Get Started/Philosophy & Trust Model

Philosophy & Trust Model

Mandated provenance over voluntary self-reporting

Most AI attribution tools rely on contributors to self-report. Ghost reverses that: the repository owner mandates provenance. Every committed line either carries a cryptographic attribution note recorded when an agent wrote it, or it is assumed human-authored.

THE CORE ATTRIBUTION CONTRACTAttribution is ground truth recorded by local filesystem hooks, or it does not exist. No heuristics, no pattern scanners, no LLMs guessing authorship.

The Double Notes Model

Every commit in a ghost-enabled repo produces up to two Git notes. Together they give a complete picture of code origins:

  • refs/notes/ghost-verified — Written on every commit by the post-commit hook. This is the installation witness that proves ghost was running.
  • refs/notes/ghost — Written only when an agent session was active during editing. Contains line-by-line attribution records.
Verified NoteGhost NoteAttribution Status
✅ Present✅ PresentVerified — AI lines attributed, human lines implicit
✅ Present❌ AbsentVerified — 100% human-written commit
❌ Absent✅ PresentPartial — AI notes present but installation witness missing
❌ Absent❌ AbsentUnverified — Ghost was not running during development

Trust Model

ScenarioTrust LevelBehavior Details
Voluntary installationHighestContributor opted in, notes generated naturally during work.
Mandated bootstrapHighBootstrap is timestamped and recorded in local Git metadata.
Bypassed hooksMediumCommits marked human, untagged commits are the contributor's responsibility.
No setupBlockedCI gates the PR and flags missing verified notes.

Ghost does not try to catch liars. No system can prevent someone from manually retyping AI output character by character. What Ghost does is make deception a conscious, recorded act tied to the contributor's Git identity — while honest contributors are authenticated automatically with zero friction.