Guides
End-to-end walkthroughs. Each one states the problem, the primitives it uses, complete code, a walkthrough of what happens, and what the resulting guarantee card does and does not promise.
| Guide | Problem | Primitives and engines |
|---|---|---|
| Support routing with an FDR gate | Auto-route tickets while keeping wrong auto-routes under 5% | Set, Gate(guarantee="fdr"), coverage and risk monitors |
| RAG claim filtering | Ship generated answers whose claims are backed by the docs | Claim (conformal factuality) |
| High-stakes approvals | Different auto-approval bars for different transaction types | Set, Gate(guarantee="risk"), Gate(guarantee="risk_high_probability") |
| LLM-judge cascade | Scale an LLM judge with an audited human-agreement rate | Judge with a cascade, Venn-Abers |
| Cost routing | Keep most traffic on a cheap model with a bound on per-request cost | Route(guarantee="cost_budget"), LocalMonitor |
| Offline calibration | Calibrate an in-house classifier with no data leaving the network | cli_sdk.stats (APS, LAC, Mondrian, IVAP), audit_coverage, LocalMonitor |
| Entity matching with Belief | Merge duplicate records only when the evidence clearly supports it | Belief, the Venn-Abers straddle rule |
| Detecting a silent model update | Notice when a vendor changes the model behind an alias | Fingerprint and coverage monitors, e-processes |
| Label-efficient calibration | Calibrate with 300 human labels instead of thousands | label_with_judge, prediction-powered inference |
The hosted-API guides assume CLI_API_KEY is set (or CLI_BASE_URL
points at a self-hosted deployment). The offline guides
need only pip install cci-sdk.