Control-free bleaching benchmark v0.1

The first signal-only baseline benchmark is a retained partial failure. Its protocol and thresholds were frozen in benchmarks/protocol-control-free-v0.1.md before aggregate execution. The complete 240-run result is benchmarks/control-free-v0.1.json.

Results

Scenario	Method	Correlation	RMSE	Event bias	Pass
Single exponential	Double exponential	0.806	0.00310	-0.25%	no
Single exponential	AsLS	0.772	0.00593	+0.27%	no
Double exponential	Double exponential	0.797	0.00318	-0.24%	no
Double exponential	AsLS	0.510	0.00879	+0.24%	no
Large transients	Double exponential	0.989	0.00363	-0.23%	yes
Large transients	AsLS	0.954	0.00817	+0.36%	yes
Slow drift	Double exponential	0.184	0.02117	-0.08%	no
Slow drift	AsLS	0.583	0.00798	+0.40%	no
Motion without control	Double exponential	0.142	0.02846	-36.85%	limitation
Motion without control	AsLS	0.132	0.04857	-34.44%	limitation
Event-locked artefact	Double exponential	0.951	0.00618	+132.85%	limitation
Event-locked artefact	AsLS	0.823	0.01095	+134.12%	limitation

Values are medians over 20 seeds. “Pass” requires correlation ≥0.90, RMSE ≤0.015 dF/F and absolute event-amplitude bias ≤20%.

Interpretation

Both estimators preserve event amplitude exceptionally well in the four bleaching scenarios. The frozen samplewise-correlation threshold nevertheless rejects the three small-transient cases because the corrected trace still contains acquisition noise: baseline correction is not denoising. This reveals that v0.1 partly conflates two jobs. The threshold remains unchanged in the frozen artifact.

The limitation cases are more important. Slow baseline estimation cannot remove oscillatory motion without another source of information. Worse, an event-locked artefact can produce a very high correlation with the true neural waveform while inflating event amplitude by about 133%. Correlation alone is therefore not a validity diagnostic.

Double-exponential fitting is a defensible compromise between under- and over-flexible bleaching models, but subtraction versus division encodes different assumptions about autofluorescence and indicator bleaching (Simpson et al., 2024). AsLS originated as a smooth lower-envelope method for spectroscopy; its transfer to photometry is experimental and its smoothness parameter is sampling-rate dependent (Eilers & Boelens, 2005).

Decision

The two methods remain available as explicitly experimental preprocessing APIs, with full provenance and signal retention. They are not yet added to the typed multiverse pipeline. A v0.2 benchmark should separate baseline error from residual high-frequency noise, test subtraction and division as distinct estimands, and vary sampling rate before either method is promoted.

The governing judgment and its alternatives are recorded separately in SDR-0002, so a future promotion or reversal cannot silently replace the reasoning available at v0.1.