Held-out IBL regularized-AsLS comparison v0.1

Status: completed; frozen aggregate gate failed; no method promotion (26 July 2026).

Result

All 24 checksum-frozen held-out sessions loaded and all five declared method paths executed. The benchmark retained all 5,920 frozen event windows for every method. Nevertheless, the overall prospective gate failed, so regularized AsLS remains experimental.

The held-out sessions had been excluded from the earlier IBL v0.3 analysis solely because they lacked enough events in both feedback conditions. Their fluorescence values had not been loaded before the v0.1 protocol and exact source hashes were frozen.

Regularization fidelity

Double exponential and the published rolling baseline were applied both before and after regularization, isolating interpolation effects from baseline-method choice.

Raw versus regularized method	Median trace r	Minimum trace r	Median normalized trace RMSE	Event-delta r	Event normalized median change
Double exponential	0.9964	0.9620	0.1136	0.9990	0.0146
Published rolling	0.9973	0.9612	0.1029	0.9990	0.0143

Trace correlations and every event-level gate passed. The prospective median normalized trace-RMSE threshold of 0.10 failed narrowly for both comparators. This is not relabelled a pass because downstream event summaries were stable: the contract explicitly required both.

The maximum target-to-source distance was 0.525 median sample intervals, exceeding the frozen 0.25 threshold. The target grid begins at the first source sample, but small interval errors accumulate phase relative to later acquired samples; almost all regularized targets are therefore interpolated rather than exact timestamp matches. Median-rate regularization is mechanically sound here, but interpolation provenance remains scientifically relevant.

Missingness and coverage

Every method retained all 5,920 selected event windows. Whole-recording finite coverage was different:

Method	Minimum fitted-baseline fraction
Raw double exponential	97.46%
Regularized double exponential	97.46%
Regularized AsLS	97.46%
Raw published rolling	77.66%
Regularized published rolling	77.31%

The 99% frozen gate therefore failed. Double exponential and AsLS preserve manually excluded recording edges as missing. The rolling baseline additionally requires a complete centred 60-second window; one short session consequently had only about 77% fitted coverage. The event-selection intervals kept every analyzed event away from these regions. This distinction validates the product decision to report candidate, gated, and complete denominators separately from global trace coverage.

AsLS sensitivity

Regularized comparison	Median baseline r	Median baseline normalized RMSE	Event-delta r	Event normalized median difference	Contrast sign agreement*
AsLS vs double exponential	0.9776	2.109	0.9997	0.0140	100%
AsLS vs published rolling	0.9816	2.863	0.9996	0.0199	100%

*Only six sessions contained usable events in both conditions; these are descriptive session contrasts, not population inference.

AsLS and the comparators produce highly correlated slow baselines but materially different baseline magnitudes relative to within-trace scale. Those differences mostly cancel in the short baseline-versus-response event contrast. This supports AsLS as an explicit sensitivity universe, not as a validated default: agreement on one short-window estimand cannot establish baseline truth or generalize to slower outcomes.

Decision

Regularization successfully makes AsLS executable on real irregular clocks and preserves these selected short-window event summaries. It did not pass the frozen whole-trace fidelity and coverage contract, and AsLS baseline identity remains method-dependent. SDR-0002 and SDR-0011 therefore remain conservative: expose the method and provenance, retain it as experimental, and do not promote it into a recommended default.

The complete machine-readable result is ibl-regularized-asls-results-v0.1.json, with fingerprint ed8ec3dc071957093790599f95122e5c4b30a738c6334c06d92a232fde6b365a.