Skip to content

Four-round adaptive synchrony experiment

Experimental result: HOLD. This verb is public for inspectable method experiments, but it is not the preferred production neighborhood selector. The current real-data gate did not support promoting a common first-break radius into a general adaptive rule. Use empirical_synchrony_decay() when the question is how synchrony changes continuously with distance.

v.adaptive_synchrony_experiment(...) separates a maximum observation domain from an estimated empirical synchrony scale. It consumes the reusable sparse relationships produced by v.local_synchrony_pairs(...); it does not recompute the underlying time-series correlation.

The four rounds are:

  1. Estimate independent cold and warm first local-to-regional breaks inside a standardized physical-distance domain.
  2. When both breaks are valid, form one configurable common radius and reduce cold, warm, and pairwise Delta over the same neighbors.
  3. Compare those adaptive reductions with a fixed-radius control from the same relationships.
  4. Retain the cold, warm, common, cold-minus-warm distance, uncertainty, and categorical status fields as scientific outputs.

The default common rule is max(cold_break, warm_break). A discovery limit is never substituted for an unresolved break. Primary Delta remains the robust reduction of pairwise cold_synchrony - warm_synchrony, not a difference of marginal medians.

Sampling contract

v.local_synchrony_pairs(...) accepts an optional distance_sampling plan. Successive (upper_km, maximum_pairs_per_focal) strata are sampled uniformly before the expensive synchrony kernel runs; None retains a dense stratum. Every retained relationship carries sampling_probability, and the adaptive reducer uses its inverse as a design weight. Validate sampled break placement and status against exhaustive focal pixels before spatial scale-up.

Current real-data gate

The bounded 2023-11-01 through 2024-01-30 PRISM experiment evaluated five exhaustive 1000 km focal domains. Exhaustive common breaks resolved at two of five representative pixels; the distance-stratified version resolved one of five, and tail-status agreement was 30%. Fixed 500 km pairwise Delta was much more stable under sampling (mean absolute error 0.00065). The predeclared representative gate therefore stopped the workflow before the 25-site Colorado, statewide Colorado, or CONUS stages. This is an implementation and evidence boundary, not a missing-value imputation opportunity.

Temporal extension

Outputs retain time_window_id and time_window_end. Repeating the bounded method over independent windows can later support median, IQR, and trend of R*(x,y,t). That extension should wait until spatial break identifiability and far-field sampling stability pass their gates.