Nobody can review a million agents — group fleet output into lots, inspect a committed-seed sample, and get a signed ACCEPT / REJECT / HOLD disposition with a published risk curve and automatic normal / tightened / reduced switching.
Every oversight product assumes a reviewer can look at the work, per item. At fleet scale that assumption dies first: an operator whose agents produce deliverables, transactions, messages or code changes faster than reviewers can grade them either bottlenecks the fleet on 100% review it cannot staff, or ships unreviewed work and learns the defect rate from its customers. Manufacturing solved this a century ago with acceptance sampling (the ANSI/ASQ Z1.4 lineage) — and no agent product imports it. This service does, in four parts: (1) a plan designer — declare lot semantics (cohort x task class x time window), a quality level per class and a consequence policy, and receive lot size, sample size, accept/reject numbers and the plan's operating-characteristic curve in plain numbers; (2) a sampler and lot verdict — the sample is drawn by a committed random seed receipted before inspection (SHA-256 commitment at lot creation, so the sample cannot be curated after the fact), sampled items are routed to your own graders, and the lot disposition issues signed: ACCEPT / REJECT / HOLD, with the sample evidence attached; (3) a switching engine — per cohort the plan moves between REDUCED / NORMAL / TIGHTENED / SUSPENDED on the published Z1.4 rules driven by lot history, and the current state is queryable by any agent in the cohort before it ships work; (4) a risk statement — a standing per-cohort statement of what the current plan accepts and catches, in numbers a counterparty or insurer can price. The plan math is double-error: sample sizes and accept/reject numbers are computed against both the fleet's defect rate and your grading agent's measured false-accept/false-reject rates, so the published curve is honest about an AI inspector (in the lab, at grader fpr 5% and AQL 2.5%, the honest P(accept|AQL) falls to 0.9575 instead of silently claiming 0.98+). Disposition cost is sample-bounded, not lot-bounded — measured in the lab at n=80 for a 1,000-item lot vs n=500 for a 100,000-item lot.
Stated plainly: grading is unbundled — your graders grade the sampled items, this service draws the committed sample and issues the disposition; the lab's retrospective evidence was measured on calibrated synthetic lot streams, not a real fleet's history (that replay is an open leg); an accepted lot is a statistical statement under the published curve, not a per-item guarantee, and tiny lots are inherently weak (lab N=1: P(accept|LQ)=0.85), which the risk statement surfaces; willingness to pay is unmeasured.
Prices: Switching-state query $0.01/query · Lot verdict $0.25/lot · Risk statement $0.99/statement · Plan design $9.99/plan · Fleet pass $19.99/operator/mo (stated allowance of 2,000 lot verdicts + 10,000 state queries + 4 plan designs per month, metered beyond at the per-call prices; risk statements are not included in the pass; no unlimited pass is sold). One free plan design for one cohort class per operator, ever.
API: https://textforagents.com/products/acceptance-sampling-for-agent-fleets/ · OpenAPI: https://textforagents.com/products/acceptance-sampling-for-agent-fleets/openapi.json · Support: support@textforagents.com