jerboa-ssd-recognizer

jerboa-ssd-recognizer is the Jerboa-owned recognizer for Star Fleet Battles SSD review data. It replaces the decision-making parts of the Python truth/OCR service with reusable .ss code.

The library does not implement an OCR engine. It consumes OCR words from a backend through jerboa-vision:

((text . "PH-1") (conf . 88) (bbox . (x1 y1 x2 y2)))

From there it owns the parts that must be identical on the server and APK:

• box-type aliases and canonical labels
• OCR text-to-box candidates
• learned truth examples from approved groups
• negative examples from deleted/suppressed groups
• race gating, such as PA Panels for Andromedans only
• feature scoring by count, position, size, aspect, and optional color
• compact truth lookup and source/page/dpi matching

Build

make test

By default the Makefile expects ../jerboa-vision beside this checkout. Override with JERBOA_VISION=/path/to/jerboa-vision if needed.

API Shape

The main entry point is:

(ssd-recognize-groups payload)

Expected payload fields:

• source: source metadata with name, path, optional source_key
• image: rendered image width and height
• groups: detected SSD groups with id, bbox, count, cell_ids
• ocr_words: backend OCR words in rendered image coordinates
• learned_examples: examples generated from approved truth files
• box_types: SFBOL box type id/name list
• inventory: optional ship inventory counts from SFBOL/SFC data

The result contains suggestions, each with group_id, label, box_type_id, confidence, reason, source, ocr_text, and candidate alternatives.

Truth service logic lives in ssd-recognizer/truth-store.ss. The important entry point is:

(ssd-matching-truth-entries entries params)

It implements the Python service behavior for exact source-key matches, source-name/page/dpi fallback matches, compact metadata, and sorting more complete truth files first.

Migration Target

The current Python service can be retired in stages:

1. Keep Python only as an OCR backend that returns words.
2. Move /api/recognize scoring to this library.
3. Move truth load/save/dump to the Jerboa SSD web service.
4. Generate Android/Kotlin or Jerboa-Android code from this same data model.
5. Replace the OCR backend with Android ML Kit for fully offline APK use.