Model Profiles · Simulation
One exact model setup per task type
The first release supports one embedding profile. Every device runs the same model, tokenizer, runtime and precision, so results can be checked against a reference.
Short English text embedding
demo-profile/minilm-l6-v2/cpu-fp32/r0
Fictional demo identifier. Not a production profile hash.
- Task type
- Text embedding
- Candidate model
- sentence-transformers/all-MiniLM-L6-v2
- Licence
- Apache 2.0 (as listed by the publisher; license and conversion review pending)
- Runtime
- ONNX Runtime Android, CPU, FP32 (proposed)
- Output
- 384 dimensions, L2-normalised
- Max tokens per record
- 256 (longer records are rejected, never truncated)
- Records per task
- up to 64
- Task payload
- up to 256 KiB
Acceptance tolerances (proposed, §9.1)
- cosine ≥ 0.9999
- max error ≤ 0.001
- norm 0.999–1.001
Not in the first release
- Text classificationDeferred
Needs approved labels and a fixed scoring method (§4.2).
- OCRDeferred
Needs document-level quality criteria and separate privacy testing (§1, §4.2).
- Short audio transcriptionDeferred
Needs language, audio and error-rate criteria (§4.2).
Why one profile
Verification compares a device's vectors with a trusted reference. That only works when every device runs the exact same model, tokenizer, runtime and precision. A changed part means a new profile ID and a new round of calibration, so results from different setups are never mixed.