Decisio v0.9.0 on Qwen3.6-35B-A3B (official FP8, self-hosted)v1.7 · 10 s rule

Composite (secondary)

9.4 · Composite rank #166

Cost and measurement conditions

$0.179 per 1,000 decisions (estimated)

Price: ESTIMATE, labelled 'base_model Qwen/Qwen3.6-35B-A3B (served: Qwen's official FP8 release at 95a723d0)'. The frozen pricing_v15 module (sha256 645fd82eac728d597048bec0995ecd3ac896566e269fb6c9a1fdb84c171a9ac4, re-frozen 26 Oct 2026) resolves the declared non-FP8 base string Qwen/Qwen3.6-35B-A3B to qwen/qwen3.6-35b-a3b at USD 0.15/1.00 per million input/output tokens. It refuses the served FP8 repository string as lacking an explicit market-listing decision, so the non-FP8 base string is the pricing key. No explicit reference or rate override is used. Estimated, not charged, not a quote from the author or any host. BILLING BASIS, billing decision: engine-processed 1,761,623 input tokens (decisio front-pads each request to its 1,056-token prefix-cache block; its own usage reports the unpadded 664,205). The engine-input total is the preregistered offline transcription of the padded prompts, including the chat template and Answer: prefill, with no prefix-cache discount; it is NOT a server-observed per-request engine count. This tag computes engine lengths internally but does not expose them in the frozen HTTP response or per-request logs; exposing them would require changing the author code or serve configuration, which was not done. The reported unpadded usage equals the offline usage count (0 mismatches), which does not independently establish the padded engine count. Difference between the two input totals: 1,097,418 tokens. Independent count-level review decomposes this as 1,075,144 front-pad tokens plus 22,274 template/prefill and token-boundary overhead tokens (15 on 1,484 answered requests and 14 on one). An offline, token-boundary-aware per-item audit of the front pad on the frozen kit (no new measurement) found a residual of 0 on all 1,485 answered items. These checks reconstruct the offline count and do not observe server engine tokens. The evaluator accepted the offline-transcription billing basis after the server-count limitation was established. The 15 bare-500 over-length refusals add no tokens. OUTPUT: decisio reports 1 per answered decision, 1,485 total; no text is generated, but the full USD 1.00/M output tariff applies to those reported tokens and no zero-output convention is invoked. On the common 1,485-item basis the engine total gives USD 0.178942 per 1,000 decisions, 2.77x the unchanged USD 0.06459465517241379 cap; separately, USD 0.177152 per 1,000 over all 1,500 rows, 2.74x the cap. The same input and output totals apply on both bases because the 15 excluded long items are exactly the 15 refusals. OVER THE OFFICIAL COST CAP, DISCLOSED: the row is included (v1.6.16 vllm-sr-d3 precedent) and falls outside the official-caps board views by construction.

x2 + 0.15 s (assumption, not measured)

Evaluator-owned disposable rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB, FIN-03 (Finland, EU), PAYG non-spot, instance type 1RTXPRO6000.30V, image 24.04.cuda13.0. Measured 26 Oct 2026, approximately 21:40-22:00Z, approximately USD 0.9 GPU rental, separate from the benchmark's estimated model tariff. Author server at loopback 127.0.0.1:8000, network cut before sealed input, offline model-hash proof, no credentials, read-only source/model/input, private output. Pre-sealed smoke included the cold first request at the 32,767-token boundary (1.043543 s), accepted and refused length probes and configuration gates; the ready gate preceded the scored process's pre-warm. TTL fit admitted the run with 5,606 s remaining against 1,529 s required (compute estimate 887 s; forecast over 10 s: 0). Frozen input sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b (v1.6-regular-20261012). Complete 1,500-row run; scored loop 91.306016 s. Raw output sha256 9e72102fef97c070a00f88d6cd22b3fe6de99f725d535df3e338c8d20d0a16b1 (evaluator-verified). Completion receipt confirms teardown and node/local shred verified. One measured attempt after five create calls failed before any instance existed.