Text set in serif type with a dotted bar in the margin was drafted by an LLM (claude-fable-5) and sometimes reviewed by the author. The rest is the author's own. How to read this site
mirror_v1
20260716_133911_mirror_v1 · complete · published 2026-07-17 · seed 42
Intent
This experiment tests directly whether the direction of an α schedule matters in itself. Each mirror pair runs the same α values for the same number of steps with the same seeds, once descending and once ascending; each segment (stage) draws its training pool fresh at its own α, so direction is the only systematic difference.
The grid covers every pair of integer endpoints from 0 to 5 in two shapes: the smooth 24-stage linear ramp of slopesweep_v1, and a jagged ladder with one rung per integer between the endpoints, so an adjacent pair degenerates to a single abrupt switch. Flat arms at each grid value calibrate what no direction looks like. All arms use the with replacement data setting. Design: the mirror_v1 spec.
Background
Every curriculum result so far points one way: a schedule that starts at high pairing concentration and moves toward low concentration builds the agreement rule, and time spent at low α erodes it, the build-and-erode account assembled by curriculum_v1 and burst_v1.
The descending arms to 0 of slopesweep_v1 collapse by the end of training, finals between 19.0 and 32.1 from every start. The collapsing arms of slopesweep_v1 showed seed bimodality, its s5_e0 finishing at 32.1 ±27.4 with seeds spanning 0.4 to 85.3.
Hypothesis
For symmetric shapes, the higher-to-lower direction always wins on unseen_mismatch: in every mirror pair, in both shape families, the descending arm's cohort mean beats the ascending arm's at its peak waypoint and at the end of training, with a 95% bootstrap interval on the difference that excludes zero. Spread is a separate competition: the descending arm is expected to be no less consistent, its peak standard deviation at most the ascending arm's.
The exposed flank is the pairs ending at 0: an ascending mirror finishes its training at high α with a freshly built rule. If ascent wins there at final, the operative variable is the final α, not the direction.
Setup
| arm | data regime | datasets | iters/dataset | epochs/dataset |
|---|---|---|---|---|
| s0_e0 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s0_e1 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s0_e1_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s0_e2 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s0_e2_blk | fresh-dataset-per-stage | 3 | 400 | 0.33 |
| s0_e3 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s0_e3_blk | fresh-dataset-per-stage | 4 | 300 | 0.25 |
| s0_e4 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s0_e4_blk | fresh-dataset-per-stage | 5 | 240 | 0.2 |
| s0_e5 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s0_e5_blk | fresh-dataset-per-stage | 6 | 200 | 0.17 |
| s1_e0 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s1_e0_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s1_e1 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s1_e2 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s1_e2_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s1_e3 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s1_e3_blk | fresh-dataset-per-stage | 3 | 400 | 0.33 |
| s1_e4 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s1_e4_blk | fresh-dataset-per-stage | 4 | 300 | 0.25 |
| s1_e5 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s1_e5_blk | fresh-dataset-per-stage | 5 | 240 | 0.2 |
| s2_e0 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s2_e0_blk | fresh-dataset-per-stage | 3 | 400 | 0.33 |
| s2_e1 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s2_e1_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s2_e2 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s2_e3 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s2_e3_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s2_e4 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s2_e4_blk | fresh-dataset-per-stage | 3 | 400 | 0.33 |
| s2_e5 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s2_e5_blk | fresh-dataset-per-stage | 4 | 300 | 0.25 |
| s3_e0 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s3_e0_blk | fresh-dataset-per-stage | 4 | 300 | 0.25 |
| s3_e1 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s3_e1_blk | fresh-dataset-per-stage | 3 | 400 | 0.33 |
| s3_e2 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s3_e2_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s3_e3 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s3_e4 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s3_e4_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s3_e5 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s3_e5_blk | fresh-dataset-per-stage | 3 | 400 | 0.33 |
| s4_e0 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s4_e0_blk | fresh-dataset-per-stage | 5 | 240 | 0.2 |
| s4_e1 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s4_e1_blk | fresh-dataset-per-stage | 4 | 300 | 0.25 |
| s4_e2 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s4_e2_blk | fresh-dataset-per-stage | 3 | 400 | 0.33 |
| s4_e3 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s4_e3_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s4_e4 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s4_e5 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s4_e5_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s5_e0 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s5_e0_blk | fresh-dataset-per-stage | 6 | 200 | 0.17 |
| s5_e1 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s5_e1_blk | fresh-dataset-per-stage | 5 | 240 | 0.2 |
| s5_e2 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s5_e2_blk | fresh-dataset-per-stage | 4 | 300 | 0.25 |
| s5_e3 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s5_e3_blk | fresh-dataset-per-stage | 3 | 400 | 0.33 |
| s5_e4 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
| s5_e4_blk | fresh-dataset-per-stage | 2 | 600 | 0.5 |
| s5_e5 | fresh-dataset-per-stage | 24 | 50 | 0.04 |
Replicate this experiment
Get the code at this exact version:
git clone https://github.com/ClaireHobbs/imagining-syntax
cd imagining-syntax
git checkout ae5f77126dc8
pip install -e ".[dev]"
Download the config (experiment.yaml) and run:
imsyn exp run experiment.yamlChecks
PASS children_complete PASS replicate_coverage PASS schedule_total_matches_final_waypoint PASS instrument_consistent
Claims
| claim | source | expectation | Δ | seeds | verdict |
|---|---|---|---|---|---|
| lin_1to0_mean_final | manifest | s1_e0 > s0_e1 on unseen_mismatch @ final | -7.0 | — | mixed 95% CI on delta [-23.6, +10.8] straddles 0 |
| lin_1to0_mean_peak | manifest | s1_e0 > s0_e1 on unseen_mismatch @ peak | +0.5 | — | mixed 95% CI on delta [-0.5, +1.5] straddles 0 |
| lin_1to0_sd_peak | manifest | s1_e0 <= s0_e1 on unseen_mismatch (sd) @ peak | +0.1 | — | rejected mean-only (no per-seed vectors) |
| lin_2to0_mean_final | manifest | s2_e0 > s0_e2 on unseen_mismatch @ final | -70.2 | — | rejected |
| lin_2to0_mean_peak | manifest | s2_e0 > s0_e2 on unseen_mismatch @ peak | +7.1 | — | held |
| lin_2to0_sd_peak | manifest | s2_e0 <= s0_e2 on unseen_mismatch (sd) @ peak | -7.8 | — | held mean-only (no per-seed vectors) |
| lin_2to1_mean_final | manifest | s2_e1 > s1_e2 on unseen_mismatch @ final | +3.3 | — | held |
| lin_2to1_mean_peak | manifest | s2_e1 > s1_e2 on unseen_mismatch @ peak | -2.4 | — | rejected |
| lin_2to1_sd_peak | manifest | s2_e1 <= s1_e2 on unseen_mismatch (sd) @ peak | +1.2 | — | rejected mean-only (no per-seed vectors) |
| lin_3to0_mean_final | manifest | s3_e0 > s0_e3 on unseen_mismatch @ final | -65.0 | — | rejected |
| lin_3to0_mean_peak | manifest | s3_e0 > s0_e3 on unseen_mismatch @ peak | +5.6 | — | mixed 95% CI on delta [-2.5, +16.3] straddles 0 |
| lin_3to0_sd_peak | manifest | s3_e0 <= s0_e3 on unseen_mismatch (sd) @ peak | -9.1 | — | held mean-only (no per-seed vectors) |
| lin_3to1_mean_final | manifest | s3_e1 > s1_e3 on unseen_mismatch @ final | +19.0 | — | held |
| lin_3to1_mean_peak | manifest | s3_e1 > s1_e3 on unseen_mismatch @ peak | -1.6 | — | mixed 95% CI on delta [-5.7, +2.5] straddles 0 |
| lin_3to1_sd_peak | manifest | s3_e1 <= s1_e3 on unseen_mismatch (sd) @ peak | -0.4 | — | held mean-only (no per-seed vectors) |
| lin_3to2_mean_final | manifest | s3_e2 > s2_e3 on unseen_mismatch @ final | +8.5 | — | held |
| lin_3to2_mean_peak | manifest | s3_e2 > s2_e3 on unseen_mismatch @ peak | -5.6 | — | rejected |
| lin_3to2_sd_peak | manifest | s3_e2 <= s2_e3 on unseen_mismatch (sd) @ peak | -1.5 | — | held mean-only (no per-seed vectors) |
| lin_4to0_mean_final | manifest | s4_e0 > s0_e4 on unseen_mismatch @ final | -49.2 | — | rejected |
| lin_4to0_mean_peak | manifest | s4_e0 > s0_e4 on unseen_mismatch @ peak | +2.7 | — | mixed 95% CI on delta [-3.4, +10.9] straddles 0 |
| lin_4to0_sd_peak | manifest | s4_e0 <= s0_e4 on unseen_mismatch (sd) @ peak | -7.2 | — | held mean-only (no per-seed vectors) |
| lin_4to1_mean_final | manifest | s4_e1 > s1_e4 on unseen_mismatch @ final | +21.5 | — | held |
| lin_4to1_mean_peak | manifest | s4_e1 > s1_e4 on unseen_mismatch @ peak | -5.0 | — | rejected |
| lin_4to1_sd_peak | manifest | s4_e1 <= s1_e4 on unseen_mismatch (sd) @ peak | +3.0 | — | rejected mean-only (no per-seed vectors) |
| lin_4to2_mean_final | manifest | s4_e2 > s2_e4 on unseen_mismatch @ final | +10.2 | — | held |
| lin_4to2_mean_peak | manifest | s4_e2 > s2_e4 on unseen_mismatch @ peak | -6.5 | — | rejected |
| lin_4to2_sd_peak | manifest | s4_e2 <= s2_e4 on unseen_mismatch (sd) @ peak | -0.6 | — | held mean-only (no per-seed vectors) |
| lin_4to3_mean_final | manifest | s4_e3 > s3_e4 on unseen_mismatch @ final | +2.5 | — | mixed 95% CI on delta [-2.6, +8.0] straddles 0 |
| lin_4to3_mean_peak | manifest | s4_e3 > s3_e4 on unseen_mismatch @ peak | +3.0 | — | mixed 95% CI on delta [-2.1, +8.0] straddles 0 |
| lin_4to3_sd_peak | manifest | s4_e3 <= s3_e4 on unseen_mismatch (sd) @ peak | -1.5 | — | held mean-only (no per-seed vectors) |
| lin_5to0_mean_final | manifest | s5_e0 > s0_e5 on unseen_mismatch @ final | -43.5 | — | rejected |
| lin_5to0_mean_peak | manifest | s5_e0 > s0_e5 on unseen_mismatch @ peak | +10.3 | — | held |
| lin_5to0_sd_peak | manifest | s5_e0 <= s0_e5 on unseen_mismatch (sd) @ peak | +3.5 | — | rejected mean-only (no per-seed vectors) |
| lin_5to1_mean_final | manifest | s5_e1 > s1_e5 on unseen_mismatch @ final | +26.9 | — | held |
| lin_5to1_mean_peak | manifest | s5_e1 > s1_e5 on unseen_mismatch @ peak | -1.6 | — | mixed 95% CI on delta [-5.6, +1.9] straddles 0 |
| lin_5to1_sd_peak | manifest | s5_e1 <= s1_e5 on unseen_mismatch (sd) @ peak | +1.6 | — | rejected mean-only (no per-seed vectors) |
| lin_5to2_mean_final | manifest | s5_e2 > s2_e5 on unseen_mismatch @ final | +13.4 | — | held |
| lin_5to2_mean_peak | manifest | s5_e2 > s2_e5 on unseen_mismatch @ peak | -5.9 | — | mixed 95% CI on delta [-12.3, +0.9] straddles 0 |
| lin_5to2_sd_peak | manifest | s5_e2 <= s2_e5 on unseen_mismatch (sd) @ peak | -1.5 | — | held mean-only (no per-seed vectors) |
| lin_5to3_mean_final | manifest | s5_e3 > s3_e5 on unseen_mismatch @ final | +3.8 | — | held |
| lin_5to3_mean_peak | manifest | s5_e3 > s3_e5 on unseen_mismatch @ peak | +1.7 | — | mixed 95% CI on delta [-2.9, +6.1] straddles 0 |
| lin_5to3_sd_peak | manifest | s5_e3 <= s3_e5 on unseen_mismatch (sd) @ peak | +1.3 | — | rejected mean-only (no per-seed vectors) |
| lin_5to4_mean_final | manifest | s5_e4 > s4_e5 on unseen_mismatch @ final | +1.0 | — | mixed 95% CI on delta [-2.4, +4.3] straddles 0 |
| lin_5to4_mean_peak | manifest | s5_e4 > s4_e5 on unseen_mismatch @ peak | +0.6 | — | mixed 95% CI on delta [-4.0, +4.9] straddles 0 |
| lin_5to4_sd_peak | manifest | s5_e4 <= s4_e5 on unseen_mismatch (sd) @ peak | +0.2 | — | rejected mean-only (no per-seed vectors) |
| blk_1to0_mean_final | manifest | s1_e0_blk > s0_e1_blk on unseen_mismatch @ final | -47.7 | — | rejected |
| blk_1to0_mean_peak | manifest | s1_e0_blk > s0_e1_blk on unseen_mismatch @ peak | -23.5 | — | rejected |
| blk_1to0_sd_peak | manifest | s1_e0_blk <= s0_e1_blk on unseen_mismatch (sd) @ peak | +11.5 | — | rejected mean-only (no per-seed vectors) |
| blk_2to0_mean_final | manifest | s2_e0_blk > s0_e2_blk on unseen_mismatch @ final | -72.7 | — | rejected
referenced by
Final α, not direction, sets the endpoint
|
| blk_2to0_mean_peak | manifest | s2_e0_blk > s0_e2_blk on unseen_mismatch @ peak | +1.5 | — | held |
| blk_2to0_sd_peak | manifest | s2_e0_blk <= s0_e2_blk on unseen_mismatch (sd) @ peak | -1.3 | — | held mean-only (no per-seed vectors) |
| blk_2to1_mean_final | manifest | s2_e1_blk > s1_e2_blk on unseen_mismatch @ final | +1.9 | — | mixed 95% CI on delta [-2.3, +5.8] straddles 0 |
| blk_2to1_mean_peak | manifest | s2_e1_blk > s1_e2_blk on unseen_mismatch @ peak | -0.4 | — | mixed 95% CI on delta [-2.1, +1.3] straddles 0 |
| blk_2to1_sd_peak | manifest | s2_e1_blk <= s1_e2_blk on unseen_mismatch (sd) @ peak | -0.2 | — | held mean-only (no per-seed vectors) |
| blk_3to0_mean_final | manifest | s3_e0_blk > s0_e3_blk on unseen_mismatch @ final | -65.7 | — | rejected |
| blk_3to0_mean_peak | manifest | s3_e0_blk > s0_e3_blk on unseen_mismatch @ peak | +0.2 | — | mixed 95% CI on delta [-5.3, +6.4] straddles 0 |
| blk_3to0_sd_peak | manifest | s3_e0_blk <= s0_e3_blk on unseen_mismatch (sd) @ peak | -4.3 | — | held mean-only (no per-seed vectors) |
| blk_3to1_mean_final | manifest | s3_e1_blk > s1_e3_blk on unseen_mismatch @ final | +5.6 | — | mixed 95% CI on delta [-1.9, +12.7] straddles 0 |
| blk_3to1_mean_peak | manifest | s3_e1_blk > s1_e3_blk on unseen_mismatch @ peak | -0.5 | — | mixed 95% CI on delta [-3.6, +2.1] straddles 0 |
| blk_3to1_sd_peak | manifest | s3_e1_blk <= s1_e3_blk on unseen_mismatch (sd) @ peak | +1.5 | — | rejected mean-only (no per-seed vectors) |
| blk_3to2_mean_final | manifest | s3_e2_blk > s2_e3_blk on unseen_mismatch @ final | +4.9 | — | mixed 95% CI on delta [-2.1, +10.8] straddles 0 |
| blk_3to2_mean_peak | manifest | s3_e2_blk > s2_e3_blk on unseen_mismatch @ peak | -11.4 | — | rejected |
| blk_3to2_sd_peak | manifest | s3_e2_blk <= s2_e3_blk on unseen_mismatch (sd) @ peak | +4.2 | — | rejected mean-only (no per-seed vectors) |
| blk_4to0_mean_final | manifest | s4_e0_blk > s0_e4_blk on unseen_mismatch @ final | -53.7 | — | rejected |
| blk_4to0_mean_peak | manifest | s4_e0_blk > s0_e4_blk on unseen_mismatch @ peak | +7.6 | — | held |
| blk_4to0_sd_peak | manifest | s4_e0_blk <= s0_e4_blk on unseen_mismatch (sd) @ peak | -10.7 | — | held mean-only (no per-seed vectors) |
| blk_4to1_mean_final | manifest | s4_e1_blk > s1_e4_blk on unseen_mismatch @ final | +20.7 | — | held |
| blk_4to1_mean_peak | manifest | s4_e1_blk > s1_e4_blk on unseen_mismatch @ peak | +5.1 | — | mixed 95% CI on delta [-1.3, +12.6] straddles 0 |
| blk_4to1_sd_peak | manifest | s4_e1_blk <= s1_e4_blk on unseen_mismatch (sd) @ peak | -8.9 | — | held mean-only (no per-seed vectors) |
| blk_4to2_mean_final | manifest | s4_e2_blk > s2_e4_blk on unseen_mismatch @ final | +9.4 | — | held |
| blk_4to2_mean_peak | manifest | s4_e2_blk > s2_e4_blk on unseen_mismatch @ peak | -12.5 | — | rejected |
| blk_4to2_sd_peak | manifest | s4_e2_blk <= s2_e4_blk on unseen_mismatch (sd) @ peak | +2.9 | — | rejected mean-only (no per-seed vectors) |
| blk_4to3_mean_final | manifest | s4_e3_blk > s3_e4_blk on unseen_mismatch @ final | +3.2 | — | mixed 95% CI on delta [-1.4, +7.7] straddles 0 |
| blk_4to3_mean_peak | manifest | s4_e3_blk > s3_e4_blk on unseen_mismatch @ peak | -1.2 | — | mixed 95% CI on delta [-4.5, +1.9] straddles 0 |
| blk_4to3_sd_peak | manifest | s4_e3_blk <= s3_e4_blk on unseen_mismatch (sd) @ peak | +0.0 | — | held mean-only (no per-seed vectors) |
| blk_5to0_mean_final | manifest | s5_e0_blk > s0_e5_blk on unseen_mismatch @ final | -40.4 | — | rejected |
| blk_5to0_mean_peak | manifest | s5_e0_blk > s0_e5_blk on unseen_mismatch @ peak | +3.0 | — | mixed 95% CI on delta [-2.2, +9.2] straddles 0 |
| blk_5to0_sd_peak | manifest | s5_e0_blk <= s0_e5_blk on unseen_mismatch (sd) @ peak | -4.8 | — | held mean-only (no per-seed vectors) |
| blk_5to1_mean_final | manifest | s5_e1_blk > s1_e5_blk on unseen_mismatch @ final | +22.7 | — | held |
| blk_5to1_mean_peak | manifest | s5_e1_blk > s1_e5_blk on unseen_mismatch @ peak | -0.6 | — | mixed 95% CI on delta [-3.3, +2.1] straddles 0 |
| blk_5to1_sd_peak | manifest | s5_e1_blk <= s1_e5_blk on unseen_mismatch (sd) @ peak | -0.1 | — | held mean-only (no per-seed vectors) |
| blk_5to2_mean_final | manifest | s5_e2_blk > s2_e5_blk on unseen_mismatch @ final | +10.6 | — | held |
| blk_5to2_mean_peak | manifest | s5_e2_blk > s2_e5_blk on unseen_mismatch @ peak | -14.2 | — | rejected |
| blk_5to2_sd_peak | manifest | s5_e2_blk <= s2_e5_blk on unseen_mismatch (sd) @ peak | +5.3 | — | rejected mean-only (no per-seed vectors) |
| blk_5to3_mean_final | manifest | s5_e3_blk > s3_e5_blk on unseen_mismatch @ final | +4.2 | — | held |
| blk_5to3_mean_peak | manifest | s5_e3_blk > s3_e5_blk on unseen_mismatch @ peak | -2.1 | — | mixed 95% CI on delta [-6.1, +1.7] straddles 0 |
| blk_5to3_sd_peak | manifest | s5_e3_blk <= s3_e5_blk on unseen_mismatch (sd) @ peak | +1.6 | — | rejected mean-only (no per-seed vectors) |
| blk_5to4_mean_final | manifest | s5_e4_blk > s4_e5_blk on unseen_mismatch @ final | +1.6 | — | mixed 95% CI on delta [-1.9, +4.8] straddles 0 |
| blk_5to4_mean_peak | manifest | s5_e4_blk > s4_e5_blk on unseen_mismatch @ peak | +0.0 | — | mixed 95% CI on delta [-3.7, +3.5] straddles 0 |
| blk_5to4_sd_peak | manifest | s5_e4_blk <= s4_e5_blk on unseen_mismatch (sd) @ peak | -0.9 | — | held mean-only (no per-seed vectors) |
Results — unseen_mismatch (mean ± sd over 10 seeds)
Evaluation data: seen_* probed at reference α = 1.4 · unseen_* = held-out pairings, uniform (α-independent) · 1000 pairs per condition
- Each probe is a minimal pair: a grammatical sentence and its verb-number-flipped twin. The model scores correct when it assigns the grammatical version higher probability; accuracy = % correct over 1000 pairs per condition.
- seen_match / seen_mismatch — noun–verb pairings that occur in training, sampled at fixed reference α = 1.4 (not at the schedule's own α, so arms stay comparable)
- @α conditions (e.g. seen_match@0) — the same seen probes regenerated at reference α = 0, 0.7, 2.1, 3
- unseen_match / unseen_mismatch — held-out noun–verb pairings that never occur in training, sampled uniformly (α = 0), so their difficulty is identical across all arms and training αs
- match vs mismatch — whether the prepositional objects agree in number with the subject; mismatch places attractor nouns between subject and verb (the AGREE-RECENT trap)
| arm | end of training (t = 1200) | per-seed | flags |
|---|---|---|---|
| s0_e0 | 26.1 ±30.2 | ⚠ seed_split: range 0.0-96.8 (3 low / 1 high of 10) | |
| s0_e1 | 21.4 ±20.9 | ⚠ seed_split: range 0.0-52.5 (5 low / 2 high of 10) | |
| s0_e1_blk | 83.0 ±20.2 | ⚠ seed_split: range 43.6-100.0 (2 low / 6 high of 10) | |
| s0_e2 | 90.0 ±11.2 | ⚠ seed_split: range 59.5-100.0 (1 low / 7 high of 10) | |
| s0_e2_blk | 94.1 ±3.9 | ||
| s0_e3 | 79.8 ±7.2 | ||
| s0_e3_blk | 81.4 ±4.6 | ||
| s0_e4 | 70.9 ±3.7 | ||
| s0_e4_blk | 72.4 ±4.0 | ||
| s0_e5 | 68.7 ±3.5 | ||
| s0_e5_blk | 66.3 ±2.7 | ||
| s1_e0 | 14.4 ±17.9 | ⚠ seed_split: range 0.0-51.1 (5 low / 2 high of 10) | |
| s1_e0_blk | 35.3 ±30.1 | ⚠ seed_split: range 0.0-96.9 (3 low / 1 high of 10) | |
| s1_e1 | 85.6 ±22.1 | ⚠ seed_split: range 25.0-100.0 (1 low / 6 high of 10) | |
| s1_e2 | 90.8 ±3.4 | ||
| s1_e2_blk | 89.1 ±3.7 | ||
| s1_e3 | 75.1 ±3.5 | ||
| s1_e3_blk | 77.7 ±3.4 | ||
| s1_e4 | 70.1 ±4.3 | ||
| s1_e4_blk | 68.9 ±2.7 | ||
| s1_e5 | 64.3 ±2.2 | ||
| s1_e5_blk | 67.1 ±2.9 | ||
| s2_e0 | 19.8 ±11.9 | ||
| s2_e0_blk | 21.4 ±15.8 | ⚠ seed_split: range 1.2-48.4 (3 low / 2 high of 10) | |
| s2_e1 | 94.1 ±3.3 | ||
| s2_e1_blk | 91.0 ±5.4 | ||
| s2_e2 | 84.3 ±4.7 | ||
| s2_e3 | 76.2 ±4.3 | ||
| s2_e3_blk | 78.4 ±4.9 | ||
| s2_e4 | 73.3 ±5.0 | ||
| s2_e4_blk | 73.8 ±4.4 | ||
| s2_e5 | 69.8 ±3.8 | ||
| s2_e5_blk | 71.7 ±3.6 | ||
| s3_e0 | 14.8 ±7.7 | ||
| s3_e0_blk | 15.7 ±10.6 | ||
| s3_e1 | 94.1 ±4.5 | ||
| s3_e1_blk | 83.3 ±11.5 | ⚠ seed_split: range 58.8-99.0 (1 low / 4 high of 10) | |
| s3_e2 | 84.7 ±5.7 | ||
| s3_e2_blk | 83.3 ±9.4 | ||
| s3_e3 | 75.2 ±5.2 | ||
| s3_e4 | 70.5 ±5.4 | ||
| s3_e4_blk | 71.4 ±5.7 | ||
| s3_e5 | 68.4 ±2.7 | ||
| s3_e5_blk | 69.9 ±4.8 | ||
| s4_e0 | 21.7 ±23.0 | ⚠ seed_split: range 1.0-82.4 (5 low / 1 high of 10) | |
| s4_e0_blk | 18.7 ±16.2 | ⚠ seed_split: range 2.8-51.5 (5 low / 1 high of 10) | |
| s4_e1 | 91.6 ±9.3 | ||
| s4_e1_blk | 89.6 ±10.9 | ||
| s4_e2 | 83.5 ±5.3 | ||
| s4_e2_blk | 83.2 ±5.9 | ||
| s4_e3 | 73.0 ±6.6 | ||
| s4_e3_blk | 74.6 ±4.6 | ||
| s4_e4 | 69.8 ±3.6 | ||
| s4_e5 | 67.2 ±2.9 | ||
| s4_e5_blk | 67.2 ±4.2 | ||
| s5_e0 | 25.2 ±20.8 | ⚠ seed_split: range 6.6-76.6 (4 low / 1 high of 10) | |
| s5_e0_blk | 25.9 ±24.3 | ⚠ seed_split: range 10.2-96.9 (7 low / 1 high of 10) | |
| s5_e1 | 91.2 ±4.7 | ||
| s5_e1_blk | 89.8 ±6.4 | ||
| s5_e2 | 83.2 ±6.7 | ||
| s5_e2_blk | 82.3 ±7.2 | ||
| s5_e3 | 72.2 ±4.7 | ||
| s5_e3_blk | 74.1 ±3.3 | ||
| s5_e4 | 68.2 ±4.5 | ||
| s5_e4_blk | 68.8 ±3.6 | ||
| s5_e5 | 66.2 ±3.9 |
Conclusions
The registered hypothesis is rejected. Of the 90 pre-registered claims, 34 held, 26 were mixed, and 30 were rejected, and the failures are not noise scattered over the table: they fall into three bands, each of which replaces part of the hypothesis with something sharper.
Descent wins the contest it was designed around, retention at a safe floor.
Across both shape families, no descending arm with a floor at α = 1 or above lost its final-value claim on unseen_mismatch: of the twenty such claims, thirteen held, with margins from 3.3 to 26.9 points, and seven straddled, the ties concentrated at short spans.
Referenced by (1 direct, 1 transitive)
Direct references:
Transitive (depth 1):
Start high and land in the working range, and the anneal ends ahead of its mirror.
Every pair ending at 0 reverses this, and not narrowly. Nine of those ten final-value claims were rejected, with the ascending arm ahead by 40 to 73 points (block 2→0: Δ −72.7, 95% interval −82.7 to −62.7); the tenth, the smooth 1→0 pair, straddled. The reading is direct: the descending arm passes through the uniform tail after it has built the rule and loses it to catastrophic interference, while its mirror spends the same uniform segments before there is anything to erase, and finishes training at peak concentration. At the horizon the final α, not the direction of travel, sets the outcome.
Referenced by (2 direct, 1 transitive)
Direct references:
Transitive (depth 1):
This is the falsification condition the hypothesis named in advance, and it fired.
The peak claims hold the result nothing in the prior work predicted.
Only four of thirty held, all on pairs ending at 0, where the descending arm's mid-run peak precedes its collapse. Eight were rejected, and seven of those are interior pairs where the ascending arm reaches the higher peak: every block pair ending at 2 (5→2: Δ −14.2, interval [−19.1, −10.0]; 4→2: −12.5; 3→2: −11.4), two of the three smooth ones (4→2: −6.5; 3→2: −5.6; 5→2 straddled at −5.9), and the smooth 4→1 and 2→1 (−5.0 and −2.4); the eighth, block 1→0 (−23.5), is a pair whose descending arm never builds the rule before the uniform tail takes it. The timing tempers the surprise without removing it: the ascending arms peak early, at 250 to 500 steps (s2_e4_blk reaches 96.5 at 300), while their descending mirrors' peaks sit at the horizon itself, 1100 to 1200 and still rising, which is the signature of undertraining under a fixed budget.
Referenced by (1 direct, 1 transitive)
Direct references:
Transitive (depth 1):
Whether the anneal would catch those peaks given a longer run is open; extending the horizon on the interior pairs is the obvious follow-up.
The consistency competition split 17 held to 13 rejected, so descent is not reliably the steadier direction, and where it collapses it is also the wilder one: the block 1→0 spread claim was rejected with the descending arm's peak standard deviation 11.5 points larger, seed bimodality again rather than uniform mediocrity. The adjacent pairs 5↔4 and 4↔3 tied every mean claim in both families; at a span of one grid step, direction is invisible at ten replicates.
Three notes on reading the table. Mirror pairs in which both arms sit at saturation produce degenerate intervals and adjudicate as mixed by construction, so a mixed peak verdict often means both arms at ceiling rather than genuine ambiguity. The ten-seed percentile bootstrap is mildly anti-conservative, so an isolated hairline hold should not be over-read. The mirror comparison figures rendered with the experiment show each pair side by side and are the fastest way to see the three bands above.
Referenced by (2 direct, 4 transitive)
Direct references:
Comparison figures
Children
| Child | peak unseen_mismatch | Status |
|---|---|---|
| s5_e5 | 66.2 | done |
| s5_e4 | 69.0 | done |
| s5_e3 | 73.0 | done |
| s5_e2 | 83.2 | done |
| s5_e1 | 93.3 | done |
| s5_e0 | 92.7 | done |
| s4_e4 | 69.8 | done |
| s4_e3 | 75.1 | done |
| s4_e2 | 83.8 | done |
| s4_e1 | 93.7 | done |
| s4_e0 | 92.6 | done |
| s3_e3 | 76.6 | done |
| s3_e2 | 86.1 | done |
| s3_e1 | 94.9 | done |
| s3_e0 | 92.4 | done |
| s2_e2 | 94.8 | done |
| s2_e1 | 97.2 | done |
| s2_e0 | 98.1 | done |
| s1_e1 | 89.6 | done |
| s1_e0 | 50.2 | done |
| s0_e0 | 49.6 | done |
| s0_e1 | 49.7 | done |
| s0_e2 | 91.0 | done |
| s0_e3 | 86.8 | done |
| s0_e4 | 89.9 | done |
| s0_e5 | 82.4 | done |
| s1_e2 | 99.6 | done |
| s1_e3 | 96.5 | done |
| s1_e4 | 98.7 | done |
| s1_e5 | 94.9 | done |
| s2_e3 | 91.7 | done |
| s2_e4 | 90.3 | done |
| s2_e5 | 89.1 | done |
| s3_e4 | 72.1 | done |
| s3_e5 | 71.3 | done |
| s4_e5 | 68.4 | done |
| s5_e4_blk | 68.8 | done |
| s5_e3_blk | 74.1 | done |
| s5_e2_blk | 82.3 | done |
| s5_e1_blk | 95.7 | done |
| s5_e0_blk | 94.9 | done |
| s4_e3_blk | 75.0 | done |
| s4_e2_blk | 84.0 | done |
| s4_e1_blk | 96.8 | done |
| s4_e0_blk | 98.2 | done |
| s3_e2_blk | 85.1 | done |
| s3_e1_blk | 95.5 | done |
| s3_e0_blk | 93.9 | done |
| s2_e1_blk | 98.0 | done |
| s2_e0_blk | 99.3 | done |
| s1_e0_blk | 59.5 | done |
| s0_e1_blk | 83.0 | done |
| s0_e2_blk | 97.8 | done |
| s0_e3_blk | 93.7 | done |
| s0_e4_blk | 90.6 | done |
| s0_e5_blk | 91.9 | done |
| s1_e2_blk | 98.4 | done |
| s1_e3_blk | 96.0 | done |
| s1_e4_blk | 91.7 | done |
| s1_e5_blk | 96.3 | done |
| s2_e3_blk | 96.5 | done |
| s2_e4_blk | 96.5 | done |
| s2_e5_blk | 96.5 | done |
| s3_e4_blk | 76.2 | done |
| s3_e5_blk | 76.2 | done |
| s4_e5_blk | 68.8 | done |