3db6dd8fef
MATRIX.md: full item 1-31 results matrix (arcs, positives, walls, open threads); overview PDF v3 updated with items 29-31 finals
NilsandClaude Fable 52026-07-17 21:24:13 +02:00
7ac7f9651a
item 31 scored: KV-memory flat (39.5 vs 39.1, p=1.0); all gates unmoved at -10 — vanishing-gradient init confound recorded; -3 rerun is the loose thread
NilsandClaude Fable 52026-07-17 21:04:46 +02:00
96afac769e
item 30 scored: readiness strongly decodable (AUC 0.80) but NOT recurrence-specific (FF 0.79); early-stop loses everywhere, oracle only +3.9 — head is instrumentation, not a stopping lever; extension/deferral gating stays live
NilsandClaude Fable 52026-07-17 20:04:06 +02:00
64a20750bd
item 29 complete: trajectory TF = adapter training signal for hybrid regimes (39.1, p=0.045), not an internalization mechanism (job 2 answer-only 15.6 < 19.1 plateau; FR term harmful)
NilsandClaude Fable 52026-07-17 14:46:54 +02:00
11604d0949
item 30 pre-registered (Nils's design): carried state as metacognitive signal — answer-readiness head on L30 at line boundaries, carry-vs-FF AUC as primary, theta sweep + oracle stop via item-20 LUT methodology
NilsandClaude Fable 52026-07-17 12:57:05 +02:00
c7167d512f
overview PDF v2: add the prior arc (items 1-20) — retrofit discovery, band-location/anchor ablations, adapter-class factorial, transfer ladder, GSM boundary, gate program
NilsandClaude Fable 52026-07-17 12:36:38 +02:00
add87cdd72
item 29 arm 1: FIRST SIGNIFICANT POSITIVE — 39.1 (p=0.045 paired), every bucket a d=1 record; ablation 8-8 p=1.0 reattributes the gain: trajectory TF is a training signal for the adapter, not an inference loop
NilsandClaude Fable 52026-07-17 12:24:18 +02:00
62ce30e557
item 25 closed (truncated): writing is not computing — lce 1.9 with 2:12 flat at 31.2 (log-only; eval json never written due to kill); λ=1.0 arm cancelled; item 27 promoted
NilsandClaude Fable 52026-07-17 00:25:32 +02:00
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)
NilsandClaude Fable 52026-07-16 23:23:13 +02:00
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
NilsandClaude Fable 52026-07-16 23:03:56 +02:00
f31e6e7e30
item 24 closed (stopped by Nils): trainable band doesn't improve fit either; E2-L internalization line closed on all four axes; LoopLoRA k=0 bit-exactness validated in passing
NilsandClaude Fable 52026-07-16 22:55:10 +02:00
7a141948fa
item 23 scored: both control axes confirm — pp30 29.3 (within ±5 of 31.6, 3x pauses bought nothing), x600 no recovery; d=1 ceiling is structural, ladder chapter closed
NilsandClaude Fable 52026-07-16 20:08:54 +02:00
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
NilsandClaude Fable 52026-07-16 19:48:58 +02:00
f143c67b3d
probe_discount: single-prompt whiteboard microscopy (lens table + digit probs, carry vs FF control) on the 80/15% problem
NilsandClaude Fable 52026-07-16 19:14:23 +02:00
fb5fb13806
tied-alpha outcome: all predictions confirmed; free B causally isolated as fidelity culprit
NilsandClaude Fable 52026-07-15 04:13:11 +02:00
0a5cd80dc1
halting probe: output-stable orbit, not state fixed point; no free ACT at state level; paper claims softened
NilsandClaude Fable 52026-07-15 03:29:26 +02:00
8981cda31d
phase diagram: add per-depth point (fidelity kept, gain depth-stranded)
NilsandClaude Fable 52026-07-15 03:25:16 +02:00
0fd93cb328
n=500 controls: paired net-effect significant (loop vs untrained merge 17-4, p=0.007); ladder and net accounting finalized
NilsandClaude Fable 52026-07-14 20:08:57 +02:00
1d4cffd6ad
integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep
NilsandClaude Fable 52026-07-14 19:44:07 +02:00
caef95f10f
budget-CoT per-item: paired ties with latent arms confirmed (p>=0.69)
NilsandClaude Fable 52026-07-14 17:29:56 +02:00
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
NilsandClaude Fable 52026-07-14 17:16:00 +02:00
e5fc453005
pin gemma-4 model revisions (upstream 12B template regression: thought channel -> empty generations)
NilsandClaude Fable 52026-07-14 17:13:35 +02:00
378bb36b8c
node_setup: retry+timeout around model downloads (xet mid-transfer stalls)
NilsandClaude Fable 52026-07-14 16:56:17 +02:00
7821f304f8
LiveCodeBench transfer harness: raw-jsonl loader (script datasets dead), stdin judge, newest-window selection, plan labeling
NilsandClaude Fable 52026-07-14 16:48:12 +02:00
dcd0b0b2dc
HumanEval transfer: trained-vs-untrained-merge paired test (p=0.80, n.s.) -- transfer gain is the merge itself, not trained content
NilsandClaude Fable 52026-07-14 16:30:36 +02:00
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
NilsandClaude Fable 52026-07-14 16:05:40 +02:00
01fd272028
panel: chat-template MC scoring (raw scoring is chance on -it model), parquet dataset mirrors, incremental dump
NilsandClaude Fable 52026-07-14 16:00:17 +02:00
NilsandClaude Fable 5
