Introduction

Conformal Logit Inference (CLI) turns the raw output of any LLM — a generated-token log-probability, a sampled answer, a judge’s verdict, a hidden-state activation — into a finite-sample statistical guarantee. It is not a model. It is a calibration and decision layer that sits between the model you already use and the code that has to act on its output.

CLI ships three engines:

Conformal prediction — prediction sets, intervals, and risk-controlled decisions with a marginal or group-conditional coverage guarantee.
Venn-Abers calibration — probability intervals $[p_0, p_1]$ with a formally proven calibration guarantee, instead of heuristic confidence scores.
E-values — anytime-valid drift monitors, safe optional stopping, and false-discovery-rate control that survives arbitrary dependence between simultaneous decisions.

All three require only one assumption: that your calibration data is exchangeable with the production traffic the guarantee is meant to cover. None of them require the underlying model to be accurate, unbiased, or trained any particular way — a worse model only produces bigger sets, wider intervals, or more escalations, never an invalid guarantee.

Where to start

If you want to	Read
Make a first guaranteed call	Quickstart
Understand the one assumption	Exchangeability
See the math behind each engine	The three engines, Venn-Abers, E-values
Pick a primitive	Primitives
Follow an end-to-end scenario	Guides
Connect your model	Backends, Provider connections
Use the hosted service: keys, limits, jobs	Hosted API
See what happened on every request	Observability, Events API, OTLP
Set up a team, roles and SSO	Organizations and SSO
Test without provider keys	Testing without keys
Read a guarantee card	Guarantee reference
Run everything inside your own network	Self-hosting
Handle failures	Errors and retries

Two ways to run it

The hosted service at https://cci.gitdate.ink/api/v1. You bring your own provider key, stored as an encrypted connection, or your own model through a CustomBackend; the service keeps calibration profiles, runs calibration jobs, monitors and audits, and records a trace of every request. Authenticate with a cci_live_... API key.
Local mode in your own process: LocalCLIClient runs the same statistics engine against a model you bring, with profiles stored as files, and the offline engine needs nothing but NumPy.

The seven primitives

Primitive	Returns	Guarantee
`Belief`	A Venn-Abers probability interval for one statement	Venn-Abers calibration
`Set`	A set of options	Coverage $\ge 1-\alpha$
`Interval`	A conformalized ordinal or continuous interval	Coverage $\ge 1-\alpha$
`Gate`	`auto_approve` or `escalate`	Risk $\le$ target, risk $\le$ target with probability $\ge 1-\delta$ , or FDR $\le$ target
`Claim`	Long-form output filtered to supported claims	$P(\text{all retained claims true}) \ge 1-\alpha$
`Judge`	A verdict, with an escalation cascade	Human agreement $\ge 1-\alpha$ on non-escalated verdicts
`Route`	Which backend in a cascade served the request	Cost budget or accuracy bound

Every guarantee in the table is marginal (averaged over exchangeable data) or group-conditional (averaged within a declared group). None is a promise about one individual decision.

Design principle

CLI never returns a guarantee it cannot back. Every primitive requires a calibration profile, and a profile with too few labelled examples for the requested $\alpha$ serves a heuristic-labelled answer instead of silently under-covering. The SDK makes that visible (answer.is_heuristic) and can refuse it outright (CLIClient(strict_guarantees=True)). Each profile calibrates exactly one query against one backend: a different backend on a strict profile fails closed rather than borrowing a calibration it was not built on. See Guarantee reference for the exact schema and Calibration profiles for the lifecycle.

Quickstart