#!RFC-AI/2;id=22;draft=5;v=0.5.0;status=draft;lang=en;dir=auto;theme=chatgpt-mono;sha256=773561f6f482ae248cdc1cddae85eeb0dfe891b3f25ffe5bd6ec52b57ea5423d
For easier reading, view RFC-AI documents using RFC-AI Reader software.

RFC-AI 22 — DEVIL'S ADVOCATE, VERIFICATION AND CORRECTIVE RESTART
DOCUMENT CLASS: AI RFC — INTELLIGENT-WORK GUIDANCE
STATUS: DRAFT FOR CEO REVIEW — ENTIRE RFC-AI SYSTEM UNIMPLEMENTED

1. PURPOSE — WHAT AND WHY
Use Devil’s Advocate review to improve actual work. DAVG challenges objectives, assumptions, implementation behavior, risks and observed outcomes; it does not enforce arbitrary document length or demand complete API schemas in guiding AI RFCs.

2. INTELLIGENT-WORK DIRECTION
Examine relevant contexts, investigate substantive weaknesses, correct and re-review. When a prerequisite is missing, collaborate with authorized qualified workers until verified satisfactory delivery, then resume. CEO release decisions remain separate.

3. PROPOSED IMPLEMENTATION APPROACH
Use the relevant authorized Interface, component and expressly designated Standard: RFC-AI 3, 11, 29, 31 and worker-coordination Interface. Investigate the complete relevant context before deciding what the assignment means (see RFC-AI 11). A competent worker may discover, invent and improve the necessary methods during implementation; this guiding AI RFC is not required to repeat low-level APIs, algorithms or technology choices. Previous and legacy systems are research evidence only, never an implementation of this draft.

4. EVALUATION AND ACCOUNTABILITY
Use independent DAVG as described in RFC-AI 22 to challenge whether the eventual real-world result meets the objective, not whether the AI RFC looks like a traditional API specification. A missing prerequisite creates an active cross-worker recovery obligation through the worker coordination Interface, not a reason to abandon the task (RFC-AI 20 and 31). Report only verified effects and preserve lessons without reintroducing old operational mistakes.

5. DESIGN CUSTODY
Original subject-specific intended contract: Independent DA/DAVG, checklist, source truth, root cause and one-fix restart from the beginning, release down, evidence-based acceptance, CEO authority for release.
Implementation tasks, legacy-derived candidate methods, proposed data schemas, tests and open engineering questions are kept with this draft's append-only objectives.txt or the responsible Interface/Standard. This RFC has NOT been implemented; it does not approve a deployment or create a new CEO release gate. Rewritten 2026-11-02T20:28:09.285Z.


## HISTORICAL FAILURE EVIDENCE — RESEARCH ONLY; NOT AN RFC-AI IMPLEMENTATION
Historical context: Self-authored tests passed despite missing real-world features; an audit finding sometimes became a passive BLOCKED label, not a completed recovery effort.
Historical reports have mixed evidence quality; verify primary facts before adopting a claimed root cause. Preserve source provenance and do NOT operationalize retired legacy policies as current RFC authority.

## MUST-DESIGN CRITERIA — CEO-DIRECTED PROPOSED FUTURE SYSTEM
MUST-22-01: MUST produce substantive, independent, different DAVG assessments; two nominally separate passes with identical unsupported reasoning do not constitute independent findings.
MUST-22-02: MUST NOT conclude with a passive BLOCKED state. Investigate, contact other qualified workers, correct and repeat differently; continue all available lawful work.
MUST-22-03: MUST treat human approval only as an explicitly authorized design-selection or other human-exclusive gate (for example professional favicon swatches or final graphic choice), never as an invented AI hold.
MUST-22-04: MUST prevent self-certification by independent simple-logic evidence at StorageInterface/real output boundaries, independently assigned auditing workers, or both.
MUST-22-05: MUST repair one substantive defect at a time and re-examine affected behaviors and objective coverage, varying the test as warranted by prior results; repeat-until-success never means identical loops.

IMPLEMENTATION STATUS: Entire RFC-AI system NOT IMPLEMENTED. Historical systems are independent past research material; no test or existing service may be counted as implementing this RFC-AI.


## ADDITIONAL MUST-DESIGN CRITERIA — 2026-11-02 (PROPOSED, NOT IMPLEMENTED)
Historical trigger: legacy materials repeatedly confused authored progress, scripts and document content with real observed execution; this is research evidence, not proof of any new RFC implementation.
MUST: A DAVG auditor MUST use independent source/boundary/physical facts, including evidence of recent actual reads and research submissions, not AI-written assertions. Audit component relationships and each approval piece; challenge reviewer/test independence.
MUST consider the concrete literal behavior of competent AI workers and enforce required hard safeguards by simple, separately running Interface logic where an actual boundary exists. MUST not delegate Interface internal details back to this guiding document. MUST preserve the requirement in relevant top-level objectives.txt and independently verify that future implementation satisfies it. New requirements here remain draft, not live controls.

## DESIGN REVIEW ADDENDUM 2026-11-03 — DAVG AS EMPIRICAL CHALLENGE
REQ-AI-22-DAVG-001: DAVG MUST examine actual artifact correctness, independent domain evidence, the full set of original objectives, authorization and current environment rather than review an AI's self-issued FULL PASS. Reviewer independence, access rights, domain competence and contrary evidence MUST be explicit.
REQ-AI-22-DAVG-002: Retest the changed defect and previously relevant regressions; unchanged already verified work need not be blindly repeated. Record FIXED with independent result, NOT_APPLICABLE with rationale or unresolved dependency with responsible next action; synthetic passing tests MUST NOT imply production approval.

DESIGN CORRECTION DAVG-14 — Independent adversarial review and real deliverable acceptance
RATIONALE: Peostectica historical DAVG packages labeled FULL PASS despite invalid SVG XML, incorrect firewall subnet assumptions and missing execution.
REQUIRED FUTURE BEHAVIOR: An independent reviewer MUST inspect the actual artifact/source version and appropriate real-world domain behavior, not just the worker narrative. Test accepted and forbidden cases, include regression checks for changed defects, preserve previous independently accepted unrelated pieces, label evidence strength and precise unresolved blockers. DAVG findings are FIXED with provider/domain evidence, NOT_APPLICABLE with reasons, or OPEN with missing facts and responsibility; a producer's own assertion cannot be an independent verdict. CEO adoption/release decisions are distinct.
ADVERSARIAL DESIGN ACCEPTANCE: Inject unescaped ampersand SVG, wrong CIDR firewall script and a non-running worker with a heartbeat file. Reject all three despite internally consistent report PDFs and passing superficial checks.
SCOPE: Draft future design and test contract; NOT installed RFC-AI software.