[cobanov]

fp16 tags for the Qwen3.5 decoders, a written model lifecycle, and every site number from results/.

Runtime and models (ADR 0008)

• Every Qwen3.5 graph now uses Split instead of SplitToSequence, which ONNX Runtime ran on the CPU. The values are unchanged: all 14 fp32 tags re-ran parity with the same numbers (results/runs/2026-10-19-parity-rtx4090-cuda.json).
• Six new -fp16 tags: kev:4b-fp16, kev:9b-fp16, nimble:9b-fp16, clef:flash-fp16, pplx-decider:flash-fp16 and strands-decider:2b-fp16. Each has its own gate: the encodings exact and at least 99.5 % of decisions the same as the fp32 reference. All pass, at 99.58–100 %. kev:0.8b-fp16 passed too, but it is not shipped because the fp32 graph is as fast.
• ollaya run --ask asks your own questions from the command line.
• The ONNX Runtime arena shrinks after each run.
• decision's upstream org was renamed to vllm-sr.

Process (ADR 0009, docs/model-lifecycle.md)

• One runbook covers every stage from intake to live: what each stage needs, its commands, the machine and a time budget.
• A rolling bench edition: a release measures only what it changed, and bench_plan decides what that is from the manifests and the parity records.
• New tools: parity_run, bench_plan, bench_run (resumable), bench_edition --update and scripts/bump-version.sh.
• The site, the library pages, the docs and the README read their numbers from results/ ({{result <tag> <path> <digits>}}, result(), generated README blocks). The build fails on a missing result or a stale README.
• Ollama 0.40.6's runs are result files. /results has a table with both servers.

Measured (edition 2026.10, updated by 0.14.0)

Measured for 0.14.0 (results/runs/2026-10-19-*): parity on the RTX 4090; typed-decisions, the median five-question
request over HTTP and the GPU memory on the RTX 5090 of Ollaya Bench 2026.10. The fp32 typed-decisions numbers are the
edition's; the fp32 graphs' outputs did not change.

Tag	Decisions the same as fp32 (parity, RTX 4090)	Typed-decisions, fp16 / fp32	Five questions, RTX 5090: fp16 / fp32	GPU memory, fp16 / fp32
kev:4b-fp16	478 / 480 (99.58 %)	0.669 / 0.669	98 / 171 ms	10.8 / 10.9 GB
kev:9b-fp16	480 / 480 (100.00 %)	0.723 / 0.722	125 / 248 ms	17.6 / 17.9 GB
nimble:9b-fp16	492 / 492 (100.00 %)	0.665 / 0.665	428 / 1015 ms	19.7 / 20.1 GB
clef:flash-fp16	570 / 571 (99.82 %)	0.703 / 0.703	150 / 268 ms	19.4 / 19.6 GB
pplx-decider:flash-fp16	695 / 696 (99.86 %)	0.726 / 0.726	90 / 126 ms	10.3 / 10.4 GB
strands-decider:2b-fp16	649 / 650 (99.85 %)	0.615 / 0.616	72 / 98 ms	5.8 / 8.5 GB
kev:0.8b-fp16 (not shipped)	478 / 480 (99.58 %)	0.461 / 0.460	67 / 62 ms	3.6 / 4.5 GB

Typed-decisions accuracy is within 0.002 of the fp32 tag's for every fp16 tag. kev:0.8b-fp16 passed its gate but is not
shipped: once Split replaced SplitToSequence, the fp32 graph of the 0.8B model was as fast.

Split alone made every fp32 graph faster, with the same parity numbers (median five-question request on the RTX 5090):

Tag	0.13.0	0.14.0
decider:0.8b	189 ms	60 ms
decider:2b	193 ms	91 ms
decider:4b	393 ms	207 ms
decision:eos	209 ms	69 ms
kev:0.8b	142 ms	62 ms
kev:4b	319 ms	171 ms
kev:9b	375 ms	248 ms
clef:flash	406 ms	268 ms
jeeves:9b	641 ms	380 ms
nimble:9b	1688 ms	1015 ms

Checks

• cargo fmt --check, clippy -D warnings and cargo test --workspace pass on Linux.
• python3 -m unittest discover -s convert/tests passes.
• npm run build passes.

The registry manifests are not in this PR: they go in as "Registry: …" commits at release time.

🤖 Generated with Claude Code