Commit Graph
73 Commits
Author SHA1 Message Date
NilsandClaude Fable 5 f1e6a628eb item 29 pre-registered (Nils's design): trajectory teacher-forcing — 10 waypoint transitions supervised independently (T[i-1]→T[i]), TF and TF+FR arms; job 1 d=1 step-span, job 2 answer-only full-CoT span
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-17 10:34:44 +02:00
NilsandClaude Fable 5 6542f80339 item 28 job: fix eval tags to ts50
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-17 00:46:57 +02:00
NilsandClaude Fable 5 bbf8a02c6b item 28 pre-registered (Nils's variant): teacher-state distillation — frozen full-cot teacher's band-exit state at step end, cosine into burst s^10; λ amended 1.0→5.0 pre-run (smoke: baseline cos-dist 0.113)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-17 00:46:19 +02:00
NilsandClaude Fable 5 44e814e78c item 27 pre-registered: zero-pause internal band looping — single M=10 in-place burst at prompt end, iteration-aligned lens targets, ii0 inference ablation; carry_steps gains inplace updates, trainer gains --inner-iters/--base-pauses
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-17 00:21:20 +02:00
NilsandClaude Fable 5 0cffce876b item 26 pre-registered: result-staging lens supervision at pre-'=' positions (no causal leakage — low loss requires computation); arms lg03 and lt03+lg03; item-25 in-flight note (lce 10.3→2.3)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 23:23:13 +02:00
NilsandClaude Fable 5 64abfe66f3 item 25 pre-registered: latent process supervision via differentiable lens readout — pause j trained to lens-encode deleted-step token j (λ=0.3/1.0 arms); carry_logits gains return_states
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 23:03:56 +02:00
NilsandClaude Fable 5 fe89820e64 item 24 pre-registered: d=1 retry with loop-only band-LoRA r=16 (4.8M params, whole band, k=0 bit-exact); trainer/eval gain --bandlora; smoke-tested train+eval
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 21:44:01 +02:00
NilsandClaude Fable 5 46b5e1b802 whiteboard microscopy on hard(795)/drop(430) specimens: FF slips at the 3-term sum (155), carry re-expands and defers (165 ✓); boards show live continuation-plan vs premature digit-commit at identical prefixes
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 19:48:58 +02:00
NilsandClaude Fable 5 f143c67b3d probe_discount: single-prompt whiteboard microscopy (lens table + digit probs, carry vs FF control) on the 80/15% problem
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 19:14:23 +02:00
NilsandClaude Fable 5 001a4ca5f3 item 23 pre-registered: d=1 ceiling controls (x600 undertrain arm, pp30 pause-bandwidth arm, +-5pt decision rule); trainer gains --tag-suffix
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 16:27:54 +02:00
NilsandClaude Fable 5 20f7014f5b item 22 pre-registered: E2-L rung B internalization ladder (front-first deletion, 10 pauses/step, warm-started d=1..3); trainer gains --drop-steps/--warm-start/--steps/--lr; smoke-tested 2 steps on Spark
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 10:13:01 +02:00
NilsandClaude Fable 5 272a7b1d1f 31B regimes parsed (early persist bump L7-16, motor only terminal — E2B signature absent, matches E4B); exp4_regimes.py takes explicit out path / defaults next to input jbar (fix pushed to node before 26B stage)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 09:54:52 +02:00
NilsandClaude Fable 5 34497b8835 item 21 McNemar scored: drop-bucket edge significant (2:6 p=0.0094, survives Bonferroni x4); overall A-vs-B ns; + session handoff, fig cleanup
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 09:49:21 +02:00
Nils 051e90e805 item 21: GSM carry-cot 57.4% (5x prior best); control isolates whiteboard to drop-bucket; E2-N/A2 planned 2026-07-16 09:40:23 +02:00
Nils 1618c206ae item 20 scored; item 21 (GSM carry-CoT + control) pre-registered and queued 2026-07-16 02:42:46 +02:00
NilsandClaude Fable 5 4b9e168838 item 20: gate threshold curve + oracle bound (probs pass, merge per-item LUT, offline sweep)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 02:03:12 +02:00
NilsandClaude Fable 5 74f04d124f node_bootstrap.sh v2: one-command node setup encoding every paid-for lesson (workspace-only, snapshot pinning, credential-aware sync, contract workers, graceful stop) + gpuq_presign for credential-less nodes
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 01:51:44 +02:00
NilsandClaude Fable 5 332f7f3622 E4B regime map (ckpt-salvaged, 150 prompts): no clear workspace signature — sensor L11-22, motor from L23, persistence bump absent; candidate band inside KV-shared zone
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 01:35:38 +02:00
NilsandClaude Fable 5 dec27a4017 gpuq worker: STOP sentinel for graceful drain (never kill mid-job again)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 01:21:00 +02:00
NilsandClaude Fable 5 550d5cbe56 gpuq data contract: declarative per-job in/out sync (pull inputs, push outputs every 120s + at exit), gpuq_sync.sh helper
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-16 01:17:33 +02:00
NilsandClaude Fable 5 684e793eef E1c: k*=0 routing + pre-loop halt target (item 19 scored, E1c pre-registered)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 23:51:44 +02:00
NilsandClaude Fable 5 f2a48ce338 E1b trainer: label-supervised halting head on frozen merge (item 19)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 23:43:37 +02:00
NilsandClaude Fable 5 357a5add06 E1 halting-gate machinery (adapter, trainer, gated eval) + pre-registration item 18
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 23:00:39 +02:00
NilsandClaude Fable 5 600f965247 regimes figure: sensor/workspace/motor shading restored (rule-derived, matches original hand-shaded bands)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 22:09:09 +02:00
NilsandClaude Fable 5 dc43769118 fig_regimes.py: reproducible cross-model depth-regimes figure (E2B+12B filled, 3 scans pending)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 22:05:29 +02:00
NilsandClaude Fable 5 3a0a6a96ae INDEX.md (task-oriented repo guide), REGIMES.json -> results/, phase diagram updated with full sweep (anchored family, seed means)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 21:52:26 +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 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 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 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 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 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 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 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 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 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