Model providers and inference accounting¶
This page consolidates the public API, model-catalog, and accounting observations from the feasibility work. Capabilities are dated observations, not guarantees of account access or harness compatibility. See the implementation results for subsequent authenticated trials.
Verified provider surfaces¶
| Candidate | Role | Verified public interface | Integration check |
|---|---|---|---|
Neuralwatt deepseek-v4.1-flash |
Generative reviewer and agent model | Chat Completions and documented Responses surface | Tool/structured-output behavior in the selected harness |
| Codex with an explicitly recorded OpenAI model | End-to-end audit baseline | Noninteractive CLI and native skills/subagents | Pinned benchmark runner, model, reasoning, and authentication configuration |
Neuralwatt clef-flash |
Typed evidence judgments and routing | POST /v1/systemone |
Preview enrollment and account-visible model access |
| TypeSafe Jev | Typed evidence judgments and routing | TypeSafe System One API | Choose and verify a versioned model available to the account |
Neuralwatt's catalog returned deepseek-v4.1-flash with source identifier deepseek-ai/DeepSeek-V4.1-Flash-1017 and tool/JSON/vision capabilities. The inspected public response did not include clef-flash, although its model page advertises the preview. That is not evidence that the user's account lacks access. The saved catalog excerpt records the observation. Public model catalog, DeepSeek model page, Clef Flash page.
Use the requested deepseek-v4.1-flash identifier rather than a moving alias such as nw-flash. Still record the served revision and date: Neuralwatt's changelog describes weight refreshes under the same model name. A fixed request string alone is insufficient provenance. Model changelog.
During the initial reference research, the local CLI reported codex-cli 0.166.0 and only version/help commands were inspected. Subsequent controlled trials are documented in the implementation results. A reproducible baseline must explicitly record its model and reasoning configuration.
Keep generative and decision adapters distinct¶
Neuralwatt documents Responses support, including structured outputs, but also differences such as no response-retrieval endpoint or background execution. Therefore “OpenAI-compatible” is a starting point for integration tests, not proof that every Codex feature works unchanged. Neuralwatt Responses API.
OpenCode has a documented Neuralwatt provider route. Verify the current model ID and settings instead of copying older model examples from the integration page. Using Codex against Neuralwatt's Responses endpoint requires separate compatibility testing; the ordinary Codex/OpenAI configuration does not establish that compatibility. Neuralwatt OpenCode integration, Harbor routing.
Neuralwatt Clef Flash uses /v1/systemone, not Chat Completions. Its documented contract accepts state and typed questions, returns distributions, and flags truncation. It documents up to 64 questions, no streaming, and preview-related 404s. Use a dedicated adapter that preserves probabilities, usage, energy, and truncation status. Reject truncated cases from automatic acceptance and record them as incomplete inputs. Neuralwatt System One API.
Do not assume identical confidence semantics across providers merely because the wire format looks similar. Neuralwatt documents its Choice confidence as the winning probability; the reviewed TypeSafe documentation describes distribution concentration. Normalize from retained raw distributions for comparisons, while retaining provider-specific fields untouched. Always supply explicit instructions and criteria rather than depending on provider-specific fallbacks from question IDs. TypeSafe confidence, TypeSafe API.
The earlier Jev review lists versioned candidates. Confirm available versions during authenticated preflight and record one; do not silently switch versions after a failed request. Comparing Jev and Clef should use the same text evidence and label meanings first. Evaluate image-assisted tasks separately where provider capabilities differ.
Measure the auditor without mixing boundaries¶
Neuralwatt documents GPU-counter-based measurement followed by request attribution, with estimated/unavailable alternatives. Its catalog figures are averages of billed energy and may reflect billing caps. Use per-request consumed-energy records and preserve attribution method, rather than treating catalog averages or discounted charges as physical measurements. This is provider-reported attribution, not an independent whole-system energy audit. Energy methodology.
Collect request IDs, requested/served model and tier, token/cache counts, duration, cost, energy units, attribution method, and available carbon metadata. The per-request usage API can support reconciliation, but is account-gated; preserve inline response extensions first. Missing energy or carbon fields remain unavailable, never zero. Per-request usage API.
A documentation inconsistency needs a unit fixture: the energy-methodology page describes a division by 3,600,000 while calling the inputs millijoules. That divisor converts joules to kWh; millijoules require 3,600,000,000. Preserve the API's declared units, check energy_joules / 3_600_000 against energy_kwh, and resolve discrepancies before interpreting results. Do not infer that this documentation error proves the service's measurements are wrong.
Keep three ledgers: provider-reported inference energy, locally measured collector/runtime resources, and unmeasured components. Record training/embodied/facility/network boundaries as covered, estimated, or excluded. Do not add overlapping provider totals or fill missing Codex/TypeSafe energy with zero. Energy comparisons across providers require compatible boundaries; cost and latency remain comparable operational measures when energy is unavailable.
Use total resources per completed, adjudicated task or accepted finding, including retries and escalations. Also report failed attempts, so a cheap system that frequently fails does not look efficient by dropping failures. Repeated trials should record cache warmth and alternate execution order to reduce confounding from changing load.
Keep the website's SWDM/DCR estimate separate from the auditor's inference accounting. They use different units of service and boundaries; a page-view estimate is not a measured offset for model usage.