teleo-infrastructure/docs/reports/leo-working-state-20260709/leo-v3-document-source-lifecycle-current.md
2026-07-13 22:21:14 +02:00

2.1 KiB

V3 document to reviewable knowledge: live VPS result

Outcome

Leo can now take the supplied V3 architecture document on the production VPS, compare it with current canonical claims, extract grounded atomic propositions, bind every claim and evidence item to exact source text, and stage one idempotent proposal for human review.

The corrected proposal is 10bc0719-1c2b-5f42-a41e-86e6478692cb. It is pending_review; applied_at is null. No canonical row was approved or applied.

What the real document exposed and fixed

  1. Asking a model to reproduce a quote was not reliable enough. Leo now asks for deterministic source-fragment IDs and restores the original text itself.
  2. Physical line numbers contained gaps at blank Markdown lines. Selectable fragments now have dense IDs, so a valid in-range ID always resolves to non-empty source text.
  3. The first strict proposal retained the atomic proposition only in metadata and planned the literal Markdown quote as claim text. The compiler now enforces the atomic proposition as the proposed claim and keeps the exact source wording as hash-bound evidence.

The flawed pending proposal a9eb4158-5aa2-5484-b5a7-3d249588b3ec was canceled before review or apply, with an audit note pointing to the corrected proposal.

Proof

  • Production deploy: 1a9c08c1a0f7a886ed9fdccb2b6912a8c55af100.
  • Tests: 1,155 full-suite passes; 62 focused source/compiler/bridge passes.
  • Source hash and extracted-text hash both match 59d596e70d26fbc491b4de3228251c308eb9c83cb86ea2dccc9b3342801c5be7.
  • Twelve claim/evidence bindings were checked as exact source substrings.
  • Canonical counts stayed at 1,837 claims, 4,145 sources, 4,670 evidence rows, and 4,916 edges.
  • Exact proposal replay created zero additional rows.
  • Gateway remained active with zero automatic restarts.

Current boundary

This proves the document-to-pending_review VPS lifecycle. It does not prove review approval, canonical apply, Telegram-visible behavior, or GCP replay. The Telegram test must be sent through the connected authenticated Chrome instance; that tab currently presents Telegram's QR login.