Skip to content

SDR-0011: Regularize irregular clocks prospectively

  • Status: Accepted
  • Date: 2026-07-26

Context

AsLS is defined over equally spaced samples, while real IBL photometry timestamps contain small acquisition jitter. The v0.3.2 analysis correctly retained those universes as structurally incompatible because no regularization policy had been declared before outcome access. Silently treating row order as a regular clock or adding interpolation after seeing effects would obscure a consequential choice.

Decision

Irregular-clock workflows may explicitly declare linear regularization before a regular-sampling operation. The prospective default candidate is:

  • target rate: reciprocal of the session's median source interval;
  • gap boundary: do not interpolate across an interval greater than 1.5 times the median source interval;
  • original timestamps and values retained on source_time;
  • Boolean acquisition masks transferred by nearest neighbour and forced false inside protected gaps;
  • interpolation distance, source jitter, resolved rate, and gap-masked fraction recorded in provenance.

Compatibility preflight uses timestamps and the declared operations only. Signal fidelity benchmarks are a separate outcome-based validation lane.

Consequences

Adding regularization makes an AsLS pipeline mechanically executable, but does not make it scientifically validated. Results must identify the interpolation policy, and robustness reports must distinguish raw-compatible and regularized workflows. The completed v0.3.2 result is not amended or rerun under this policy.

Revisit trigger

Revisit before promotion if target-rate inflation is requested, protected gaps remove material event coverage, interpolation distance approaches the event timescale, or validation on sharp transients and held-out recordings fails.

Evidence added later

The checksum-frozen 24-session held-out IBL comparison made AsLS mechanically executable and retained all 5,920 event windows. Raw-versus-regularized event deltas were highly stable, but the aggregate protocol failed its strict whole-trace normalized-RMSE, interpolation-distance, and finite-coverage gates. Maximum source distance reached 0.525 median samples, and AsLS baseline magnitude remained materially comparator-dependent despite near-identical short-window event deltas. Regularization therefore remains explicit and experimental rather than a silent or recommended default. See the held-out report.