ouroboros/docs
Anton Razzhigaev 7bf5f71e47 fix(v6.71.1): acceptance-review convergence — evidence-parity, obligation rebuttal wire, honest-partial quorum, ordinary-round acceptance path
The required+blocking acceptance loop terminates by reviewer agreement, not
identical resubmits (field pathology: 61-94% of task cost burned on byte-identical
resubmits; 21/17 consecutive dups on the GAIA loop carriers).

- Evidence-parity: the acceptance reviewer sees each tool result at the actor's
  own per-tool window (SSOT TOOL_RESULT_LIMITS, fallback DEFAULT_TOOL_RESULT_LIMIT),
  with the actor's own truncation marker preserved; llm_trace stores the actor view
  instead of a hidden 700-char head+tail copy.
- Budget ladder: trajectory re-cap runs FIRST (equal-share, floor 700, disclosed),
  then an escape-proof 700-floor backstop; artifact previews and the agent_supplied
  rebuttal channel are last resorts, never collateral; __immutable_core_overflow__
  can only mean genuine core dominance.
- Rebuttal wire: host-attested acceptance_obligations catalog (open rows always
  ship; disposed history count-capped, disclosed) + a commit-gate-strength reviewer
  rule; an agent disposition is PENDING until host settlement (obligation_is_pending
  SSOT), a re-raised finding reopens it, an accepted rebuttal keeps agent provenance
  (disposed_rebuttal_accepted), and a rebuttal is never itself criterion evidence.
- Honest partial: _criteria_shape_valid lets a non-solved-tier partial/missing/
  rejected criterion contribute a valid NON-clean vote; task_acceptance_is_clean
  and the FAIL veto are untouched.
- Ordinary-round acceptance: delivery-control is not armed on the acceptance
  improvement path, removing the stacked-directive freeze.
- Reflection: bounded head+tail marker scan (pre-parity trigger surface preserved;
  doc bodies quoting marker strings no longer false-positive), error snippets are
  redacted before the reflection prompt and pre-capped for breadth.

Bench-verified on the original loop carriers: 21/17 identical resubmits -> 0,
one task at -54% cost; blocking honesty preserved on a genuinely red criterion.
Converged through 2 adversarial rounds, 4-task bench verification, codex review,
and 7 triad+scope rounds (final authorizing pass: clean).
2026-07-18 11:21:53 +00:00
..
ARCHITECTURE.md fix(v6.71.1): acceptance-review convergence — evidence-parity, obligation rebuttal wire, honest-partial quorum, ordinary-round acceptance path 2026-07-18 11:21:53 +00:00
CHECKLISTS.md feat(v6.69.0): review-economics — cacheable reviewer prompts, honest router-rejection accounting, and review-wave budget admission 2026-07-17 08:03:16 +00:00
CREATING_SKILLS.md feat(v6.71.0): UI polish — a shared rich-content contract, one composer dock, calm charts, and whole-row disclosures 2026-07-17 11:49:01 +00:00
DEPLOYMENT.md v5.8.3-rc.2: centralize state helpers 2026-05-08 02:12:09 +03:00
DEVELOPMENT.md feat(v6.70.0): owner-facing honesty — full reviewer rationale, self-locating tool errors, and a decision-turn outcome contract 2026-07-17 10:18:10 +00:00