0005 — Interpretation is scored and multi-candidate — never one "detected format"

Status: Accepted

Context

A single raw timestamp value is usually underdetermined: the same integer is a plausible reading under many formats. Returning one "detected format" would present a guess as a verdict — the opposite of forensic rigor.

Decision

Interpretation is first-class and scored. interpret_int returns all plausible readings as Candidates, each with a score, named components, and assumptions. It never returns a single detected format.

• Scoring components are emitted verbatim (a named set), never collapsed to a bare rank: representable, in_window, granularity_match, magnitude_fit, epoch_distance, not_sentinel always; and, when context is supplied, byte_width_match, endian_match, artifact_match, neighbour_monotonicity. epoch_distance is a low-weight MAGNITUDE/RECENCY prior: a real timestamp sits well past the format's own epoch, so a reading hugging the epoch (a value too small for that unit, e.g. a 13-digit Unix-ms read as ns-since-2001 → 2001 plus minutes) is demoted. It re-orders only; it never hides a reading.
• Epistemic stance: a reading is evidence, not a verdict. Output says "consistent with", never "is" / "detected". A leap-smear disclaimer is surfaced where relevant (a smeared source is indistinguishable from a raw value without clock-policy metadata).

Consequences

• The default output stays ranked candidates; the zero-context result is unchanged when no width/endian/artifact/neighbour context is available.
• The registry must carry the evidence metadata scoring depends on (citations, tz and leap semantics, plausibility windows).