Roadmap

1. ✅ Spike: does the model separate dangerous from safe at all?
2. ✅ MVP: Claude Code hook, Tier 0, context (git + paths + redaction), Jev client, policy, audit, install, explain
3. ✅ Breadth: eval harness + corpus (80), adversarial fuzzer, exec shim, enriched targets, doctor hardening
4. ◻️ Tuning: 300-entry corpus, calibration report, tuned bands, BENCHMARK.md
5. ◻️ Phase 2: Codex adapter verification, daemon with LRU cache, goreleaser + Homebrew
6. ◻️ Polish: log viewer, Write/Edit matchers, exec coverage

Explicitly not doing early: hosted service, GUI, Windows support, custom model, multi-agent orchestration.

Design doc: plan.md. Where the shipped MVP deviates from it: DEVIATIONS.md.