Bev v0.1.2 — measured RAM and portable runtime

Bev v0.1.2

Memory first: the 7.21 GB GGUF reached 7.28 GiB peak resident system RAM in a short CPU-only inference with one 4,096-token slot on Stallion. The earlier 8.30 GiB number is an observed GPU-memory snapshot for the separate Linux CUDA benchmark setup with two 16,384-token slots; it is not host RAM or a measured peak. Allow additional headroom.

Bev brings Jevfire-style one-token decision scoring to Prism ML's unchanged Ternary-Bonsai-2-27B PQ2_0 model. This release packages the inference service and reproducible configuration; it does not introduce trained or fine-tuned weights.

• Typed decisions: /v1/decisions for flat boolean/enum schemas; /v1/systemone for Choice, Noul and Score. Original option identities and rubric descriptions are preserved. Score is the probability-weighted rubric position.
• Complete candidate scores: a small Prism runtime patch returns selected raw token log probabilities while retaining full-vocabulary normalization. Missing candidates and oversized contexts fail explicitly.
• Auditable service: startup-captured source hashes, pinned model/runtime files, and a benchmark wrapper that checks source stability across evaluation.
• Validation: 46 Bev tests passed and 7 optional-dependency tests were skipped on macOS; 14/14 real Metal API checks passed. The v0.1.1 Linux CUDA run previously passed 53 Bev/package tests and 66 upstream Persian benchmark tests. The full Persian run completed 624/624 valid answers, with 229/240 Choice, 152/160 Noul, and 70/80 Score within ±0.5.
• Portable build: the same pinned Prism source and scoring patch can now be built through scripts/install-portable.py for macOS Metal or Linux/Windows CPU. 14/14 API smoke checks passed on an Apple M2 Mac; Windows remains an untested source-build path. The Linux CUDA path remains available.
• Measured footprint: 7,206,168,928-byte GGUF; 7,630,416 KiB CPU process high-water RSS on Stallion and 8,504 MiB observed GPU memory in the original CUDA setup. Persian benchmark median latency was 2.136 seconds per request, with six questions in main/repeat requests.

Candidate probabilities are relative preferences, not calibrated correctness guarantees. Fields are evaluated independently. Quality trails the published Jev reference on this Persian suite, and the earlier 120-request general diagnostic shows mixed results. No controlled full-precision comparison or ternary speedup claim is made.

See benchmark results and limitations, aggregate Persian results, and provenance. Benchmark texts, labels and upstream scorer code are not bundled.