Jeffan-4B-2010v1.7 · 10 s rule

Composite (secondary)

41.1 · Composite rank #62

Cost and measurement conditions

$0.016 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen 22 Oct 2026): base string Qwen/Qwen3.5-4B priced as Qwen/Qwen3.5-4B USD 0.03/M input, USD 0.15/M output. The module selects the deprecated-after-cutoff DeepInfra reference retained under the frozen policy (price snapshots of 25 Sep 2026), USD 0.03/M input and USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the author engine through the shim: input_tokens includes the rendered prompt, chat template and options; output_tokens = 0 is our shim's labelled convention because the engine reports only input usage. Reported usage is 776,399 input tokens and 0 output tokens, equal to the preregistered offline single-read count, offset 0 and 0 item-level mismatches. The 15 refused items carry no usage and count 0 tokens. Those are the standard long items excluded from every row's common 1,485-item cost basis (v1.6.30), so common-basis input usage is also 776,399, with 0 output tokens. Measured arithmetic: USD 0.015684828 per 1,000 decisions on the common basis = 0.242819290x the unchanged USD 0.06459465517241379 cap; separately USD 0.015527980/1,000 over all 1,500 items = 0.240391097x. COST ON THE COMMON BASIS: USD 0.015685 per 1,000 decisions on 1,485 items, 0.243 times the official cost cap USD 0.064595; estimated, not charged.

x2 + 0.15 s (assumption, not measured)

Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB (97,887 MiB), FIN-03 (Finland, EU), PAYG non-spot, instance type 1RTXPRO6000.30V, USD 2.5741917808219177/h recorded by the runner. One VM for both rows back to back, 31b then 4b. Engine and loopback shim in a --network none --read-only container per row; serial requests, read-only code/model/input, blank credential variables, HF offline flags. In-container network cut proven before sealed input: only loopback, no default route, DNS and TCP fail; code inventory/hashes and model files checked. Attempt 2 measured 24 Oct 2026 07:59-08:36Z; this row's measurement container 94 s, run cap 183 s, complete=true, deadline_hit=false, rc 0; scored loop 50.604333 s over 1,500 requests. Frozen v1.6-regular-20261012 self-hosted input sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b, checked in the container. Raw output sha256 e32cbcf090bccf4750a0225b30d9cd73aa0636be787f9f517d3e97946691121a. Row cleanup verified; final exact-id VM teardown confirms provider absent and guard closed at 08:36:06Z. GPU rental is not the benchmark's estimated model tariff.

Model and serving disclosure

THE AUTHORS' OWN DECISION INDEX ENGINE, run unmodified: tattrongvu/Jeffan-4B-2010 at 937a008f86f4a20a903d628f22ffad3e3f431d2c, three author package files (__init__.py, engine.py, render.py), the three DI-kit base files at f8552a13b643d39f61185f5951bb5fed16bedf51 and our transport-only jeffan_shim.py. Exact ten-file code manifest sha256 f38f38d69571180283539bb9d8fad4412a6376c5d010bbe53919f8a81a5b8004, hash-checked in the container. jeffan_engine.DecisionEngine(max_len=65536, budget=65536, device="cuda"), Transformers bf16, SDPA with explicit bidirectional masks for full-attention layers and causal Gated-DeltaNet layers, one read. Python 3.12.3, torch 2.13.0+cu130 (author note: 2.14.1+cu130; pyproject requires >=2.9), transformers 5.17.0, tokenizers 0.23.3, flash-linear-attention 0.5.2, fla-core 0.5.2, huggingface-hub 1.33.0. Optional causal-conv1d absent: torch fallback. vLLM is installed in the image but is not the 4b serving backend. Base image vllm/vllm-openai:v0.29.0@sha256:c2914767605584b6d8f45686b82de173ecc99e781897aa3d0a66dacd72c51ae1; derived image sha256:4f4d8a799589b6137954bd2fd7d0f2c4ac8b126ad08e5d31d261efe409bd4b51, hash-pinned wheel overlay. The submitted code was reviewed and cleared at the exact pins with the reviewed safeguards; evaluator exposure and the device were cleared for both attempts. Source review records dormant Hub-capable hf_hub_download/from_pretrained paths; the model argument is a hash-checked local directory, revision=None, offline flags set and network disabled. No author shell/subprocess, eval/exec, pickle or trust_remote_code path on the reviewed serving path. The 4b vLLM plugin was excluded and never executed. The runner and device selection change were reviewed; the amended runner was rechecked with the suite rerun and amendment recorded before retry. Weights: Apache-2.0 at this pin, public, ungated; bf16, 4,205,751,296 parameters, 8,411,558,400 weight bytes. FIRST MEASUREMENT of Jeffan-4B-2010 (our own intake). AUTHOR'S OWN ENGINE on our rented hardware, one request at a time; our harness replaces the author's Decision Index pipeline. NATIVE OUTPUTS: next-token probabilities over the option codes, choice argmax and probabilities, noul P(true), score expectation and probabilities; the first-party transport shim adds no prompt text or probability arithmetic. AUTHOR EXPOSURE: the preregistration records training data as undisclosed (Decision Index note) and author-reported Decision Index suite figures; no overlap or contamination audit was run by us. The author's reported figures are not independently verified by us. NOT ANSWERED, counted wrong: 15 items exceed the author engine's 65,536-token max_len/budget, HTTP 422 Unsupported, exactly the pre-registered standard long positions in the pre-registered long-item set; no truncation. Answered 1,485 of 1,500 items HTTP 200; no unexpected refusal. 10 s RULE: 0 decisions over 10 s; no additional decision is counted wrong by that rule in the v1.7 share view (frozen v1.6 and v1.7-off views apply no time limit). Maximum 0.083058 s, median 0.028291 s, p95 0.044957 s. DETERMINISM: first 50 positions run twice, 50/50 bit-identical, 0 argmax flips, maximum probability difference 0.0. ATTEMPT HISTORY: attempt 1, pod window 24 Oct 2026 05:52-06:21Z (about USD 1.24) on a separate disposable GPU instance, stopped the 31b row after 1,057 of 1,500 items at the runner's own time cap. The smoke under-predicted long-item latency because repeated probe prompts hit the prefix cache. The partial is retained as a receipt only, never scored or merged; the 4b row had not started. Amendment 3 corrected the first-party runner to use cold nonce-bearing probes, maximum anchor latency and a long-item budget floor; a further review preceded the retry. evaluator exposure and device ruling admitted the fresh full 31b run as the final same-use retry; 4b remained a first use. Attempt 2 is this measurement, both rows back to back on new a separate disposable GPU instance, 24 Oct 07:59-08:36Z, about USD 1.57 rental. Between attempts only our first-party runner (smoke probes and fit gate, pre-registration Amendment 3) changed; the scoring harness, model, author code, input and engine settings are unchanged, and the pinned image stack was rebuilt on the new VM.