Skip to content

Prospective IBL feedback expansion v0.2

Status: stopped at the frozen readiness gate (26 July 2026)

Outcome-blind cohort lock

The frozen protocol required at least 12 eligible animals beyond development animals fip_13–fip_16 before loading fluorescence outcomes or calculating feedback contrasts.

The reproducible query found 546 public sessions from 22 animals:

  • 139 sessions from the four development animals, excluded as specified;
  • 407 sessions from 18 nominally new animals;
  • zero eligible held-out sessions and zero eligible held-out animals.

Every new session contained labelled 470-nm GCaMP rows and wavelength-0 rows named None, but no labelled 415-nm isosbestic rows. The package's versioned IBL adapter requires a 415-nm reference for the nine frozen reference-correction universes, so all 407 new sessions received the same machine-readable exclusion reason: adapter_schema_failure.

The complete session/lab/subject inventory, exclusions, query environment, and fingerprint are in ibl-feedback-cohort-v0.2.json. It was committed before any dF/F or feedback contrast was calculated.

Why wavelength 0 was not treated as a reference

IBL's extraction source maps wavelength 0 to None / “No additional signal” and maps 415 nm separately to Isosbestic. The public loading guide likewise instructs users to select frames by the recorded wavelength. Although ROI columns contain finite values on wavelength-0 rows, neither metadata nor official semantics identify those values as a control channel. Relabelling them after the gate failure would be an outcome-adjacent schema amendment and could manufacture eligibility.

Sources:

Decision and interpretation

The held-out estimate, confidence interval, multiverse, and specification curve do not exist for v0.2. This is a readiness failure, not a null scientific result. The apparent increase from four to 22 public animals did not increase the eligible population for the frozen paired-signal analysis.

Any analysis of the 18 new animals requires a v0.3 protocol with one of two forms:

  1. provenance establishes that another stored field is genuinely the isosbestic measurement, permitting an updated adapter; or
  2. a declared signal-only estimand uses experimental control-free preprocessing, clearly separated from the paired-reference v0.1/v0.2 analysis.

The second option cannot be described as a held-out replication of the original nine-universe analysis because it changes every correction workflow and its identifiability assumptions.

The subsequent channel-provenance audit found that the study methods explicitly describe 470-nm-only acquisition and that the producing extractor maps wavelength 0 to no-LED frames. The first route is therefore closed on current evidence. A signal-only v0.3 design is being developed separately and has not yet been frozen or executed.