Governance, security and compliance
The technical, governance and security specifications for SOW Task 4, set for a standalone system that is not connected to the Department. They are also designed to carry forward to FedRAMP authorisation if a later phase needs government hosting.
Model card
Technical specification: model architecture, data inputs and performance benchmarks.
- Encoder
- CLIP ViT-B/32 (OpenAI), 8-bit quantised · MIT licence
- Weights fingerprint
- 583fd1110a514667812fee7d684952aaf82a99b959760c8d7dca7e0ab9839299
- Method
- Nearest-neighbour retrieval over 411 reference embeddings, fused with zero-shot similarity to written category descriptions, then a temperature softmax over 13 designated categories plus an out-of-scope control. Parameters were fit by leave-one-out on the reference split.
- Inputs
- A single RGB photograph, resized to 224 px by the model's processor. Nothing else: no metadata, location or case data.
- Outputs
- Ranked categories with probabilities, an advisory risk band, the closest exemplars with sources, and a written rationale.
- Held-out performance
- Top-1 85.6%, top-3 95.0%, source country 90.6%, ECE 0.035 (n=139).
- Runtime
- ONNX Runtime WebAssembly in the browser, CPU only. No GPU, server, API key or network needed after the first load.
- Intended use
- Triage support for trained officers and investigators: flagging objects that may fall within a designated list, so they can be referred to a subject-matter expert.
- Not for
- Legal determinations, seizure decisions, provenance or authenticity findings, valuation, or clearing an object. A low score is not evidence that an object is unrestricted.
- Known limitations
- Studio-photograph corpus of 40 images or fewer per category; some categories use cross-border or stylistically close stand-ins (see the caveats); not tested on field imagery; forgeries and replicas are not modelled.
- Retraining path
- Replace or extend the corpus through Task 3 data-sharing agreements, rebuild the index, and re-run the benchmark. The same adapter contract accepts a fine-tuned model.