Skip to content

Control-free bleaching benchmark v0.2

All frozen v0.2 acceptance gates passed. The protocol was committed before aggregate execution in benchmarks/protocol-control-free-v0.2.md, and all 440 runs are retained in benchmarks/control-free-v0.2.json.

Baseline fidelity

Scenario Method Baseline relative RMSE at 10/20/40 Hz Pass all rates
Single exponential Double exponential 0.090% / 0.092% / 0.089% yes
Single exponential AsLS 0.507% / 0.510% / 0.511% yes
Double exponential Double exponential 0.089% / 0.091% / 0.089% yes
Double exponential AsLS 0.805% / 0.801% / 0.805% yes
Slow drift Double exponential 2.074% / 2.074% / 2.075% no
Slow drift AsLS 0.724% / 0.722% / 0.724% yes

Event-amplitude median absolute bias remained below 1.7% in every cell. Scaling the AsLS penalty by (rate / 20)^4 made its baseline error nearly invariant from 10 to 40 Hz. Full-trace correlations remain in the artifact but are descriptive: their lower values in small-transient cases still reflect residual acquisition noise.

The methods are complementary rather than interchangeable. A constrained double-exponential model is substantially more accurate when its bleaching shape is correct. Its failure on sinusoidal slow drift is structural. AsLS is less accurate on exponential baselines but flexible enough to pass the slow-drift case.

Subtraction versus division

Bleaching mechanism Division: late-vs-early change Subtraction: late-vs-early change Matched choice
Indicator fluorescence -0.2% -25.9% division
Additive autofluorescence +34.4% -0.4% subtraction

This is not a cosmetic normalization choice. Division yields dF/F and is stable when baseline and physiological amplitude scale together. Subtraction retains fluorescence units and is stable when the bleaching component is additive. The API semantics and alternatives are recorded in SDR-0003.

Decision

v0.2 resolves the metric defect identified in v0.1 and supports both experimental implementations under their stated assumptions. It does not satisfy the independent real-data-control requirement in SDR-0002, so neither method is promoted to the recommended typed pipeline yet.

The next validation should compare control-free baseline estimates against a real recording with an independent stable fluorophore or acquisition control, while keeping motion correction conceptually separate.