allec-oavg (plantwaterer; commercially restricted)v1.7 · 10 s rule

Composite (secondary)

74.0 · Not ranked on this board

Cost and measurement conditions

$0.030 per 1,000 decisions (estimated)

GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 735,102 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0295320334008779/1,000; old list-price estimate USD 0.0444319543317663/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0295320334008779/1,000. Estimated, not a bill. USD 0.0295320334 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0295061687 over all 1,500 items).

p50 latency: 0.4 s

Self-hosted latency estimate under the shared method (assumption, not measured)

Measured 30 Oct 2026 on an evaluator-owned Verda VM in FIN-03 (Finland, EU), one NVIDIA RTX PRO 6000 Blackwell Server Edition, nominal 96 GB, image 24.04.cuda13.0. THE AUTHOR'S OWN MODEL AND DECISION LIBRARY, with a first-party loopback System One wire shim: plantwaterer/allec at e54027aa1f865bdd25724e9b3c1acb51b2bbc3c3. Library-only submission, with no author HTTP server. The shim calls the author's unchanged README entry point LenAvgBayesEngine(model=local_model, device=cuda) from lenavg_bayes_engine; it does not substitute the older single-pass DikitEngine. Full Gemma-4-31B-derived weights and trained decision head are measured together. Native defaults remain BF16 backbone, CPU FP64 L01-post head, bidirectional global jev_mem attention and causal sliding layers, max_len 32,768, tokens_per_batch 32,768, l_star 939 and averaging enabled. Short-request routing uses structural characters, not token counts. Native four-order choice averaging for menus of at most 52 options, two-order noul averaging, deterministic author temperatures and Bayesian head correction are retained. Menus above 52 retain native chunk scoring, model-selected finalists and the final round, including total 0.001 probability mass on discarded final options. Requests are serialized because native forward hooks and averaging flags are mutable. Native probabilities are transported unchanged; no generated text, host retry, answer repair, stock causal generation, quantization or stock-vLLM equivalence. Score criteria are unsupported: the shim returns native-domain HTTP 422 before the author's length helper can hit .items() on score-list criteria, without score-to-choice conversion. Native context/state-shown refusals also return HTTP 422. No truncation, raised context cap, split state or fallback inference. The native renderer's crop followed by check_state_shown is transcribed as a refusal before forwarding. The first-party observer additionally refuses any alternative forward missing the full normalized state or exceeding 32,768 tokens; this is a disclosed no-truncation transport adaptation. Structured state uses native JSON conversion, preserving native structural-length routing on the original state. USAGE IS FIRST-PARTY OBSERVED ACCOUNTING: numeric input_tokens sums the tokenizer lengths of every actually forwarded prompt, including averaged option orders, packed sequences and second rounds; it is expanded prefill, not a single read and not an author-native billing contract. Output_tokens is 0 because no continuation is generated. The frozen TypeSafe adapter and measurement driver remain unchanged, including the 120 s transport timeout. Decision Index base.py is pinned at 9eb2dbe2a358004c8782c66e40a83ac07b953fec; no Decision Index suite or training code runs. Runtime receipts retain source, package RECORD and tokenizer hashes; torch 2.14.1, CUDA 13.0, transformers 5.12.1, tokenizers 0.22.2, numpy 2.2.6, safetensors 0.8.0 and accelerate 1.15.0. Source review found no benchmark-item answer lookup in the serving closure; framework internals were not exhaustively audited. Author code and weights ran only on the disposable rented pod, with verified local artifacts, read-only mounts, empty credentials, offline flags and network cut before sealed input. Separate public warm-ups preceded three fresh timed repetitions each at 256 and 2,048 tokens and two at 4,096. Native typed probes passed, including 255-choice, Unicode JSON and structural-routing boundaries. Unsupported score and over-limit probes returned HTTP 422. These are synthetic observations, not scored latency aggregates. Separate public pre-warm passed. The complete 1,500-request run took 208.734721660614 s; raw output sha256 185c8f534d2e8b21d38a39af9a3152f0f4754c41bf921128bb4d2b768698aee7 was verified. Post-run health was ready. The VM was closed and temporary node and local copies deleted, with no cleanup errors reported. GPU rental spend is separate from the benchmark's model-cost estimate.