From 0a5cd80dc137fe57313c39f2d8ca28eddb1f9315 Mon Sep 17 00:00:00 2001 From: Nils Date: Wed, 15 Jul 2026 03:29:26 +0200 Subject: [PATCH] halting probe: output-stable orbit, not state fixed point; no free ACT at state level; paper claims softened Co-Authored-By: Claude Fable 5 --- PAPER.md | 13 +- results-loop/PROTOCOL_UNIFIED.md | 14 + results-loop/eval_code_halt_probe.json | 1266 ++++++++++++++++++++++++ 3 files changed, 1290 insertions(+), 3 deletions(-) create mode 100644 results-loop/eval_code_halt_probe.json diff --git a/PAPER.md b/PAPER.md index 821b96a..3af8b84 100644 --- a/PAPER.md +++ b/PAPER.md @@ -81,7 +81,8 @@ not laws.)* without recurrence matches it; stacking buys nothing (loop-training a distill-warmed adapter: 34.5%, below distill alone; running the distilled adapter in loop mode: drops to 20.0%); deeper k at inference is flat - (k=8: 40.0%). The recurrence is a *training-time scaffold* that lets the + (k=8: 40.0%; output-stable despite residual state drift, §3.8). The +recurrence is a *training-time scaffold* that lets the adapter find plan-shaped content — content that can equally be put there by distillation if plans are available. @@ -346,8 +347,14 @@ plan-representation overlaps; task-specific training still dominates. ### 3.8 Mechanism, verification, deployment The trained loop takes a large first step (cos(s₁,s₀)=0.926 vs 0.977 -untrained) and converges bit-exactly by k≈3–4; accuracy and lens-sharpening -plateau there. P(latent concept) under the J-lens at the band exit rises +untrained); accuracy and lens-sharpening plateau by k≈3–4. A population +probe (n=250, state-cosine threshold 0.9995) shows the plateau is +*output-level*: half the prompts' states are still drifting at 1e-3–1e-4 +cosine scale at k=8 while generation is already depth-stable — an +output-stable orbit rather than a literal state fixed point, with no +difficulty gradient in state-convergence depth. Consequently, +convergence-based early exit ("free ACT") does not fall out of the state +trajectory; halting would need an output-level signal. P(latent concept) under the J-lens at the band exit rises 0.015→0.13 across iterations (~8× the untrained hold) — the lens that placed the implant also renders its silent content inspectable. The STaR labels train a free difficulty gate (route predicted-hard to k=4, else k=0); diff --git a/results-loop/PROTOCOL_UNIFIED.md b/results-loop/PROTOCOL_UNIFIED.md index 590d71e..ae63141 100644 --- a/results-loop/PROTOCOL_UNIFIED.md +++ b/results-loop/PROTOCOL_UNIFIED.md @@ -230,3 +230,17 @@ training depth) and no capability; map-sharing is load-bearing for the anytime-usable gain. Depth-4 adapter overfit visible in val (hard k4 CE 0.188@99 -> 0.371@599) — LTV concentrates small-pool overfitting into single depths. + +--- Outcome, item 13b (scored 2026-07-15): accuracy prediction CONFIRMED +(k=8 halt run 52.0/90.2/42.9 = plateau level); convergence predictions +REFUTED. Per-item state-cosine (thresh 0.9995, k=8 cap): k_conv +distribution 4:3, 5:57, 6:47, 7:17, never-within-8:126 — mean ~7, and NO +difficulty gradient (easy 7.01 vs hard 7.00). The earlier "bit-exact by +k~3-4" was the single dynamics-probe example, not the population: outputs +plateau by k~2-4 while the state keeps drifting at 1e-3..1e-4 cosine +scale; the fixed point is an OUTPUT-stable orbit (suffix layers + decode +wash out residual state motion), not a literal state fixed point for most +prompts. Free-ACT via state-cosine therefore yields no early exit at this +threshold, and no ACT-like difficulty allocation falls out for free — +output-level halting signals would be needed. Paper's dynamics claims +softened accordingly. diff --git a/results-loop/eval_code_halt_probe.json b/results-loop/eval_code_halt_probe.json new file mode 100644 index 0000000..2ad1983 --- /dev/null +++ b/results-loop/eval_code_halt_probe.json @@ -0,0 +1,1266 @@ +{ + "tag": "halt_probe", + "ks": { + "8": { + "acc": 0.52, + "by_label": { + "easy": 0.9016393442622951, + "hard": 0.42857142857142855, + "drop": 0.08 + }, + "per_item": [ + { + "task_id": 11, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 12, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 13, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 14, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 15, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 16, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 17, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 18, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 19, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 20, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 21, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 22, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 23, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 24, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 25, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 26, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 27, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 28, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 29, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 30, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 31, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 32, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 33, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 34, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 35, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 36, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 37, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 38, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 39, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 40, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 41, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 42, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 43, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 44, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 45, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 46, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 47, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 48, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 49, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 50, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 51, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 52, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 53, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 54, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 55, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 56, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 57, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 58, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 59, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 60, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 61, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 62, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 63, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 64, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 65, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 66, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 67, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 68, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 69, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 70, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 71, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 72, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 73, + "ok": false, + "k_conv": 7 + }, + { + "task_id": 74, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 75, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 76, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 77, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 78, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 79, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 80, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 81, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 82, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 83, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 84, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 85, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 86, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 87, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 88, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 89, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 90, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 91, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 92, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 93, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 94, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 95, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 96, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 97, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 98, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 99, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 100, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 101, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 102, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 103, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 104, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 105, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 106, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 107, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 108, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 109, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 110, + "ok": false, + "k_conv": 4 + }, + { + "task_id": 111, + "ok": false, + "k_conv": 4 + }, + { + "task_id": 112, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 113, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 114, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 115, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 116, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 117, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 118, + "ok": false, + "k_conv": 7 + }, + { + "task_id": 119, + "ok": false, + "k_conv": 7 + }, + { + "task_id": 120, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 121, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 122, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 123, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 124, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 125, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 126, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 127, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 128, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 129, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 130, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 131, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 132, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 133, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 134, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 135, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 136, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 137, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 138, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 139, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 140, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 141, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 142, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 143, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 144, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 145, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 146, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 147, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 148, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 149, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 150, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 151, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 152, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 153, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 154, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 155, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 156, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 157, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 158, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 159, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 160, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 161, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 162, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 163, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 164, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 165, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 166, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 167, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 168, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 169, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 170, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 171, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 172, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 173, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 174, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 175, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 176, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 177, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 178, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 179, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 180, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 181, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 182, + "ok": false, + "k_conv": 7 + }, + { + "task_id": 183, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 184, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 185, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 186, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 187, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 188, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 189, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 190, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 191, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 192, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 193, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 194, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 195, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 196, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 197, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 198, + "ok": false, + "k_conv": 7 + }, + { + "task_id": 199, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 200, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 201, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 202, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 203, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 204, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 205, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 206, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 207, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 208, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 209, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 210, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 211, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 212, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 213, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 214, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 215, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 216, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 217, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 218, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 219, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 220, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 221, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 222, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 223, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 224, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 225, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 226, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 227, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 228, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 229, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 230, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 231, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 232, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 233, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 234, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 235, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 236, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 237, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 238, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 239, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 240, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 241, + "ok": false, + "k_conv": 4 + }, + { + "task_id": 242, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 243, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 244, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 245, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 246, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 247, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 248, + "ok": false, + "k_conv": 8 + }, + { + "task_id": 249, + "ok": true, + "k_conv": 6 + }, + { + "task_id": 250, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 251, + "ok": true, + "k_conv": 7 + }, + { + "task_id": 252, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 253, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 254, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 255, + "ok": false, + "k_conv": 5 + }, + { + "task_id": 256, + "ok": true, + "k_conv": 5 + }, + { + "task_id": 257, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 258, + "ok": true, + "k_conv": 8 + }, + { + "task_id": 259, + "ok": false, + "k_conv": 6 + }, + { + "task_id": 260, + "ok": false, + "k_conv": 5 + } + ] + } + }, + "n": 250 +} \ No newline at end of file