Evidence

OpenRemap explains why it believes something, not just what it believes. Every layer of the pipeline carries evidence that you can inspect:

Layer	Evidence example
Detection	`SIZE_MATCH`, `MAGIC_MATCH`, `IDENT_BLOCK`, `FAMILY_STRING` — what made an extractor claim a binary
Identification	confidence tier + signal list (each with its point delta)
Checksums	per-entry verdicts: `ok` / `stale` / `disabled`, with stored vs expected values
Maps	structural score, classifier label + probability (`fuel 0.72`), layout region
Health	per-check verdicts (`ok` / `warn` / `fail` / `skip`) with details
VINs	WMI whitelist, ISO check digit, year, numeric tail, mirror count, ident context

The principle: a verdict without evidence is a guess. Anything probabilistic must say how confident it is and what supports it.

Detection evidence

When an extractor's can_handle() decides a binary belongs to a family, it records why as evidence tags:

Tag	Meaning
`SIZE_MATCH`	File size matches a known family size
`MAGIC_MATCH`	Header / descriptor magic bytes match
`HEADER_MATCH`	A known header magic matches
`POINTER_TABLE`	A pointer table with plausible targets was found
`LAYOUT_FINGERPRINT`	The block layout matches the family
`BOOT_BLOCK`	The boot block carries the expected shape
`IDENT_BLOCK`	Identity block found at the expected offset
`FAMILY_STRING`	Family-specific ASCII string found (e.g. `Copr.DENSO`)
`FAMILY_ANCHOR`	A family-specific anchor string was found
`DETECTION_SIGNATURE`	A declared detection signature matched
`SYNC_MARKER`	A known sync-marker sequence was found
`MANUFACTURER_CONFIRM`	The manufacturer's own string confirms the file
`EXCLUSION_CLEAR`	No conflicting family signatures found
`FILL_PATTERN`	The fill-byte ratio matches the family's fill style

An extractor may also emit free-form tags of its own (for example MOTRONIC_M2_LABEL or TSW_STRING); those carry the default weight.

Tags are weighted — structural ones (magic bytes, pointer tables, layout fingerprints, ident blocks) are worth 4; identity/detection tags (FAMILY_STRING, SYNC_MARKER, …) 3; soft ones (SIZE_MATCH, EXCLUSION_CLEAR, FILL_PATTERN) 2; any free-form tag 2. The total is capped at +20, so a long tag list cannot inflate the score. See Confidence scoring for the table.

Confidence signals

score_identity turns the identity dict plus the detection evidence into a score and a tier (High / Medium / Low / Suspicious / Unknown). Each contributing factor is a signal with a point delta:

  Tier   MEDIUM
  Signal  +  detection evidence (4 tags: SIZE_MATCH, MAGIC_MATCH, POINTER_TABLE, BOOT_BLOCK)
  Signal  +  canonical SW version
  Signal  +  calibration ID present (86CAU_AT)
  Signal  -  tuning/modification keywords in filename

The terminal prints the sign only; the point values are in the JSON output and in the one-line summary above the section (see Confidence scoring).

Negative signals matter as much as positive ones — a filename saying sport or stage1 costs points, and the warning is shown.

Probabilistic labels

Where a real name cannot be proven (no A2L/DAMOS), labels stay probabilistic: the map classifier emits fuel 0.72, never a bare claim. The number says how well the structure matches the label's profile; the label is a suggestion, not a fact.

Provenance

Every result records where its information came from:

Identity fields — extracted from the binary at known offsets (the extractor states them).
Checksums — from community-documented algorithms (RomRaider, NefMoto, IronFelix, EcuFlash defs), validated against factory files.
Map labels — inferred structurally; an A2L/DAMOS source would upgrade them, never be replaced silently.
Verdicts — always explainable by the evidence list above.

Where evidence is exposed

Command	Where to look
`identify`	the confidence section (tier, score, signals, warnings)
`scan`	per-file confidence in the table and reports
`checksum`	per-scheme and per-entry statuses with stored/expected values
`health`	per-check verdicts with details
`scan-maps --classify`	`label` + `label_confidence` per table
`scan-vins`	per-candidate evidence flags and confidence

→ Confidence scoring — how tiers and signals are computed → How it works — the pipeline that produces the evidence