Inbox triage agent: daily evaluation

Runs every day on GitHub Actions. Samples labelled messages (synthetic, not customer data), runs the agent (claims guardrail, then an LLM classifier with a local fallback, then a confidence router) and scores it. Last run: 2026-11-11 · mode: llm · model: gemini-3.1-flash-lite

Accuracy, familiar wording
96.0%
Accuracy, new wording
80.0%
Auto-handled (familiar)
56.0%
Claims sent to a human
100.0%

New-wording accuracy over time

History (last 30 runs)

DateModeAcc. familiarAcc. new wordingAuto acc. (new)Auto rateClaims to human
2026-11-11llm96.0%80.0%92.3%56.0%100.0%
2026-11-10llm92.0%72.0%83.3%44.0%100.0%
2026-11-09llm96.0%92.0%92.9%48.0%100.0%
2026-11-08llm88.0%76.0%90.0%68.0%100.0%
2026-11-07llm92.0%80.0%88.9%60.0%100.0%
2026-11-06llm96.0%72.0%69.2%60.0%100.0%
2026-11-06fallback-only97.5%20.0%0.0%52.5%100.0%

Today's errors (up to 15)

MessageTruePredictedConf.Source
when is my next payment due also our driver rear-ended someonebillingclaim_report1.0guardrail
thanks need the acord 25 form also one of our trucks was in an accidentcertificate_requestclaim_report1.0guardrail
got a late payment notice also need the acord 25 formbillingcertificate_request0.9llm
thanks is there a fee for paying in installments also we bought another vanbillingpolicy_change0.9llm
following up - client won't pay until they get our insurance certificate also closing the business, please cancel coveragecertificate_requestcancel0.95llm
following up - please terminate effective end of month also employee got hurt on the job todaycancelclaim_report1.0guardrail

Source: github.com/f20220970-debug/llm-triage-agent