M-value/PLS trajectory specificity pilot
Run date: 2026-07-29 to 2026-07-30
Code revision: 7cffff1 (main)
Configuration: examples/trajectory_power_study/pilot_50x199.json
Purpose
This pilot tested the production PLS trajectory pipeline after correcting methylation integration to convert B values back to M values before standardisation and latent-space construction. The primary question was whether the previously observed magnitude-to-orientation cross-talk remained after that correction.
This is a medium-sized diagnostic run, not the paper-grade study. It used 50 replicates per cell, so a rejection rate near 0.05 has a Monte Carlo standard error of approximately 0.03.
Run configuration
| Parameter | Value |
|---|---|
| Integration | PLS, using M-value methylation |
| Samples per replicate | 300 |
| Stages | 4 |
| Replicates per cell | 50 |
| Permutations per test | 199 |
| Effect sizes | 0.00, 0.25, 0.50, 0.75, 1.00 |
| Trajectory modes | magnitude, orientation, shape, translation |
| Significance level | 0.05 |
| Total evaluations | 1,100 |
| Failed evaluations | 0 |
Primary result
The M-value correction removed the targeted magnitude cross-talk. At the largest magnitude effect, the intended delta test had complete power, while the angle and shape rejection rates remained below the nominal 0.05 level.
| Injected mode | Delta | Angle | Shape |
|---|---|---|---|
| Magnitude | 1.00 | 0.02 | 0.02 |
| Orientation | 0.72 | 0.76 | 1.00 |
| Shape | 1.00 | 0.84 | 1.00 |
| Translation | 0.02 | 0.04 | 0.08 |
Values are rejection rates at effect size 1.00, with 50 replicates per entry. Bold entries are the statistic intended to detect that mode. Translation is a negative control.
Magnitude specificity was stable across the entire effect-size sweep:
| Effect size | Delta | Angle | Shape |
|---|---|---|---|
| 0.00 | 0.04 | 0.02 | 0.08 |
| 0.25 | 1.00 | 0.02 | 0.02 |
| 0.50 | 1.00 | 0.00 | 0.02 |
| 0.75 | 1.00 | 0.00 | 0.00 |
| 1.00 | 1.00 | 0.02 | 0.02 |
Type I error and controls
The ordinary null cell was consistent with the nominal 0.05 level:
| Baseline | Delta | Angle | Shape | Combined rule |
|---|---|---|---|---|
| No difference | 0.04 | 0.02 | 0.06 | 0.10 |
| Translation control | 0.06 | 0.02 | 0.10 | 0.14 |
All preregistered per-statistic Type I checks passed their two-Monte-Carlo-standard-error tolerance. The combined rule is descriptive here and was not among those individual Type I acceptance checks.
Acceptance results
- Passed: all six Type I checks for the ordinary and translation baselines.
- Passed: magnitude/delta power and monotonicity; top-effect power was 1.00.
- Passed: shape/shape power and monotonicity; top-effect power was 1.00.
- Failed: orientation/angle power target and monotonicity; rates across increasing effects were 0.00, 0.60, 0.66, 0.56, and 0.76.
- Passed: all translation specificity checks and both magnitude off-diagonal checks.
- Failed: orientation-to-delta, orientation-to-shape, shape-to-delta, and shape-to-angle specificity checks.
Interpretation
The pilot supports the rung-ladder conclusion that magnitude-to-angle leakage was caused by input representation rather than the trajectory estimator. Converting methylation to M-value space before PLS integration restores the expected separation: magnitude strongly affects delta without elevating angle or shape rejection.
The study does not establish general specificity of all three statistics. Orientation and shape injections still move multiple statistics. Those failures may reflect the definitions of the feature-space manipulations rather than a defect in the estimators: a global feature permutation can alter Procrustes shape as well as orientation, and relocating stage-specific feature support can change step lengths and angles as well as shape. This requires a separate geometry audit before interpreting those cells as estimator cross-talk.
The orientation curve is also non-monotone and misses the preregistered 0.80 top-effect power threshold (observed 0.76). With only 50 replicates, the standard error at 0.76 is approximately 0.06, but the strong off-diagonal rejection rates are too large to attribute to Monte Carlo noise.
Decision and next work
The targeted magnitude finding is strong enough to retain M-value conversion as the production contract. Before launching the 500-replicate paper-grade grid, audit orientation and shape injections in the pre-integration generative frame and verify that each is a pure change in its intended geometric quantity. Then repeat this pilot with corrected manipulations. Running the full study now would make the remaining cross-talk estimates more precise without resolving their interpretation.
Reproduction
From the repository root, run four resumable shards:
for i in 0 1 2 3; do
uv run python scripts/run_study_shard.py \
--config examples/trajectory_power_study/pilot_50x199.json \
--out-dir results/mvalue-pls-pilot-50x199 \
--shard-index "$i" --n-shards 4 --n-jobs 1 \
--error-policy record
done
Then merge and report:
uv run python scripts/motco_study.py merge \
--out-dir results/mvalue-pls-pilot-50x199
uv run python scripts/motco_study.py report \
--config examples/trajectory_power_study/pilot_50x199.json \
--out-dir results/mvalue-pls-pilot-50x199