Seen 0.14 adds the bounded host and accelerator contracts required before the Qwen3.8-27B engine can be implemented. These are infrastructure APIs, model locks, and certification gates; the model engine itself is not part of 0.14.
inference/scalars defines stable C-layout storage representations for Float16, BFloat16, Float8E4M3, and Float8E5M2, plus checked 64-bit tensor geometry. The existing fixed-width integer names remain canonical. inference/span provides checked borrowed spans and bounded 2/3/4/8-bit packed views. memory/aligned_buffer owns 16–256-byte-aligned storage and closes idempotently. The matching public C declarations are in seen_runtime/seen_inference.h.
memory/mapped_file owns read-only files and bounded mapped windows with offsets above 4 GiB. A file cannot close while windows remain active. Window validation detects truncation; sequential, random, prefetch, huge-page, page-lock, and Linux NUMA requests return stable results rather than silently falling back. Unsupported advisory controls are explicit.
crypto/sha256 is incremental and reads files in bounded chunks. Digests use canonical lowercase hexadecimal and can be parsed for fixed-work comparison. json/strict rejects invalid UTF-8, duplicate names, unpaired surrogates, limit violations, and non-canonical or overflowing Int64 text. json/canonical deterministically orders object names by UTF-16 code units for canonical manifest output.
formats/safetensors validates the little-endian header length, bounded metadata, exact geometry, dtypes, non-overlap, local shard-index paths, and deterministic name ordering. Tensor byte windows remain zero-copy and are invalidated by explicit owner closure.
The CUDA subsystem is separate from Vulkan and is built only with -DSEEN_ENABLE_CUDA=ON. Disabled configuration stops before CUDA language or SDK discovery, so ordinary CPU builds do not locate or link NVIDIA libraries. The initial backend is Linux x86-64, targets compute capability 8.9, and has experimental-hardware maturity.
The versioned seen_cuda_* ABI owns only device/resource adaptation. Seen owns allocation bounds, move-only lifetime, fallback policy, algorithm identity, and diagnostics. It covers devices, device and pinned memory, asynchronous copies, streams, events, graph capture/replay/update, and deterministic cuBLASLt F16/BF16 algorithm selection. Silent fallback is not supported.
Model-specific native kernels borrow the exact Seen-owned stream through the fixed-width seen-cuda-stream-launch-token-v1 contract. A token contains the native cudaStream_t value, device ordinal, capture flags, and a generation; it is valid only for one immediately nested adapter call while its move-only CudaStream owner remains open. Borrowing validates the opaque handle and expected device without allocation, fallback, host blocking, or CUDA synchronization. The borrower may enqueue capture-compatible work but may not retain the token, destroy or synchronize the stream, change its device, or make scheduling policy. Closed, stale, malformed, and cross-device requests return typed CUDA statuses before the native value is exposed.
projects/seen_ml/qwen38 contains the strict model/source lock contracts, exact configuration validator, tensor-role classifier, tokenizer asset pins, content-addressed artifact promotion, and environment capture. The official model revision and every required file are immutable inputs. The concrete source lock also pins the Qwen and Transformers references and the initial llama.cpp/vLLM comparison revisions. Remote model code is never executed.
Benchmark evidence requires at least 30 measured samples, correctness and quality state, host and VRAM usage, transfers, effective bandwidth, qualified hardware conditions, and the shared five-percent kernel regression gate. Generated model assets and machine evidence remain outside Git.
Release acceptance is serial and runs inside the repository's current-memory derived, swap-disabled hard scope. It includes a 60 GiB sparse mapping and truncation test, hostile JSON/Safetensors cases, native-boundary inventory, CPU-only dependency scanning, and real RTX 4090 memory, stream, event, graph, F16, and BF16 cuBLASLt execution. Compile-only evidence cannot replace the hardware result.