drogna

A demonstration harness. Every number in it is invented.

V1 record — describes retired software; superseded for V2 by srd.md, constitution 2.0.0 and ADR-0027.

ADR-0002: Decorrelation timescale is a field, authored per feature

Status: Accepted Date: 26 August 2026 Requirement: SRD FR-05 (v0.3); consumed by FR-08, FR-32, FR-34 Supersedes: the provisional per-feature position taken in SRD v0.2

Context

Something must govern how fast a location loses memory of a measurement. That quantity — the decorrelation timescale, tau — drives the revisit cadence during loiter (FR-08) and is an input to every cell the planner scores (FR-32, FR-34).

SRD v0.2 left the question open and started with tau as a property of each seeded feature. That was recorded as an open question rather than a decision, and it did not survive contact with the requirements that consume it.

Three formulations were available.

Per-feature. Each of the four seeded features of FR-03 carries its own timescale. Simple to author, simple to record in the ground-truth manifest, and it matches the intuition that an eddy decorrelates faster than open water. But it leaves the background water with no timescale at all — and FR-08 explicitly requires that quiet water be left alone, which is a statement about the background, not about the features. The planner would have to invent a value for every cell outside a feature, and an invented value is not ground truth and cannot be scored.

Per-region. A static map over the domain, giving every location a timescale including the background. This fixes the gap, and it is defensible physically. But FR-03 requires a feature of known drift velocity, and a static map cannot follow it: the fast-decorrelating patch would stay put while the feature it describes moves away, so the revisit pattern of FR-08 would track the wrong water.

A field, authored per feature. tau(latitude, longitude, depth, time), evaluated at every location, where the value is a domain-wide background blended with the contribution of any feature overlapping that location, and a moving feature's timescale advects with the feature.

Decision

The decorrelation timescale is a field, tau(latitude, longitude, depth, time). It is authored per feature over a domain-wide background value, and evaluated per location. A moving feature's timescale advects with it. Both the background value and the per-feature timescales are ground truth and are recorded in the FR-04 manifest.

Authoring stays per feature, so a scenario is still written the way a person thinks about it: a background, and four features that disturb it. Evaluation is per location, so every consumer gets a defined answer everywhere.

Consequences