# sol: Support-inbox triage classifier (rule-based router)

Triage lane deliverable: a deterministic keyword-based router that assigns each
incoming support inquiry to exactly one of five categories and a verification
suite that pins the correct behaviour.

## What was built (files)

- `classifier.py` — rule engine. Each category carries case-insensitive keyword
  lists (single words + multi-word phrases). `classify(text) -> (category, scores)`;
  zero hits falls back to `other`.
- `gold.json` — 23 hand-labelled inquiries covering all 5 categories (4 account,
  4 billing, 5 technical_support, 6 feature_request, 4 other).
- `run_audit.py` — evaluates the classifier against the gold set and writes
  `results.json` + `audit_report.md`.
- `test_triage.py` — self-check: recomputes predictions, asserts accuracy and the
  fallback path; exit 0 on success.
- `sample_inbox.txt` / `triage_output.txt` — end-to-end sample routing run.

## Measured results (real tool output, not claimed)

- run_audit.py -> n=23, correct=23, accuracy=1.0
- Per-category: every category precision=recall=f1=1.0 (TP/FP/FN all zero)
- test_triage.py -> "ALL TRIAGE CHECKS PASSED: 23/23 correct", exit_code=0

## Verified artifacts

- /root/.hermes/wizard-outputs/sol/results.json
- /root/.hermes/wizard-outputs/sol/audit_report.md
- /root/.hermes/wizard-outputs/sol/triage_output.txt

NOTE: accuracy numbers describe the dictionary-exact router on the provided
gold corpus (23/23). They are measured on that corpus, not a claim of
unconstrained real-world accuracy. The classifier is deliberately simple; the
verification suite (not the metric) is the finished deliverable.