nuckcrews/inference-audit

Inference Audit: Provider Delivery and Metering

This dataset contains 3,932 controlled observations from OpenAI-compatible endpoints serving openai/gpt-oss-120b through 18 pinned providers. The runs measure what an API returned and reported at the HTTP boundary: delivery, parameter compliance, token accounting, caching, streaming behavior, latency, and repeatability.

The records do not identify a model from its outputs, prove billing fraud, or establish why two endpoints differ. They preserve observations that can be reproduced and disputed.

What is included

Surface	Evidence in the records
Delivery	HTTP errors, empty output, finish reason, strict-prefix truncation
Contract	Exact/contains expectations, JSON validity, reasoning-off compliance
Context	Deterministic needle recall at roughly 6k, 20k, and 45k words
Metering	Prompt, completion, reasoning, cache tokens, and reported cost
Cache	Reported reuse across controlled shared prefixes and time gaps
Stability	Agreement across repeated byte-identical requests
Streaming	Time to first token, terminal usage, and [DONE] delivery
Experimental	Whether top-logprob evidence was returned; raw values are not published here

Five campaigns were run on September 1, 2026 using inference-audit:

• A 47-case provider comparison with repeats
• A higher-retry standard run
• A controlled cache-reuse run
• Streaming smoke tests
• A logprob-capability run

The included summary.json records source-file SHA-256 digests and per-campaign reports.

Fields

Each row represents one requested case, repeat, and provider. Important columns include:

• campaign, run_id, case_id, category, and repeat
• model, requested_provider, and served_provider
• started_at, status_code, latency_ms, and ttft_ms
• prompt, prompt_sha256, and prompt_chars
• expectation_kind and expectation_value
• reasoning_mode, temperature, finish_reason, and stream_done
• Reported prompt, completion, reasoning, cache, and cost values
• content and error

Provider-generated reasoning text, response IDs, and raw logprob arrays were removed from the public export. Presence and character-count fields remain so the compliance checks are auditable.

Example

from datasets import load_dataset

rows = load_dataset("nuckcrews/inference-audit", split="train")
print(rows.column_names)

Methodology

Requests pin one provider and disable fallbacks. The standard suite uses 47 cases across short answers, code, JSON extraction, reasoning traps, multilingual prompts, and long-context needles. Repeated calls make endpoint instability visible without assuming that cross-provider consensus is truth.

The full definitions and scoring semantics are in the inference-audit specification.

Limitations

• One run describes one routing path, region, model revision, and time window.
• Provider-reported token and cache counts are not independently attested.
• Token disagreement can come from a tokenizer, chat template, serving stack, or accounting.
• Consensus is not correctness.
• A single needle does not establish an endpoint's complete context-window behavior.
• Endpoint behavior and provider configuration can change after collection.
• Model and provider names are identifiers, not endorsements.

Licensing

The dataset compilation, prompts, schema, and original annotations are released under Apache 2.0. Model-generated text and third-party names may be subject to applicable model, provider, or trademark terms; no additional rights to those materials are granted. The repository metadata therefore uses license: other rather than implying one blanket license for every field.

Citation

@dataset{crews2026inferenceaudit,
  author = {Nick Crews},
  title = {Inference Audit: Provider Delivery and Metering},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/nuckcrews/inference-audit}
}