Eve library

AUD-008 adversarial-review reconciliation — generated

Evidence

Revision 1

On this page

Canonical truth: V1_V9_AUTONOMOUS_CONTENT_SOTA_AUDIT_2026-07-15.md. Adversarial provenance: V1_V9_SOTA_AUDIT_REVIEW_2026-07-15.md. This disposition ledger is generated; it does not promote the review into a second current truth source.

Disposition summary#

Accepted Corrected Rejected Total
7 4 0 11

Finding ledger#

Review Disposition Finding Canonical resolution Follow-up / closure
§3.1 accepted The inherited V3 29/40 gate figure was wrong. Canonical and V3 audit text now derive 30/41 required and 24/39 named counts from the release record. AUD-001
§3.2 corrected The stored V3 record was stale at review time. The stale consumed artifact was accepted; the claim that the verifier itself was never run was withdrawn. A 24-hour freshness contract now applies. AUD-001, AUD-002
§3.3 accepted Unreal availability advice was host-ambiguous. The audit now distinguishes a local observation from project-level availability and names the designated execution profile. AUD-004
§3.4 accepted The original verification table omitted commands, boundaries, and source identity. AUD-003 publishes an authenticated record, exact commands, immutable source, host observations, and bounded logs. AUD-003
§4.1 corrected No in-repository operator UI consumer of the V1 autonomy review API was found. The API remains credited as a tested component; the absolute no-UI claim is narrowed because an external private client cannot be excluded without deployed-consumer evidence. OPS-003, V1-002
§4.2 corrected V1 has partial planner token accounting, not complete end-to-end measured cost. Persisted token/call totals are retained as evidence; generator, research, critic, judge, tool, media/GPU, monetary accounting, feedback, and hard full-path caps remain open. COST-001, COST-002, COST-003
§4.3 accepted Enforcement comments overstated executed-run and V8 byte-binding behavior. AUD-005 corrected the comments, preserved hash scope and signed state, generated field references, and added drift CI. AUD-005
§4.4 accepted Hathor returns an always-empty memoriesTriggered field alongside constant confidence. The canonical prompt audit and remediation order now cover both fields; implementation remains open. V2-001, PROMPT-003
§4.5 accepted The coverage registry can be seeded mechanically from existing catalogs and ledgers. COV-002 now mechanically seeds all 28 V1 surfaces, four explicitly incomplete scheduled categories, every V1–V9 launch/open/blocked ledger row, and both canonical and conflicting V8/V9 gate contracts plus required artifacts and placements. Reviewed migration rules retain every duplicate/contradictory source and the report keeps canonical completeness at zero. COV-001, COV-002
§4.6 corrected The repository declares a single-Hetzner-box topology, but live topology was not probed. The declaration informs box-loss recovery requirements without being promoted to live deployment proof. OPS-001, OPS-002
§4.7 accepted Locale, accessibility, V1 snapshot policy, and test-coherence coverage were absent at review time; repository inventory, release-closure enforcement contracts, and test-coherence reporting now exist while live conformance remains open. COV-004 publishes the source-bound locale/accessibility inventory, COV-005 provides fail-closed release reconciliation when invoked, and AUD-007 publishes test coherence; live artifact conformance, snapshot policy, authoritative shipped observation, and product release-path integration remain open. AUD-007, COV-004, COV-005, PRIV-001, REL-001

Claim boundary#

  • accepted means the finding survived re-verification and is reflected in canonical truth.
  • corrected means the material issue survived but an absolute or unsupported boundary was narrowed.
  • rejected means re-verification did not support the finding. The current count is zero.
  • Follow-up IDs can remain open even when reconciliation itself is complete.