v0.2.0

Our Notes BoxLens v0.2.0 recognises member and Snap list screenshots with seven neural models running on CPU through ONNX Runtime. Source code and model weights use MIT; third-party notices are retained. The repository and attachments contain no game client assets, screenshots, fonts, master-data exports, or ready-made card gallery.

What changed

• Card-frame locator finds member and Snap tiles on the whole screenshot, including tiles cut by the screen edge.
• Per-kind artwork encoders match each tile against a gallery built locally from your own artwork. Identities below the similarity and margin thresholds are returned in the new unidentified list with a review candidate instead of a guess. New cards need only updated master data, artwork and prepare; the models stay unchanged.
• Per-kind field readers and card-rank classifiers read level, training count and card rank, and emit a value only when the crop is fully inside the image and passes its threshold.
• models/recognition.json pins all seven models by SHA-256 with their thresholds. Observations use schema v2 and the HTTP response is ournotes-boxlens.scan-result/2.

Included

• Source archive and Python CPU wheel.
• ournotes-boxlens-v0.2.0-weights.zip: models/ (seven ONNX files and recognition.json, ready to copy into DATA/models), matching PyTorch state dicts, model card, model-selection record and a hash manifest.
• Scene generator, training, export, selection and evaluation code for every model; training recipes with seeds; schemas and evaluation summaries.

Evaluation

Models and thresholds were frozen on development sets, then scored on two synthetic final sets of 2,000 full screenshots each, with screen resolutions and compression settings not used in training:

• Identity: 96.80% and 97.02% of gallery cards identified, with no wrong identity; the rest are reported as unidentified. Identities whose artwork never entered encoder training: 91.9% and 93.0%, no wrong identity. Foreign artwork was accepted as a gallery card in 1.32% and 0.92% of cases.
• Fields on gallery cards: visible level 95.9% / 96.2%, training count 96.0% / 96.0%, card rank 99.6% / 99.6%, with no misread in either set.
• Real-sample regression: eight screenshots from one collection gave 129/129 identities and all visible training counts and ranks correct. This is a regression set, not a representative real-world test.

Tiles cut by the screen edge or covered by overlays are mostly reported as unidentified, and small, low-resolution or heavily compressed cards have lower field coverage. Scroll cards into view and recapture, then review unidentified tiles and unknown values. Synthetic and single-user results do not establish general accuracy across players, devices, or future client versions. Details and strata are in docs/EVALUATION.md.

Verify the attached source, weights, and wheel using SHA256SUMS.txt.

Our Notes BoxLens v0.1.0

Our Notes BoxLens v0.1.0 is the first public release. Source code and model weights use MIT; third-party notices are retained. The repository and attachments contain no game client assets, screenshots, fonts, master-data exports, or ready-made card gallery.

Included

• Source archive and Python CPU wheel.
• Encoder and field model weights in ONNX and PyTorch formats, plus verified training initializers and hash manifests.
• Training recipe, frozen code required to trace the released checkpoint lineage, schemas, evaluation summaries, and CPU benchmarks.

Evaluation

• Clear card-rank matrix: all 122 catalog identities at ranks 1–5, across four resolutions; 2,440 observations, 2,440 identities and ranks correct. This isolates rank and fixes level to Lv.1. Rank is read by sprite-template matching, not a separate neural classifier.
• Training coverage: scene data contains all levels 1–90; numeric field crops cover 1–100. The five rank counts are 7,208 / 7,089 / 7,371 / 7,357 / 7,372.
• Independent synthetic full-screen tests: 560 normal and 280 stress screenshots across five languages and multiple parameter views. Emitted field values had zero errors in these sets, while rejected and unlocalized fields remain; the stress set has lower coverage. Detailed counts and strata are in the repository reports.
• Real-sample regression: ten screenshots from one collection gave 156/156 identities, 20/21 readable levels, 120/120 visible ranks, and 18/18 training counts. This is calibration data, not a representative real-world test.
• CPU scan benchmark: on Windows 10 / Intel i5-12400F, median 1.82 s with two logical CPUs and 1.66 s with four for ten images; P95 was 2.43 s and 2.90 s respectively. Other machine load was not controlled.

Clear, fully visible list screenshots are the supported target. Occlusion, overlays, fade effects, and heavy compression can cause omissions or misreads; reposition and recapture, then review or manually correct uncertain values. Synthetic and single-user regression results do not establish general accuracy across players, devices, or future client versions.

Verify the attached source, weights, and wheel using SHA256SUMS.txt.