SDR-0017: Couple preprocessing outputs and separate incompatible units
- Status: Accepted
- Date: 2026-07-27
- Decision owners: project maintainers
- Related contract: Multiverse scientific contract v0.1
Context
Signal-only preprocessing can produce divisive ΔF/F or subtractive acquired- fluorescence outcomes. Normalization therefore changes both the ordered processing operations and the variable summarized around events. Independent decision nodes would create invalid combinations, while pooling their magnitudes would imply that incompatible units share a scale.
Decision
Represent preprocessing operations and their event-summary output as one atomic typed multiverse choice. Apply response-window choices first and the coupled output choice last, so another node cannot restore an incompatible variable.
Partition scientist-facing robustness reports by output units. Do not display a pooled median or range across lanes. Reject a project-wide practical-effect threshold when preprocessing alternatives span units. Require signal-only project recipes to remain within a signal-only primary analysis and materialize operations in the order resampling, filtering, baseline estimation.
Consequences
Divisive and subtractive workflows can inhabit one transparent robustness analysis without being numerically conflated. The low-level result retains every universe, but interpretation of magnitude must occur within report lanes. A future schema is needed for per-lane practical-effect thresholds and summaries.
Alternatives considered
- Cross preprocessing and output-variable nodes: rejected because most generated combinations would be structurally invalid rather than scientific alternatives.
- Convert subtractive values to ΔF/F after execution: rejected because that changes the declared transformation and can conceal baseline assumptions.
- Forbid mixed normalization: rejected because robustness to normalization is a substantive methodological question.
Revisit trigger
Add typed, per-unit-lane effect thresholds and machine-readable lane summaries before presenting practical-effect stability for mixed-normalization multiverses.
Evidence added later
On 2026-07-27, the revisit trigger was met with typed per-lane thresholds and the
normative multiverse_lane_summary artifact. Denominator and serialization policy
are governed by SDR-0018.