ADR-0006 — Recovery likelihood is not assessment confidence

Status: accepted · Date: 2026-09

Context

Consumer tools show a single colour. That conflates "how likely is this file intact?" with "how much does the tool actually know?".

Decision

RecoveryHealth carries two independent values in the range 0–100:

likelihood — the estimated probability that recovery yields the original content;
confidence — the quality of the evidence behind that estimate.

Both are produced by a deterministic model of hard constraints plus weighted evidence, and every score is accompanied by concrete, machine-generated reasons. Cluster reallocation is described as "allocated to active data", never as "definitely overwritten", because $Bitmap proves the former only.

Consequences

"Likelihood 38 / confidence 96" and "likelihood 75 / confidence 31" are distinct, meaningful outcomes.
Thresholds and weights are provisional and must be calibrated empirically against a controlled corpus; the UI labels them as estimates until then.