AJev lm6q (Gemma 4 31B)v1.7 · 10 s rule

Composite (secondary)

43.7 · Composite rank #55

Cost and measurement conditions

$0.093 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen on 22 Oct 2026 with launch-list keys added and nothing else changed; the rates for this base string are identical in the earlier module 81e3578f): base string google/gemma-4-31B-it (the card's base_model) priced as google/gemma-4-31b-it USD 0.09/M input, USD 0.34/M output. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the usage block of our shim, which mirrors the author's server (input_tokens = length of the rendered prompt token ids incl. system message, options, repeated user text and chat template; output_tokens = 0, as the author's server reports with thinking off, although vLLM samples and discards one token per request: 1,485 completion tokens in total, kept in our per-row audit log, not billed). Per row, shim usage = vLLM prompt_tokens = our offline count through the same shim and the pinned tokenizer: 1,530,069 input tokens over the 1,485 answered items, and 15 refusals at the pre-registered ids and positions (HTTP 422, 0 tokens). OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.092731 per 1,000 decisions on the common basis (USD 0.091804 over all 1,500 items) is 1.436x the official cap USD 0.064595; the row is ranked (v1.6.16 vllm-sr-d3 and v1.6.24 imajev-9b precedent) and falls outside the official-caps board views by construction. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them); their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence.

x2 + 0.15 s (assumption, not measured)

Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB (driver 580.178.04), FIN-03 (Finland, EU), pay-as-you-go non-spot, instance type 1RTXPRO6000.30V at USD 2.549/h. Stock vLLM 0.30.0 and our shim (mounted read-only, sha re-checked on the node) in ONE docker container with network none (the cut proven from inside at start and again by docker exec from outside: loopback only, no default route, DNS and TCP fail), read-only root, tmpfs scratch, only the output directory writable, env -i with every credential-named variable empty, HF offline flags, the sealed input not present until the smoke and the time gate had passed (input sha256 dfd36720... checked on the node and inside the container); one request at a time, timed over loopback on the VM. Measured 22 Oct 2026 (VM reserved 09:44Z, scored run 09:55-10:01Z, released 10:01Z, about USD 0.76) on the frozen v1.6-regular-20261012 self-hosted input; node shredded and VM removed afterwards (teardown confirmed).

Model and serving disclosure

NO AUTHOR CODE WAS EXECUTED OR IMPORTED. The Hub repository holds a rank-64 LoRA adapter (safetensors + configs only, no .py, no auto_map): andyzhang232/ajev-gemma4-31b-lm6q at bdcfb05db02f48729f5f5a0b6b1c8cf6d6406d06 (Apache-2.0 per card, not merged), applied directly by stock vLLM 0.30.0 (image vllm/vllm-openai:v0.30.0@sha256:8a69ffad015f138d7170c4ddc429e230a3bc1c1719f67e14324749df200a4b90, no derived image, no overlay wheels) on google/gemma-4-31B-it at 842da3794eaa0b77d5f08bae87a17459d91ff475. The request/response layer is OUR first-party shim ajev_systemone.py (sha256 e849fcdbaf0740f3a22b61108053572b3a76d3b35030aa39ed6b00f560bcbb9a), a rule-by-rule port of the author's published inference package (cmzy/ajev-infer at e2e723aba95a1f15d51966f43b53430372f49a4d, read as text only; the unit tests compare the author's constants by ast, never exec): request normalisation, prompt v3 (v2 layout with system message for choice questions of up to 26 options, v1 layout otherwise, user text repeated once when it is at most 12,000 characters), Gemma-4 chat template with thinking off, no truncation, lettered options A-Z then tokenizer-verified two-letter codes up to 255, readout over both single-token forms of each label combined by logsumexp, softmax over labels at T=1 (the adapter stores no temperatures). One /v1/completions call per decision (max_tokens 1, temperature 0, allowed_token_ids = all label forms, logprobs = their count). Server flags: --enable-lora --max-lora-rank 64 --max-loras 1 --lora-modules ajev-lm6q=<adapter snapshot> --dtype bfloat16 --max-model-len 40960 --gpu-memory-utilization 0.88 --enable-prefix-caching --max-logprobs 512 --logprobs-mode processed_logprobs --generation-config vllm --max-num-seqs 16 --max-num-batched-tokens 8192. Review: independent reviews of the shim and the runner preceded the run; their findings were fixed and re-reviewed, and the shim hash above is the one that ran. Weights: both snapshots downloaded revision-pinned and every file sha256/inventory-verified against MODEL_MANIFEST.json (sha256 dade7dfa4333fd15ef79c5a1682e4c3a3a6fd79a45035b251d6704a0492ad086): base 12 files, 62,578,686,256 bytes; adapter 5 files, 1,959,002,147 bytes; verdict OK for both. Smoke before any item: a separate smoke-only Verda pod (passed on attempt 12) and an in-container synthetic probe on the scored VM; no test item was ever used in a smoke. FIRST MEASUREMENT (found by our own intake: Decision Index 0.4.1 entry). SERVED BY OUR SHIM, NOT BY THE AUTHOR'S SERVER: the row measures the author's documented vLLM direct-LoRA serving path as re-implemented by us (stock vLLM 0.30.0 + our shim), not necessarily the configuration behind the author's self-reported 65.08 (the card says only that it was self-run with the official kit); there is no numeric parity check against the author's code (running it is not allowed here), only a rule-by-rule port plus tests. DISCLOSED DEVIATIONS from the author's server: D1 score probabilities are a dict keyed by level (the adapter needs a dict; the author returns a list), rounded to 4 decimals like the author; D2 a missing label token is an error instead of the author's -1e4 fill (cannot change an answer here, all label ids are requested); D3 requests go through vLLM's HTTP /v1/completions route instead of in-process generation; D4 an upstream 'maximum context length' 400 would be mapped to 422 (audited separately; never happened); D5 batch parameters --max-num-batched-tokens 8192 --max-num-seqs 16 are our pins, because the author's engine most likely runs 2,496 / 128 (source-derived, not verified live) and an explicit 2,048 is rejected by vLLM 0.30.0 for Gemma-4; this changes chunk boundaries and batch shape only; the context limit 40,960 follows the author's inference README for this 31B submission (the author's server default is 131,072); output_tokens is reported as 0 like the author. OPEN PARITY RISK (not tested): the adapter has v_proj LoRA on 50 of 60 layers and none on the 10 global layers where K and V share one projection; vLLM's fused qkv_proj loads K into both slots, so a K delta may not reach V as it would under Hugging Face PEFT. The smoke proved the adapter is really applied (LoRA-vs-base total variation up to 0.95 on a choice case) and that our readout equals an independent recomputation from raw upstream logprobs; during the smoke attempts the probe-only tolerances were loosened by us (a looser bound for cases with more than 26 options, 1e-6 to 1e-3, then one gross-error bound of 0.03 / TV 0.10) after float-level noise, the shim stayed byte-identical and no item was involved. CONTEXT: our shim answers HTTP 422 above 40,960 tokens (no truncation); the 15 items of 76,562-81,737 tokens were refused exactly as pre-registered by id and position and are scored wrong; the largest answered item is 3,595 tokens. DETERMINISM: over the first 50 positions, two passes, 6/50 were bit-identical, 0 argmax flips, max probability difference 0.029 (vLLM batching and prefix-cache bf16 noise); the probabilities need not equal the author's or the DI lab's. TRAINING DATA, the author's own statement: the card names no training datasets and warns that some training data carries non-commercial or attribution terms ('check them yourself before commercial use'; not a conclusion about the weights); the training-base revision is undeclared (PEFT revision null), so our base pin is a serving pin, not provenance. The author states it ran the full public Decision Index 0.3 suite (140,620 requests) with the official kit and compared adapters on it; the author's code comments describe benchmark-guided prompt-format changes (our static triage: development exposure, no item-specific rule found); no JevBench use or denial is declared; not independently verified. SELF-REPORTED, unverified: public Decision Index 65.08 (the DI kit rescore is 65.46; the DI lab's Decision Index 0.4.1 full score is 67.57, decision agreement with the submitted rows 99.80% on 831 rows against a 98.5% bar), latency median 139 ms on one RTX PRO 6000. 10 s RULE: max measured decision latency 0.81 s; 0 decisions over 10 s.