NilsandClaude Fable 5 f1bd5daf62 MATRIX.md: item 32 row + final architecture verdict ...

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

2026-07-18 02:10:40 +02:00

..

eval_bestof3_pi.json

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

2026-07-14 17:16:00 +02:00

eval_bestof3_select.json

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

2026-07-14 17:16:00 +02:00

eval_budgetcot_pi.json

budget-CoT per-item: paired ties with latent arms confirmed (p>=0.69)

2026-07-14 17:29:56 +02:00

eval_code2gsm_loop.json

items 16 + 15-closure scored: cross-task transfer toxic (task-locality confirmed 3-point); noise-s0 50%+ was a lucky seed

2026-07-15 20:57:44 +02:00

eval_code2gsm_ns.json

items 16 + 15-closure scored: cross-task transfer toxic (task-locality confirmed 3-point); noise-s0 50%+ was a lucky seed

2026-07-15 20:57:44 +02:00

eval_code_code_ff_full.json

n=500 controls: paired net-effect significant (loop vs untrained merge 17-4, p=0.007); ladder and net accounting finalized

2026-07-14 20:08:57 +02:00

eval_code_distill_s7.json

CPU endgame: stats pass (Wilson/McNemar/pooled hard), final figures, PAPER.md rewrite around amortizable-content thesis

2026-07-14 15:46:58 +02:00

eval_code_gatehead_c_e300.json

E1c scored: easy routing solved (95.9% at 0.11 iters), hard recall regressed; threshold calibration = identified next knob

2026-07-16 01:48:14 +02:00

eval_code_halt_probe.json

halting probe: output-stable orbit, not state fixed point; no free ACT at state level; paper claims softened

2026-07-15 03:29:26 +02:00

eval_code_merge_h2048_e400.json

h2048 outcome: per-depth exonerated (sharing >> time-variation at matched params); ceiling nudged upward

2026-07-15 15:00:32 +02:00

eval_code_merge_lut.json

item 20 scored; item 21 (GSM carry-CoT + control) pre-registered and queued

2026-07-16 02:42:46 +02:00

eval_code_merge_ns_e400.json

randk + noise-s0 eval JSONs (landed pre-crash)

2026-07-15 14:15:26 +02:00

eval_code_merge_ns_s1_e400.json

items 16 + 15-closure scored: cross-task transfer toxic (task-locality confirmed 3-point); noise-s0 50%+ was a lucky seed

2026-07-15 20:57:44 +02:00

eval_code_merge_ns_s2_e400.json

items 16 + 15-closure scored: cross-task transfer toxic (task-locality confirmed 3-point); noise-s0 50%+ was a lucky seed

2026-07-15 20:57:44 +02:00

eval_code_merge_rk16_e400.json

randk + noise-s0 eval JSONs (landed pre-crash)

2026-07-15 14:15:26 +02:00

eval_code_parcae16_s1_e400.json

parcae seed 1: fidelity refutation replicates (easy ~71-73% both seeds)

2026-07-15 20:56:28 +02:00

eval_code_rec16_partial.json

rec arm outcome: rho->4.5 (norm-projected churn), prediction (a) confirmed through k=8; k16/32 cancelled

2026-07-15 01:39:22 +02:00

eval_code_ta_e400.json

tied-alpha outcome: all predictions confirmed; free B causally isolated as fidelity culprit

2026-07-15 04:13:11 +02:00

eval_code_tap23_rerun.json

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

eval_code_untrained_full.json

n=500 controls: paired net-effect significant (loop vs untrained merge 17-4, p=0.007); ladder and net accounting finalized

2026-07-14 20:08:57 +02:00

eval_gsm_carrycot_b1_ii10_ablate.json

item 27 scored: 34.0 best-in-series with zero pauses, but paired ns (burst 10-6 p=0.45; vs positional 44-38 p=0.58) — ceiling holds; pause tape confirmed ~worthless (ablation 32.4)

2026-07-17 02:12:54 +02:00

eval_gsm_carrycot_b1_ii10.json

item 27 scored: 34.0 best-in-series with zero pauses, but paired ns (burst 10-6 p=0.45; vs positional 44-38 p=0.58) — ceiling holds; pause tape confirmed ~worthless (ablation 32.4)

2026-07-17 02:12:54 +02:00

eval_gsm_carrycot_b1_kvm512.json

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

2026-07-17 21:04:46 +02:00

eval_gsm_carrycot_b1_lg03.json

item 26 scored: staging supervision actively harmful (24.2 alone, 15.6 combined) — blanket pre-'=' forcing corrupts the board's just-in-time schedule; overnight program complete, items 25-28 all scored

2026-07-17 07:18:09 +02:00

eval_gsm_carrycot_b1_lt03_lg03.json

item 26 scored: staging supervision actively harmful (24.2 alone, 15.6 combined) — blanket pre-'=' forcing corrupts the board's just-in-time schedule; overnight program complete, items 25-28 all scored

2026-07-17 07:18:09 +02:00

eval_gsm_carrycot_b1_pp30.json

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

2026-07-16 20:08:54 +02:00

eval_gsm_carrycot_b1_scst.json

item 32 scored: latent paper net-harmful — TF 20.3 (exposure catastrophe), ST 27.0 (quantization noise), analog 39.1 wins; loop=plan-machine / tokens=executor division final; architecture tree closed

2026-07-18 02:10:02 +02:00

eval_gsm_carrycot_b1_sctf.json

item 32 scored: latent paper net-harmful — TF 20.3 (exposure catastrophe), ST 27.0 (quantization noise), analog 39.1 wins; loop=plan-machine / tokens=executor division final; architecture tree closed

2026-07-18 02:10:02 +02:00

eval_gsm_carrycot_b1_tjs50_00_ablate.json

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

2026-07-17 12:24:18 +02:00

eval_gsm_carrycot_b1_tjs50_00.json

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

2026-07-17 12:24:18 +02:00

eval_gsm_carrycot_b1_tjs50_50.json

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)

2026-07-17 14:46:54 +02:00

eval_gsm_carrycot_b1_ts50_ablate.json

item 28 scored: distillation succeeded (cos 0.113→0.044), function didn't follow — 30.1 matched / 29.7 ablated, easy damaged to 48.3; state-side supervision 2x2 complete and uniformly null; read-side clamp is item 29 candidate

2026-07-17 03:41:43 +02:00

eval_gsm_carrycot_b1_ts50.json

item 28 scored: distillation succeeded (cos 0.113→0.044), function didn't follow — 30.1 matched / 29.7 ablated, easy damaged to 48.3; state-side supervision 2x2 complete and uniformly null; read-side clamp is item 29 candidate

2026-07-17 03:41:43 +02:00

eval_gsm_carrycot_b1.json

item 22 scored: ladder breaks at d=1 (57.4->31.6), plateaus ~19 at d=2/3 — 2x cold floor (d=3 vs base McNemar p=0.0025); latent loop trades easy reliability for hard/drop reach

2026-07-16 15:29:42 +02:00

eval_gsm_carrycot_b2.json

item 22 scored: ladder breaks at d=1 (57.4->31.6), plateaus ~19 at d=2/3 — 2x cold floor (d=3 vs base McNemar p=0.0025); latent loop trades easy reliability for hard/drop reach

2026-07-16 15:29:42 +02:00

eval_gsm_carrycot_b3.json

item 22 scored: ladder breaks at d=1 (57.4->31.6), plateaus ~19 at d=2/3 — 2x cold floor (d=3 vs base McNemar p=0.0025); latent loop trades easy reliability for hard/drop reach

2026-07-16 15:29:42 +02:00

eval_gsm_carrycot_b99_tjf50_00.json

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)

2026-07-17 14:46:54 +02:00

eval_gsm_carrycot_e400.json

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

eval_gsm_carrycot_ff_e400.json

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

eval_gsmonly_po_e400.json

item 17 scored: GSM-only current recipe fails as predicted; GSM chapter closed

2026-07-15 21:14:58 +02:00

eval_lcb_distill.json

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

eval_lcb_loop.json

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

eval_lcb_untrained.json

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

eval_panel_base_k0.json

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

eval_panel_distill_ff.json

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

eval_panel_loop_k2.json

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

eval_rung2_s0.json

CPU endgame: stats pass (Wilson/McNemar/pooled hard), final figures, PAPER.md rewrite around amortizable-content thesis

2026-07-14 15:46:58 +02:00

eval_rung2_s1.json

CPU endgame: stats pass (Wilson/McNemar/pooled hard), final figures, PAPER.md rewrite around amortizable-content thesis

2026-07-14 15:46:58 +02:00

fig_kcurves.png

kcurves final: all 10 panels filled

2026-07-15 20:57:45 +02:00

fig_ladder.png

n=500 controls: paired net-effect significant (loop vs untrained merge 17-4, p=0.007); ladder and net accounting finalized

2026-07-14 20:08:57 +02:00

fig_loop_vs_ff.png

ladder figure: base vs FF vs loops vs distill (fig_loop_vs_ff)

2026-07-15 12:26:22 +02:00

fig_phase.png

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

2026-07-16 09:49:21 +02:00

fig_placement.png

CPU endgame: stats pass (Wilson/McNemar/pooled hard), final figures, PAPER.md rewrite around amortizable-content thesis

2026-07-14 15:46:58 +02:00

fig_scale.png

CPU endgame: stats pass (Wilson/McNemar/pooled hard), final figures, PAPER.md rewrite around amortizable-content thesis

2026-07-14 15:46:58 +02:00

fig_transfer.png

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

gate_probs.json

item 20 scored; item 21 (GSM carry-CoT + control) pre-registered and queued

2026-07-16 02:42:46 +02:00

gate_threshold_curve.json

item 20 scored; item 21 (GSM carry-CoT + control) pre-registered and queued

2026-07-16 02:42:46 +02:00

gsm_cot_data.json

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

kvmem_carrycot_b1_ii10_kvm512_fm_p0_e200.pt

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

2026-07-17 21:04:46 +02:00

lcb_labels.json

integrate LCB dissociation (untrained merge wins far-transfer, p=0.001 vs trained loop), capability panel (no MC damage), L23 exit completes sweep

2026-07-14 19:44:07 +02:00

MATRIX.md

MATRIX.md: item 32 row + final architecture verdict

2026-07-18 02:10:40 +02:00

metacog_auc.json

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

2026-07-17 20:04:06 +02:00

metacog_lut.json

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

2026-07-17 20:04:06 +02:00

probe_discount_ff.pt

probe_discount: single-prompt whiteboard microscopy (lens table + digit probs, carry vs FF control) on the 80/15% problem

2026-07-16 19:14:23 +02:00

probe_discount_prompt.pt

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

2026-07-16 19:48:58 +02:00

probe_discount.pt

probe_discount: single-prompt whiteboard microscopy (lens table + digit probs, carry vs FF control) on the 80/15% problem

2026-07-16 19:14:23 +02:00

probe_gsm430_ff.pt

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

2026-07-16 19:48:58 +02:00

probe_gsm430_prompt.pt

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

2026-07-16 19:48:58 +02:00

probe_gsm430.pt

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

2026-07-16 19:48:58 +02:00

probe_gsm795_ff.pt

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

2026-07-16 19:48:58 +02:00

probe_gsm795_prompt.pt

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

2026-07-16 19:48:58 +02:00

probe_gsm795.pt

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

2026-07-16 19:48:58 +02:00

PROTOCOL_UNIFIED.md

item 32 scored: latent paper net-harmful — TF 20.3 (exposure catastrophe), ST 27.0 (quantization noise), analog 39.1 wins; loop=plan-machine / tokens=executor division final; architecture tree closed

2026-07-18 02:10:02 +02:00

REPORT_wave3.md

track protocol pre-registrations and interim report in git

2026-07-14 02:54:52 +02:00

stats_final.json

n=500 controls: paired net-effect significant (loop vs untrained merge 17-4, p=0.007); ladder and net accounting finalized

2026-07-14 20:08:57 +02:00

STATS.md

n=500 controls: paired net-effect significant (loop vs untrained merge 17-4, p=0.007); ladder and net accounting finalized

2026-07-14 20:08:57 +02:00

symproj_carrycot_b1_lt03_ii10_scst_fm_p0_e200.pt

item 32 scored: latent paper net-harmful — TF 20.3 (exposure catastrophe), ST 27.0 (quantization noise), analog 39.1 wins; loop=plan-machine / tokens=executor division final; architecture tree closed

2026-07-18 02:10:02 +02:00

symproj_carrycot_b1_lt03_ii10_sctf_fm_p0_e200.pt

item 32 scored: latent paper net-harmful — TF 20.3 (exposure catastrophe), ST 27.0 (quantization noise), analog 39.1 wins; loop=plan-machine / tokens=executor division final; architecture tree closed

2026-07-18 02:10:02 +02:00

train_carrycot_b1_blr16_log.json

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

2026-07-16 22:55:10 +02:00

train_carrycot_b1_ii10_kvm512_fm_p0_log.json

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

2026-07-17 21:04:46 +02:00

train_carrycot_b1_ii10_tjs50_00_p0_log.json

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

2026-07-17 12:24:18 +02:00

train_carrycot_b1_ii10_tjs50_50_p0_log.json

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)

2026-07-17 14:46:54 +02:00

train_carrycot_b1_ii10_ts50_p0_log.json

item 28 scored: distillation succeeded (cos 0.113→0.044), function didn't follow — 30.1 matched / 29.7 ablated, easy damaged to 48.3; state-side supervision 2x2 complete and uniformly null; read-side clamp is item 29 candidate

2026-07-17 03:41:43 +02:00

train_carrycot_b1_lg03_log.json

item 26 scored: staging supervision actively harmful (24.2 alone, 15.6 combined) — blanket pre-'=' forcing corrupts the board's just-in-time schedule; overnight program complete, items 25-28 all scored

2026-07-17 07:18:09 +02:00

train_carrycot_b1_log.json

item 22 scored: ladder breaks at d=1 (57.4->31.6), plateaus ~19 at d=2/3 — 2x cold floor (d=3 vs base McNemar p=0.0025); latent loop trades easy reliability for hard/drop reach

2026-07-16 15:29:42 +02:00

train_carrycot_b1_lt03_ii10_p0_log.json

item 27 scored: 34.0 best-in-series with zero pauses, but paired ns (burst 10-6 p=0.45; vs positional 44-38 p=0.58) — ceiling holds; pause tape confirmed ~worthless (ablation 32.4)

2026-07-17 02:12:54 +02:00

train_carrycot_b1_lt03_ii10_scst_fm_p0_log.json

item 32 scored: latent paper net-harmful — TF 20.3 (exposure catastrophe), ST 27.0 (quantization noise), analog 39.1 wins; loop=plan-machine / tokens=executor division final; architecture tree closed

2026-07-18 02:10:02 +02:00

train_carrycot_b1_lt03_ii10_sctf_fm_p0_log.json

item 32 scored: latent paper net-harmful — TF 20.3 (exposure catastrophe), ST 27.0 (quantization noise), analog 39.1 wins; loop=plan-machine / tokens=executor division final; architecture tree closed

2026-07-18 02:10:02 +02:00

train_carrycot_b1_lt03_lg03_log.json

item 26 scored: staging supervision actively harmful (24.2 alone, 15.6 combined) — blanket pre-'=' forcing corrupts the board's just-in-time schedule; overnight program complete, items 25-28 all scored

2026-07-17 07:18:09 +02:00

train_carrycot_b1_lt03_log.json

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

2026-07-17 00:25:32 +02:00

train_carrycot_b1_pp30_log.json

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

2026-07-16 20:08:54 +02:00

train_carrycot_b1_x600_log.json

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

2026-07-16 20:08:54 +02:00

train_carrycot_b2_log.json

item 22 scored: ladder breaks at d=1 (57.4->31.6), plateaus ~19 at d=2/3 — 2x cold floor (d=3 vs base McNemar p=0.0025); latent loop trades easy reliability for hard/drop reach

2026-07-16 15:29:42 +02:00

train_carrycot_b3_log.json

item 22 scored: ladder breaks at d=1 (57.4->31.6), plateaus ~19 at d=2/3 — 2x cold floor (d=3 vs base McNemar p=0.0025); latent loop trades easy reliability for hard/drop reach

2026-07-16 15:29:42 +02:00

train_carrycot_b99_ii10_tjf50_00_p0_log.json

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)

2026-07-17 14:46:54 +02:00