v1.0.0 - Disposition v1

Disposition v1: a personalization layer for AI-assisted coding. It learns how you write and steers your AI coding tool to match, so AI output reads like you wrote it and stays maintainable by hand. Fully local. Target: Claude Code, Java, solo plus a shared team layer.

The loop, end to end

• Capture - Bootstrap (author-filtered git mining), Interview (scenarios, /voice narration, adaptive follow-ups), Corrections (explicit and passive git-anchored watcher), Ambient Capture.
• Distill - Induction into rules (auto-accept mechanical, triage the rest) plus a consolidation pass to merge near-duplicates.
• Steer - Forced Injection into Claude Code or a proxy, with three strategies (full / dynamic / hybrid).
• Check - Verification Gate: deterministic tier then adversarial LLM judge, bounded regeneration, human escalation.
• Keep honest - self-aging profile, Drift delta-queries, a shared in-repo Project layer, and cold-start archetypes.

Milestones

• M0 skeleton, M1 thin thread, M2 richer capture, M3 signal quality, M4 phase 2. See CHANGELOG.md.

Notes

• Source install (not on PyPI): see the README. LLM steps need ANTHROPIC_API_KEY.
• 132 tests, all offline (FakeLLM). Proof of concept: pragmatic choices are documented in the ADRs and the implementation plan.