DecisionNode Edge Mini 0.9.2v1.7 · 10 s rule

Composite (secondary)

5.3 · Composite rank #165

Cost and measurement conditions

$0.0046 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/Qwen3.5-0.8B-Base (the base named in the repo's LICENSING.md; the module's DeepInfra base-reference table maps it to the Qwen/Qwen3.5-0.8B listing) priced as Qwen/Qwen3.5-0.8B USD 0.01/M input, USD 0.05/M output (the non-Base sibling listing; the module pins the DeepInfra/OpenRouter price snapshots of 25 Sep 2026). Estimated, not charged, not a quote from the author or any host. Tokens as reported by the usage block of the author's server: input_tokens = the rendered prompt of the one question (chat template and options included); output_tokens = 0 on every row (the server generates no text: one forward pass and a readout head). The server-reported total of 677,924 input tokens over the 1,485 answered items equals the pre-registered offline count (first-party transcription of the author's prompt renderer and the pinned tokenizer, made before the run), with 0 mismatches item by item and offset 0; the same total applies to both DecisionNode rows. The 15 refused items (HTTP 400) carry no usage and count 0 tokens. Those 15 are the 15 standard very long items, which count for no row's cost on the common 1,485-item cost basis (v1.6.30); the common-basis input tokens are therefore also 677,924, with 0 output tokens. The preregistration (Amendment 2) predicted USD 0.004519 per 1,000 decisions on the 1,500-item basis (0.0700x the cost cap) and USD 0.004565 on the common 1,485-item basis (0.0707x); the measured token total equals the predicted one. COST ON THE COMMON BASIS: USD 0.004565 per 1,000 decisions on 1,485 items, 0.071 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 2.574/h as recorded by the runner (300 GB disk configured). One VM for both rows back to back with verified node shredding after each row and a final shred. The author's server (127.0.0.1:8000) in a --network none --read-only container per row; driver requests serial over loopback. Read-only code, model and input, empty credential variables, HF offline flags; cut proof inside the container before any sealed input (only loopback, no default route, DNS and TCP fail). Measured 24 Oct 2026 (reservation 05:33:57Z, VM created 05:35:01Z, teardown confirmed 05:43:05Z); this row's measurement container 91 s (run cap 109 s), scored loop 49.297 s over 1,500 requests. Frozen v1.6-regular-20261012 self-hosted input sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b, checked in the container. Raw output sha256 46b320c66dc0821d94fcd4a60a58a837d06f331357002b244e7e6c7fa4176cb3. Node-shred receipt verified for each row; exact-id VM and volume teardown confirms provider absent and guard closed. GPU rental is not the benchmark's estimated model tariff.

Model and serving disclosure

THE AUTHORS' OWN SERVER, run unmodified: bynn-intelligence-com/decisionnode-edge-mini at b1ca52601cd95ef543bbe2bd0cebae2ed3f3b3fa (tag v0.9.2) with the authors' PyPI package decisionnode-edge 0.9.2 (wheel sha256 dfb0b48c1233c32a106d8eee4d6c320dc3454026a1119851787f345b2f5c6317, sdist sha256 2f1da8de0be9edb3a58c3804dc024cede75b53dcea9aa82a42645e1a189ccbd0): 11 Python modules, installed from the hash-pinned wheel with --no-deps --no-index under a network-free image build; the 11 modules equal (git blob oid) runtime/src/decisionnode_edge in both Hub repos at their pins. Only the 10 data files of the pinned revision were mounted read-only at /model (config.json, decisionnode_edge.json, calibration.json, box_calibration.json, tokenizer.json, tokenizer_config.json, preprocessor_config.json, video_preprocessor_config.json, model.safetensors, readout.safetensors); no Python, GGUF or shell file of the Hub repo was fetched. Command: env HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 USE_HUB_KERNELS=NO DECISIONNODE_EDGE_BACKEND=torch PYTHONDONTWRITEBYTECODE=1 python3 -m decisionnode_edge serve /model --host 127.0.0.1 --port 8000 --backend torch. PyTorch bf16 reference backend, one request at a time, no optional kernels (flash-linear-attention, causal-conv1d, Hub kernels, openvino and mlx absent; checked in the image). Image: vllm/vllm-openai:v0.30.0@sha256:8a69ffad015f138d7170c4ddc429e230a3bc1c1719f67e14324749df200a4b90 used as the CUDA 13.0 / PyTorch base (vLLM 0.30.0 is installed in it but was not used: the author's default --backend auto would select vllm there, so torch was set on the command line and in the environment, and /healthz reported backend torch after the sealed run), derived image sha256:b1f31eab61d184033ae78a92aa778c7faa2c909f8566725b4716d7803535dd47 built offline from three hash-pinned wheels (decisionnode_edge 0.9.2, transformers 5.18.0 sha256 d79e5a515a572ee3deb33eb0d578be47d91e70c75d54cf21bf6eede6edef14a9, huggingface_hub 1.33.0 sha256 04e434b06e100eddbce9a6e817d72693a7884b10a79bd67ab48080d5c07eb899). Python 3.12.3, torch 2.13.0+cu130, transformers 5.18.0, tokenizers 0.23.2, safetensors 0.8.0, numpy 2.2.6. Code gate: cleared by the benchmark owner on the disclosed patterns. DISCLOSED PATTERNS: dormant subprocess start of a llama-server child in the llamacpp backend (backends.py:353), reachable only with --backend llamacpp; GGUF build subprocesses (build.py:134,136), a file imported by nothing and run only as python -m decisionnode_edge.build; Hub-capable from_pretrained loaders (model.py:49,92,235) that receive the local hash-checked model folder, under HF_HUB_OFFLINE=1, TRANSFORMERS_OFFLINE=1 and a container without network. No runtime download, trust_remote_code, string eval/exec, pickle, telemetry or request-body logging in the author code. The source and runner were reviewed before the run, with the required conditions enforced. A leftover .cache directory that would have stopped the run before any sealed item was removed and the fix rechecked; the exact device selection change was also reviewed. Weights: AGPL-3.0, public, ungated, 853,051,456 bf16 parameters, 10 data files of 1,727,281,260 bytes in total hash-checked file by file (verdict OK); prereg manifest sha256 4c7373fe2e2f94ef2f0df038708580dd2acbdae840f595142155c2cae1e27abb. The Hub main branch of the repo is one commit ahead of the pinned tag (0e77dfdc), with zero file differences against the pinned tag commit (added 0, removed 0, changed 0 in our file-by-file comparison); the measurement stays at the pinned commit b1ca5260. FIRST MEASUREMENT of DecisionNode Edge Mini 0.9.2 (our own intake; filed as GitHub issue 231 by the account @bynnit on 23 Oct 2026). No DecisionNode row exists on the live JevBench board (v1.6.42, checked 24 Oct 2026). AUTHOR'S OWN SERVER on our rented hardware, PyTorch bf16 reference backend, one request at a time. NATIVE OUTPUTS: probabilities over the item's options from the model's readout head with the author's calibration applied by the server (choice argmax and probabilities; score level, expectation and probabilities; noul P(true) from the server's truth answer reported as a noul); the server rounds probabilities to 2 decimals; nothing is generated. NAMING NOTE: the package, the PyPI release and the Hub tags are 0.9.2, but the served model name reads decisionnode-edge-mini-0.9.1 (/healthz, /v1/models and every response): the name string comes from decisionnode_edge.json in the model folder. The release owner's triage of 23 Oct 2026 found model.safetensors, readout.safetensors, config.json, tokenizer.json and calibration.json byte-identical across the tags v0.9.1 and v0.9.2; v0.9.2 adds the /v1/systemone server, box answers and documentation. The measured model tensors are therefore the 0.9.1 weights, served by the 0.9.2 runtime. PYTORCH PATH: the author's default backend selection would have used vLLM in this image; we forced the PyTorch reference path, without the optional flash-linear-attention and causal-conv1d kernels. The author's README says each backend agrees with the PyTorch reference on at least 98% of top answers and that the PyTorch path without the kernels gives the same answers about twice as slowly on long inputs; we did not test either statement. The README's timing (7.3 ms at about 500 tokens, vLLM backend, RTX PRO 6000) does not transfer to this PyTorch measurement. AUTHOR EXPOSURE (volunteered by the author in the submission and recorded in the release owner's triage; not verified by us): no JevBench item is in the training data (every training record was checked against the public items by exact text and overlapping word windows, and matches were dropped), but the public items were one of the authors' development suites: they read scores on them while comparing training runs, which is model selection against the public split. No overlap or contamination audit was run by us. LICENCE NOTE: AGPL-3.0 for the weights and the runtime; Bynn Intelligence Inc. also offers a separate commercial licence (LICENSING.md); whether the page needs any licence line is a release-owner decision. NOT ANSWERED, counted wrong: 15 items exceed the author's 16,384-token context limit (HTTP 400, error_type max_tokens_exceeded), exactly the 15 standard very long items (76,309-80,466 tokens by our first-party offline count) at positions in the pre-registered long-item set, as pre-registered by id and position; never truncated. Answered: 1,485 of the 1,500 items sent, HTTP 200 (status counts 200: 1,485, 400: 15); the longest answered item was 1,794 tokens. 10 s RULE: 0 decisions over 10 s, so no decision is scored wrong under the 10 s rule in the v1.7 share view (the frozen v1.6 and v1.7-off views apply no time limit); maximum 0.108 s over the 1,500 requests (median 0.0314 s, p95 0.0454 s). DETERMINISM: first 50 positions run twice, 50/50 bit-identical, 0 argmax flips, maximum probability difference 0.0 (compared on the server's 2-decimal outputs). CONTEXT: author-imposed limit 16,384 tokens; no truncation. Pre-sealed synthetic probes (not benchmark items): a 16,384-token request was accepted (0.64 s), requests of 16,385 and 82,000 tokens were refused with HTTP 400 and the server stayed healthy. The vendor's reported figures are not independently verified by us.