Decisio 12B tuned commit (variant of the Decisio family)v1.7 · 10 s rule

Composite (secondary)

72.2 · Composite rank #5

Cost and measurement conditions

$0.024 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference, explicit reference with NO base key: the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-freeze of 22 Oct 2026) has no listing for the card's base string google/gemma-4-12B-it (it is in the module's UNLISTED_BASE_MODELS set, so there is no exact rate and no floor; reference() returns nothing for it, checked on this module). The reference used is therefore google/gemma-3-12b-it USD 0.05/M input, USD 0.15/M output, the nearest publicly hosted 12B Gemma sibling, at its FULL catalog pair, the same explicit reference that the published google/gemma-4-12B-it rows cygnet, jev-omni, winnow-12b, torchcast-decision-12b and the live decisio v0.8.0 and v0.9.0 rows already carry. It is applied explicitly, as a labelled estimate; no base key is registered for it. Estimated, not charged, not a quote from the author or any host. Tokens as reported natively by the usage block of the author's server: 715,443 input and 1,485 output tokens across the 1,485 answered items (1 reported output token per answered row for the option-label readout; no generated prose). The native input unit counts header, state and question separately without the system turn or chat template. It is not a count of every token processed internally: the server can also register the state prefix through a warm-up, and that work is not added to native usage. Native input usage equals our first-party offline count exactly (0 mismatches over the answered items; our sealed-input count before the run was also 715,443). The 15 items refused HTTP 422 carry no usage and count 0 tokens, so the common 1,485-item basis (v1.6.30) has the same totals, 715,443 input and 1,485 output tokens. The zero-output convention for native-probability rows is NOT applied: the server reports a real output count of 1 per answered row (the same treatment as the live decisio rows, seb-9b-v2 and KnowLine, which price their 1 output token per row), and the full output price is used. Measured arithmetic: USD 0.024239 per 1,000 decisions on the common 1,485-item basis = 0.375248x the unchanged USD 0.06459465517241379 cap; separately USD 0.023997 per 1,000 over all 1,500 rows = 0.371495x. WITHIN COST CAP on both bases: within 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, USD 2.558095890410959/h recorded by the runner, 150 GB disk. One VM for this row, one attempt. Measured 25 Oct 2026, VM window about 23:26-23:39Z, about USD 0.55 GPU rental (elapsed time at the recorded hourly rate). Author server at 127.0.0.1:8000, network cut before sealed input (offline proof and cut proof receipts), read-only model and input, private output, HF offline flags, no credentials. Before sealed input: model files re-hashed against SHA256SUMS; the dependency Triton gate, a pre-warm and the smoke passed (the smoke includes both 422 over-limit probes at 32,768 and 80k tokens); the TTL fit gate admitted the run. Input sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b (v1.6-regular-20261012) was re-checked inside the container (match). Complete 1,500-row run, rc 0. Raw output sha256 e9c889e1a138c875eb52521748e63b4344043c7a5598527c073172dd1c81594f (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.