Philly Decider v2 (PhillyLabs)v1.7 · 10 s rule

Composite (secondary)

71.2 · Composite rank #10

Cost and measurement conditions

$0.015 per 1,000 decisions (estimated)

Price: ESTIMATE, labelled 'est., base-model price; the model is a fine-tune of a fine-tune (H2O-Lightning-4B v1.1) of this backbone'. The registry base is the ultimate backbone Qwen/Qwen3.5-4B at USD 0.03 / 0.15 per 1M (input / output); output tokens 0 (one label-logprob read per decision, nothing is generated). Effective scoring tariff: the frozen pricing_v15 module (sha256 645fd82eac728d597048bec0995ecd3ac896566e269fb6c9a1fdb84c171a9ac4, re-frozen 26 Oct 2026) lists the backbone string Qwen/Qwen3.5-4B and prices it as deepinfra:Qwen/Qwen3.5-4B at USD 0.03/0.15 per million input/output tokens; neither the repository id PhillyLabs/philly-decider-v2 nor the intermediate model h2oai/h2o-lightning-4b has a listing of its own (the module refuses both with 'lacks an explicit market-listing decision', checked on this module), so the backbone string is the only key used, the same basis as the live H2O-Lightning-4B v1.1 row. No explicit reference or override is used. Estimated, not charged, not a quote from the author or any host. INPUT BASIS: the server's usage.input_tokens is the prompt-token count of the FINAL prompt (instructions, rendered state, options and the 'Answer:' prefill); for the 15 items whose state was trimmed (see disclosure) it is the trimmed prompt, not the original state, which is the live H2O-Lightning-4B v1.1 row's convention. 1,234,232 input tokens across the 1,500 answered items, of which the 15 trimmed prompts carry 482,224 (32,114 to 32,197 tokens each) and the 1,485 other items 752,008; all equal our first-party offline count exactly (0 mismatches over 1,500 items, the 15 trimmed items included; for those, the offline count rests on a transcription of the server's trim and could not be proven offline, so it was a record-only check and it matched). OUTPUT: output_tokens is 0 natively on every row (the shim reads one label distribution with max_tokens 1 and generates no text); it is priced at the module's output rate on 0 tokens, so USD 0; no zero-output convention is invoked. Arithmetic over all 1,500 rows on native usage: USD 0.024685 per 1,000 decisions = 0.38x the unchanged USD 0.06459465517241379 cap. The common 1,485-item basis figure is computed by the scorer. WITHIN COST CAP: included under the automatic cost treatment; no cap exception or top-up.

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, driver 580.178.04), FIN-03 (Finland, EU), PAYG non-spot, instance type 1RTXPRO6000.30V, image 24.04.cuda13.0, USD 2.583095890410959/h recorded by the runner, 150 GB disk. One VM for this row, one measured attempt. Measured 26 Oct 2026, VM window about 19:32-19:43Z (639.6 s from create to closed), about USD 0.46 GPU rental (USD 0.459 at the recorded hourly rate). Author server (stock vLLM at 127.0.0.1:8000 behind the author's shim at 127.0.0.1:8741) in a --network none --read-only container, network cut before sealed input (cut proof and offline model-hash proof receipts: only loopback, no default route, DNS and TCP failed; all 17 model files verified), read-only model, source and input, private output, HF offline flags, no credentials. Before sealed input: the image digest, the stack versions, the Triton gate, the weights fetch with a node re-verify (17 of 17), the pre-sealed smoke in its own server process (/health equal to the preregistered configuration, a cold trimmed probe at the longest original state, anchors from 83 to 40,959 tokens, over-limit and 256-option refusals with recovery, typed probes) and the TTL fit gate all passed (forecast of items over 10 s: 0; forecast compute 674 s); a public pre-warm preceded the scored loop, which ran in a fresh server process. Input sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b (v1.6-regular-20261012) was re-checked inside the container (match). Complete 1,500-row run, rc 0; scored loop 59 s. Raw output sha256 6d02e677e68cd68383a789e4f720e025b4c14845a02c831a2ba95f8b6f7eaf67 (the pull-hash receipt). Node-shred and local-shred receipts verified; exact-id teardown receipt confirms the VM closed. GPU rental is separate from the benchmark's estimated model tariff.