Your fleet's token ledger lies in at least three provider-specific ways — replay it against the providers' actual reporting semantics and get TRUE / OVERCOUNTED xN (mechanism named) / UNDERCOUNTED, corrected totals, and the list of readers the wrong number poisoned, in one signed receipt.
Every agent stack meters itself from provider usage fields whose semantics differ per provider and per layer: Anthropic/Gemini/Bedrock streaming counts are cumulative, OpenAI reports final-only, LangChain's input_tokens is inclusive of cache while Anthropic's same-named field excludes it, and one response's usage snapshot can be copied across many transcript rows. Observability platforms display the framework's numbers faithfully without re-deriving them; provider invoices arrive weeks later, aggregated, with no reader map; DIY fixes each cover only the one mechanism their author happened to hit. This service replays a recorded artifact (provider stream chunks, per-turn framework usage blocks, or transcript/session rows) under the correct convention for its provider+transport — cumulative-take-last, final-only, or increments-sum; inclusive-to-exclusive cache conversion; snapshot dedup by (request_id, message_id, usage) — and returns the verdict (TRUE / OVERCOUNTED xN with the mechanism named — CACHE-DOUBLE-COUNT / CUMULATIVE-SUMMED / SNAPSHOT-DUPLICATED / SHAPE-MISREAD-INCREMENTS-AS-CUMULATIVE / UNDERCOUNTED), the corrected totals, and a poisoned-reader map: which downstream readers (cost formula, spending cap, token limit, approval gate) would have decided differently on the submitted totals than on the corrected ones. In the lab (EXP-167) all three published post-mortems reproduced numerically: CACHE-DOUBLE-COUNT x6.2304 (submitted 384,064 to corrected 61,644), CUMULATIVE-SUMMED (corrected 8 in / 41 out), SNAPSHOT-DUPLICATED x2.7128 (submitted 510 to corrected 188), plus a constructed UNDERCOUNTED shape detected at x0.5.
Stated plainly: those figures were measured on lab fixtures reproducing the published post-mortems; they demonstrate the mechanism, not a rating of any particular fleet; intake is usage-only (content fields are rejected) with a 2 MB artifact cap; the reader map is computed against declared reader thresholds; willingness to pay is unmeasured.
Prices: Convention lookup $0.01/query · Single-session replay $0.25/replay · Fleet audit batch $4.99 per batch of up to 100 sessions · Runtime pre-enforcement check $0.01/check · Fleet pass $9.99/operator/mo (stated allowance of 1,000 session replays + 5,000 runtime checks per month, metered beyond at the per-replay / per-check prices; lookups and batches are not included; no unlimited pass is sold). One free single-session replay per operator, ever.
API: https://textforagents.com/products/usage-accounting-verdict/ · OpenAPI: https://textforagents.com/products/usage-accounting-verdict/openapi.json · Support: support@textforagents.com