needle-point-v10

Needle Point V10

This is the owner-approved interim production release of Needle Point V10. It is a self-contained Needle Point .np artifact for CPU PyTorch inference. The public model repository contains the model package and release manifests only; the sealed evaluation corpus is stored separately in the private dataset repository named in the release record.

Runtime policy

• PyTorch on CPU only
• FP32 weights and manual attention
• strict constrained decoding
• fail-closed fallback policy (fallbacks=[])
• batch-one inference
• intra-op CPU threads exactly 1, 2, 3, or 4; canonical/default 4
• inter-op CPU threads fixed at 1

The quarantined native C++/HIP and ROCm paths are not production options or fallbacks for this release.

Artifact and provenance

• artifact: automaticity-v10.np
• artifact bytes: 108,589,056
• artifact SHA-256: 13966196b2efff5ca6bf5cefc19127dfad56e261f937e7f9684f8f12a6639672
• package-build manifest SHA-256: 20aa1326200fa9c0d3c63f7cd83c9988c1be5fcdd29002bd6a540e8a2a1ddfc4
• V10 package manifest SHA-256: e5c84a61fa1034299133a0164c6900baba6fb5d2df26599e94d2844de3e350ee
• artifact checksum sidecar SHA-256: 69687b7267860f9e9d4475d78a06fc52ef9159baae125de8a3d8ee03d4bd4d92
• private V10 corpus manifest identity: 3d7d275cde7fd6f9d3d6e2960fe318d2f2b168212ae35f18c80fe1faeb3cd8c2
• private V10 corpus manifest bytes SHA-256: f78efbf78d2ac9bec41cd2ef29c03f1c71c454e44d39003272721ce8d9315285
• public Needle base-model SHA-256: c5f9a3016e4537e492c362da5cb8ba05107d8595bec0d5ea5d8a65801db46531
• tokenizer SHA-256: 0823f5b9133c68a8140addc5d7a425fa9119c4c8cb4a550363b4bffa4ba1c8c7
• upstream training commit: ffb1c5144c5a16cb8ec650dbc8a6f6fd3854f8f2

The complete provenance and immutable Hub revision are recorded in the repository's docs/V10_RELEASE.md.

Evaluation status

This is an interim release and is not a claim that the mandatory gates passed. The one-shot V10 FINAL-BENCH result recorded:

• schema-valid: 1,050/1,050
• action full semantic exact: 268/312 (85.897%)
• action tool exact: 294/312 (94.231%)
• no-op exact: 733/738 (99.322%), failed against the required 100%
• no-op overcalls: 5, failed against the required 0
• total misses: 49 (44 action and 5 no-op)

The frozen target-host latency matrix also failed its raw <500 ms p50/p95/p99 tails. At canonical four CPU threads the raw warm tails were 344.667 / 1,733.279 / 1,849.877 ms (p50/p95/p99). The owner-approved latency exception is latency-only and does not rewrite these failures. The one-shot FINAL sealed-run tails were 416.886 / 1,980.465 / 3,051.189 ms and are kept as a separate identity, not mixed with the matrix.

Next-version work

Future work must improve no-op discrimination and action misses, reduce CPU PyTorch prefill and decode tail latency, and complete real FP32/upstream-JAX parity and production concurrency validation. A later release requires a newly authorized sealed evaluation identity; this model card does not authorize or claim a rerun.