Skip to content

Engineering Design and Integrity Safeguards

lukeskytorep-bot edited this page Sep 30, 2026 · 3 revisions

Engineering Design and Integrity Safeguards

Current boundary

  • v0.7.13 current public release;
  • SQLite schema 27;
  • modularization Stages 1–9 completed;
  • UX-DATA and architecture gates active.

Blind evidence

Target and Reveal data must remain outside Viewer and Monitor requests until the blind evidence is sealed. Post-Reveal review, Judge evaluation and later discussion are separate evidence domains.

Immutable and frozen state

Session snapshots, sealed evidence, Research Lock, frozen Judge scores and immutable Viewer Learning versions cannot be silently replaced by later Profile changes.

Provider execution layers

Three mechanisms remain separate:

  1. transport retry for classified network/upstream failures;
  2. analytical output recovery for incomplete/length-limited semantic output;
  3. domain recovery such as JSON repair or learning-document capacity retry.

They must not recursively multiply one another.

Provider continuation

Continuation state is persisted outside visible transcripts and tied to the exact assistant message or session event. Replay requires a compatible provider configuration, credential, endpoint, model and format fingerprint.

OpenRouter streaming canonicalization merges only confirmed contiguous text/summary deltas. Opaque encrypted/signature blocks remain unchanged. Limits are applied to the canonical logical state, with independent byte/depth/node guardrails.

Database compatibility

verify:ux-data and native tests protect the exact-green 23→24→25→26→27 path, historical data ownership, flat Conversation hierarchy, controlled purge and factory-target classification.

Release integrity

Full frontend, architecture, Rust and Clippy checks are mandatory. Windows and Linux packages must be produced from the same accepted main commit and verified through their attestations and desktop runtime smoke.

Clone this wiki locally