2B/12B/26B),
plus the workspace-loop retrofit line: merge adapter, prompt-only latent
planning (MBPP), carry variant, attribution controls (FF/pause/untrained),
band-location ablation, Blocksworld harness, 12B replication scripts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 00:54:12 +02:00