DANDI 000351 timestamp-aligned parity audit v0.2
Status: completed; exact parity failed (26 July 2026)
Question and provenance
Can the archived percentage dF/F in DANDI:000351 be reproduced by fitting its raw 405-nm control to raw 470-nm fluorescence, after explicitly aligning the different recorded clocks?
The v0.2 protocol was frozen after the v0.1 structural failure and before numerical comparison. It retained the same four checksum-pinned draft assets and changed only the declared alignment rule. The study identifies these data with Mesolimbic dopamine release conveys causal associations and links its analysis code through the ANCCR repository. The inspected public repository and checksum-verified Zenodo v1.0 archive contain downstream analysis but no raw-to-archived dF/F conversion implementation. Their bundled extension schema explicitly describes raw timestamps as not yet behavior-synchronized and processed dF/F timestamps as synchronized, corroborating the v0.1 mismatch without specifying how the transformation was performed.
Results
All assets passed integrity checks, all sessions had a non-empty shared time domain, and every candidate executed.
| Method | Sessions | Median correlation | Minimum correlation | Median RMSE | Maximum RMSE | Exact parity |
|---|---|---|---|---|---|---|
| OLS | 4 | 0.591 | 0.128 | 0.813 pp | 3.266 pp | Fail |
| Huber IRLS | 4 | 0.590 | 0.127 | 0.813 pp | 3.295 pp | Fail |
The two fits were nearly indistinguishable, so robust fitting alone does not explain the archived output. Session-level OLS correlations were 0.720, 0.560, 0.128, and 0.621. Only one session met the 0.50-percentage-point RMSE threshold; none met the correlation threshold. Between 0.024% and 0.580% of archived samples fell outside the shared finite time domain and were excluded without extrapolation.
Complete metrics are retained in the
JSON result, generated by
run_dandi_000351_parity_v2.py.
Interpretation
The archived dF/F is not reproduced by full-session affine 405-to-470 fitting plus linear clock alignment. Plausible missing choices include preprocessing filters, fit windows, resampling implementation, segment removal, or a different correction formula. Those are hypotheses, not findings, and were not tuned after observing the result.
This failure is valuable for library design: NWB co-location and matching labels do not guarantee raw-to-processed reproducibility. FiberPhotometry will preserve archived outputs as externally supplied derived data, record explicit alignment policies, and avoid presenting approximate parity as validation.
The Dandiset remains an unpublished mutable draft, so this audit pins assets by
digest and observation date rather than treating draft as a stable version.
Next investigation
- Search the archived Zenodo release and repository history for the missing conversion routine or parameter record.
- If primary provenance is recovered, freeze a v0.3 exact reconstruction before execution.
- Otherwise use this cohort for analysis robustness—not numerical ground truth— and proceed to the prospectively gated 18-animal IBL expansion.