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

schedopt_bo_r8

20260722_142849_schedopt_bo_r8 · complete · published 2026-07-23 · seed 1000

part of investigation alpha-curriculum → schedule-opt → schedopt-v2

Intent

Intent: Round 8 of the Phase-4 search over the modulation family \@{schedopt-bo-r8-intent} Step E of the schedopt re-freeze continuation (Phase 4), round 8: eight candidate schedules from the four-parameter modulation family, the three-parameter ramp multiplied by (1 + A·w) with amplitude A in [0, 1] and w pinned to burst_16's realized tooth wave (period 400/3 steps, duty 0.5, peak-first), proposed by Thompson sampling from a Gaussian process warm-started from every three-parameter evaluation re-expressed at A = 0 and scored at the re-frozen operating point v2 (m_star=95.5, k=350, dwell 2) over the shared search seed pool. The pre-registration (experiments/schedopt/phase4_prereg.md) precedes every run of this phase; v1 and v2 numbers are never merged.

Hypothesis

Hypothesis: The tooth amplitude helps at a speed-measuring objective \@{schedopt-bo-r8-hypothesis} At least one modulated candidate beats the best smooth schedule on the search pool, and the running optimum's amplitude A is greater than zero. Search-phase s-hat values are never reported as results.

Setup

model: GPT (nanoGPT-derived), weight-tied embeddings — 2 layers × 4 heads × 256 dim (~1,635,584 params) · block 50 · dropout 0.1 · vocab 167 tokens
training: AdamW (β 0.9/0.95, wd 0.1) · lr 0.0006 constant (no warmup, no decay) · batch 32 · 350 iterations · fresh init per replicate (seeds derived from base seed) · 25 seeds (base 1000)
training data: PCFG, zipfian noun–verb pairing (oneshot where noted) · 12,000 sentences per dataset (9,600 train — 1 epoch = 300 iters) · unseen holdout 10
code: commit 13bd2ea93ea8

data regime (all arms): fresh-dataset-per-stage — 14 dataset(s) × 25 iters each = 0.02 epochs per dataset

Replicate this experiment

Get the code at this exact version:

git clone https://github.com/ClaireHobbs/imagining-syntax
cd imagining-syntax
git checkout 13bd2ea93ea8
pip install -e ".[dev]"

Download the config (experiment.yaml) and run:

imsyn exp run experiment.yaml

Checks

PASS children_complete PASS replicate_coverage PASS schedule_total_matches_final_waypoint PASS instrument_consistent PASS objective_computable

Results — objective: P(unseen_mismatch ≥ 95.5 by t = 350, dwell 2, smoothing 1)

arm	ŝ	wilson95	median crossing [CI]	crossed
r8_c0	0.920	[0.750, 0.978]	227.4 [222.1, 256.4]	23/25
r8_c3	0.920	[0.750, 0.978]	224.0 [221.1, 228.6]	23/25
r8_c5	0.880	[0.700, 0.958]	276.9 [220.1, 298.5]	22/25
r8_c1	0.760	[0.566, 0.885]	221.5 [215.2, 289.5]	19/25
r8_c2	0.760	[0.566, 0.885]	221.6 [214.9, 283.5]	19/25
r8_c4	0.080	[0.022, 0.250]	265.4 [234.0, 296.8]	2/25
r8_c6	0.080	[0.022, 0.250]	217.9 [217.9, 218.0]	2/25
r8_c7	0.040	[0.007, 0.195]	148.4 [148.4, 148.4]	1/25

Trajectory summary — unseen_mismatch

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)
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 = 350)	flags
r8_c0	96.7 ±9.0
r8_c1	89.3 ±18.0	⚠ seed_split: range 43.6-100.0 (3 low / 17 high of 25)
r8_c2	96.5 ±8.2
r8_c3	99.8 ±0.3
r8_c4	44.6 ±26.6	⚠ seed_split: range 2.4-98.5 (4 low / 2 high of 25)
r8_c5	99.7 ±0.6
r8_c6	46.6 ±25.5	⚠ seed_split: range 0.0-98.3 (3 low / 3 high of 25)
r8_c7	91.6 ±2.3

Conclusions

Conclusion: Teeth on the smooth winner's shape break the plateau \@{schedopt-bo-r8-conclusion} Round 8 breaks the two-round plateau at 0.84 and sharpens the picture of what the amplitude buys: the best schedules are now toothed versions of the smooth winner's shape, a mid-high ramp that dives to the floor only at the end. Result: Late-diving toothed descents reach 0.92 \@{schedopt-bo-r8-toothed-descent} r8_c0 (ramp α 2.61 to 0.10 with late curvature γ 3.07, amplitude A = 0.43) and r8_c3 (3.42 to 0.16, γ 3.65, A = 0.69) each crossed on 0.92 of the pool seeds, Wilson interval 0.75 to 0.98, an improvement of 0.08 over the round-7 plateau; r8_c5 (3.19 to 4.44, γ 3.00, A = 0.98) reached 0.88. Both winners hold their teeth inside the productive range for most of the training budget and reach the floor only in the last segments. The failures repeat the round-7 pattern from both directions: teeth whose valleys sit near deep-uniform exposure late in the run (r8_c6, ramp 1.42 to 0.44 at A = 0.84, crossed 0.08) and shapes that move high early and stay there (r8_c7, 3.46 to 4.32 with early curvature, 0.04). With an improvement above 0.02 the pre-registered stop rule does not fire, and round 9, the pre-registered cap, is the last either way. These are search-phase values on the v2 objective, not reportable results.

Comparison figures

Children

Child	peak unseen_mismatch	Status
r8_c0	96.8	done
r8_c1	95.8	done
r8_c2	96.5	done
r8_c3	99.8	done
r8_c4	55.2	done
r8_c5	99.7	done
r8_c6	62.3	done
r8_c7	91.6	done

experiment.yaml