@inferock/measure

TypeScript primitives behind inferock-bench: canonical provider events, token-count checks, billing-integrity signals, pricing, and receipt-ready loss rows.

Use it for local Claude/GPT token usage checks, OpenAI/Anthropic/Gemini/OpenRouter cost evidence (OpenRouter dollar pricing is endpoint-evidence-gated in the public app), billing-integrity tests, and receipt math when you do not need the local proxy process.

Use this package when you want the measurement math without running the local proxy.

| Need | Start here | | --- | --- | | Event shape | Canonical Event Schema | | Signal semantics | Public Signal Semantics | | Receipt ledgers | The Inferock Standard | | Local proxy and captures | inferock-bench README |

import { resolveCostEvidenceForEvent, runStatelessDetectors } from "@inferock/measure/stateless";

const cost = resolveCostEvidenceForEvent(canonicalEvent);
const costUsd = cost.amountUsd; // null when unpriced or partial; never a fabricated zero
const signals = runStatelessDetectors(canonicalEvent);

The CLI and full public docs live at https://github.com/inferock/inferock-bench.

What ships

• Canonical provider event types, including OpenAI, Anthropic, Gemini, and OpenRouter-compatible events; OpenRouter public-app dollar pricing is restricted to pinned observed endpoints.
• Broken-output, billed-empty, refusal, latency, availability, token-count, and billing-integrity detectors.
• Tagged billed, estimated, allocated, and unpriced cost evidence for observed provider usage.
• Receipt-oriented signal fields for spent dollars, bill-bounded money loss, time loss, provider-recognized recovery, recognition gap, and separate invoice-check exposure accounting.

Anthropic Token Recount

Anthropic output-token cross-checks use anthropic_count_tokens_recount_v1 only after runtime per-model calibration between Anthropic messages/count_tokens and the pinned local estimator. Billed usage is excluded from calibration and the learned residual tolerance is capped at 16 tokens. Evidence is capped at provider-assisted grade B. Offline estimates use the vendored MIT Xenova/claude-tokenizer at revision cae688821ea05490de49a6d3faa36468a4672fad; see src/vendor/claude-tokenizer/LICENSE-Xenova-claude-tokenizer.md.

Anthropic does not publish a local tokenizer for Claude 3 or later models, and no API returns an independent recount of billed output tokens. Anthropic-side recounts use Anthropic's count_tokens estimate on delivered output. Per-model overhead and tolerance are calibrated only against the pinned local Xenova tokenizer and delivered text; billed usage is excluded, and tolerance is capped at 16 tokens. The residual limitation is that count_tokens remains provider-assisted and the local Claude 1/2-era tokenizer is approximate, so correlated or systematic tokenizer bias can remain. Evidence is capped at grade B pending an official current public Anthropic tokenizer.

@inferock/measure is Apache-2.0. The inferock-bench CLI is the local proxy that captures real calls and feeds this library.

What to read next

• Pricing methodology for source dates, pricing versions, and pricing_unknown.
• Evidence grade methodology for how detector posture affects provider-recognized recovery.
• Coverage test methodology for the local benchmark path that produces receipt evidence.

inferock-bench is source-available under FSL-1.1-ALv2, and @inferock/measure is Apache-2.0. Together they are the source-available measurement layer of Inferock, the hosted billing-integrity and reliability gateway.