---

title: Pomodoro v0.8.0 — Low Validation lDDT Root Cause Analysis tags: [journal, pomodoro, edm, bug, sampling, lddt, v0.8] created: 2026-05-11 updated: 2026-05-11 status: active related:

• “pomodoro-v0.7-training-plan”
• “edm-pomodoro-algorithm-review”
• “edm-noise-exploration”

---

Pomodoro v0.8.0 — Low Validation lDDT Root Cause Analysis

Problem

Validation lDDT during EDM sampling was ~35, while the debug workspace reference achieved ~60. Training-step lDDT (single-step denoising) was unaffected.

Root Cause

sample_schedule had reversed interpolation direction — ascending (sigma_min → sigma_max) instead of descending (sigma_max → sigma_min). The sampler started from near-zero noise and progressively added noise, then jumped to 0.0 at the final padded step. This inverted the entire denoising trajectory.

Buggy code (all versions v0.6.0–v0.8.0):

# Ascending — WRONG
sigmas = (
    cfg.sigma_min**inv_rho + steps / (num_steps - 1) * (cfg.sigma_max**inv_rho - cfg.sigma_min**inv_rho)
) ** cfg.rho

Corrected code (matching debug reference):

# Descending — CORRECT
sigmas = (
    cfg.sigma_max**inv_rho + steps / (num_steps - 1) * (cfg.sigma_min**inv_rho - cfg.sigma_max**inv_rho)
) ** cfg.rho
sigmas = sigmas * cfg.sigma_data

Secondary Issue: Missing sigma_data Scaling

The debug reference (workspace/debug/edm/debug_training.py) multiplies the schedule by cfg.sigma_data (line 46). All production versions were missing this scaling. Added alongside the direction fix.

Why Training lDDT Was Unaffected

sample_schedule is only called by edm_sample(), which is only invoked in validate(). Training step() uses noise_distribution() for single-step denoising — no schedule involved.

Other Differences Reviewed (Not the Cause)

Parameter	Debug	v0.8.0	Impact
fragment_length	32	8	Affects context but not schedule bug
S (model width)	32	64	Wider model needs more training
r (stability noise)	1e-3	0.0	Could affect convergence
compute_lDDT arg order	Swapped in all versions	Swapped	Symmetric metric — doesn’t affect value

These may explain residual gaps after the fix, but the reversed schedule was the primary cause.

Fix Applied

Same commit pushed to all 10 branches:

fix: reverse EDM sample_schedule direction (sigma_max→sigma_min) and add sigma_data scaling

Branch	Commit
main	341497d
v0.6.0	c7e05c0
v0.6.1	6b7d733
v0.6.2	f43d68c
v0.6.3	e733e29
v0.7.0	8c3497e
v0.7.1	27c5eb1
v0.7.2	b1bcb25
v0.7.3	fabb922
v0.8.0	fe2703f

Remaining Open Items

• compute_lDDT argument order is swapped in all callers (X0 passed as predicted, X_hat as ground truth) — the mask M is computed from predicted distances instead of reference distances. Affects local selection behavior but not the reported metric value since |D0 - D| is symmetric.
• Fragment length difference (32 vs 8) may need further investigation for optimal lDDT.
• Stability noise r=0.0 in v0.8.0 vs r=1e-3 in debug — worth testing.