Skip to content

Evidence model

Every value the app shows carries its provenance:

Evidence(value, unit, confidence, source, raw)
  • value, unit: normalised per the field (volts, amps, watts, codes, connector sizes).
  • confidence: 0 to 1, from the model's structured output or from the matcher's rule.
  • source: where the value came from. photo is a value read from a whole photo; label_crop is a value read when the label was found first by SAM 3.1 and its crop went to the model with the photo. measurement is a dimension computed from marks on a photo with a scale reference. voice_note is the owner's spoken note, transcribed by Muse Voice Transcribe and kept as their own words. user_edit is a value the owner typed. listing_text is a value read from a pasted listing and screenshot one read from a screenshot of it, on a candidate rather than on an owned item. grounding_citation, video_frame and document arrive with grounding, video and cited documents.
  • raw: the text as it was read, before normalisation, so the user can check the clean-up.

The evidence table in the UI, the listing generator and the matcher all consume this one shape, so they work for any field set.

How a reading becomes evidence

  1. Muse Spark returns one entry per field it could read, with the value as printed, the unit and a confidence. Fields the photos do not support are left out.
  2. Entries whose name is not in the field set are dropped with a warning; blank values are dropped silently; a field read twice keeps the more confident reading.
  3. The field's normaliser cleans the value: 19,5 V becomes 19.5 with unit V, 3330mA becomes 3.33 A, P/N OP065-19-333 becomes OP065-19-333.
  4. Warnings are added when no part or model number was read and when any field is below 0.5 confidence. Warnings travel with the result; nothing is silently filled in.

What the model is shown

Before a label crop (or a whole photo, when the label was not found first) goes to the model, its contrast is stretched and its edges sharpened a little, so faint print reads better; the reading says so with enhanced. What is kept is always the photo as taken. The sharpest of the images read is scored (the variance of its Laplacian); under EXTRACT_BLUR_THRESHOLD the reading carries a warning that the label looks blurred or shows little detail, so the person retakes it closer and steadier rather than trusting a weak read.

Regions

A segmentation result is the other half of the evidence: one region per match with its concept (item, label, connector), its box on the prepared photo, the model's confidence, how much of the box the mask covers, the crop, and the mask outline. The overlay drawn over the photo is the same data rendered, so what the user sees is what the next step read.

Provenance in the inventory

Saved privately, each evidence row becomes an assertion on the item with a provenance: observed for a value read from a photo or a crop, owner_reported, documented, tested or estimated for values that arrive later. An observed value is proposed until the owner accepts or rejects it. A correction never replaces an observation: the owner's value is added beside it and becomes the current one, and both keep their source, their confidence and their time. The item's fields are the current value per field, chosen as the latest accepted assertion, else the latest proposed one; rejected ones never show but are never erased.