Persome

Local-first Personal Model Runtime for macOS

View the Project on GitHub Intuition-Lab/personal-model

Gator quality boundary

Gator is the Runtime’s cross-stage admission policy, not a second capture daemon or a parallel model writer. It keeps raw local observation, normalized activity, and promoted personal-model geometry distinct so recall-oriented capture does not imply first-sighting memory.

capture evidence
  -> privacy boundary
  -> S0 event/content suppression
  -> S1 focused content
  -> timeline provenance
  -> reducer / memory-delta evidence gate
  -> independent-session Point candidates and retry-safe Lines
  -> independent-input promotion for Face, Volume, Root

Current enforced gates

Evidence and compatibility

Raw captures retain their configured local retention policy. A TimelineBlock is a bounded evidence projection, not a promise that every frame is stored forever. Existing blocks migrate to legacy because their original normalization path cannot be reconstructed safely. No migration guesses from entry wording or silently deletes prior model state.

Pre-item-ledger memory deltas keep the same conservative boundary. A legacy row explicitly marked not_requested can take its historical context-free apply path. A legacy pending or failed row with Point/Line/event effects is not replayed automatically: some effects may already have committed, and inventing a receipt after the fact could count an additive Line twice. It remains failed and owner-auditable until an explicit repair can establish what happened. The affected window is quarantined and its session watermark advances, so one irrecoverable historical receipt cannot block all later, fully receipted windows.

An llm block has passed content-signal and response-shape checks; it is not yet a per-entry cryptographic grounding guarantee. Source receipts plus locally verified evidence spans are part of the next promotion hardening slice.

Next hardening slices

  1. Per-entry source receipts and evidence-span validation before normalized claims can promote.
  2. Durable source-event receipts and trigger-to-captured-content snapshot alignment when queueing delays observation.
  3. An owner-visible dirty-data repair workflow for legacy open-Line collision reports.
  4. Candidate/receipt diagnostics in owner-facing model health surfaces without exposing raw capture content.
  5. A batch/indexed PersonGraph identity lookup so large legacy stores do not rescan every current Point for each bounded source event during enrichment.