/magi

Devil's Advocate Reference

Purpose: Formal opposition protocol that assigns a rotating role to argue against the prevailing view, stress-testing high-stakes proposals before commitment. Originating in the Catholic canonization process (advocatus diaboli, 1587-1983) and refined by RAND Corporation strategic planning, the protocol institutionalizes dissent so unanimity is earned, not assumed. Counters groupthink without depending on a naturally contrarian member.

Scope Boundary

• magi devil: structured red-team deliberation. A formal DA role is assigned (rotated), tasked with the strongest possible case against the proposal. Output is a stress-test report feeding Magi's standard VOTE, with explicit groupthink check on any 3-0 unanimous result.
• magi decide / tradeoff / arbitrate / strategic (default lanes): balanced three-perspective evaluation. DA is auto-triggered only on 3-0 unanimity; devil runs the full protocol regardless of consensus shape.
• magi sixhat (elsewhere): parallel single-mode rotation including a Black Hat. DA is deeper than Black Hat — it sustains opposition across all modes with formal role separation.
• magi delphi (elsewhere): anonymous multi-round convergence. DA is named (or pseudonymous within a session) and oppositional, not consensus-seeking.
• flux (elsewhere): reframes the problem. DA holds the frame fixed and attacks the proposal within it.
• flux (elsewhere): collaborative idea-deepening. DA is adversarial-by-charter; Flux is generative.
• omen (elsewhere): pre-mortem failure-mode enumeration with FMEA scoring. DA is broader (logic, evidence, assumptions, motivations) and qualitative; Omen is quantitative and failure-focused.

Workflow

ASSIGN   →  rotate DA role; never the proposal author, never the same person twice running
         →  brief DA: "build the strongest case against this, in good faith"

PREPARE  →  DA gets isolated prep time (15-30 min); reads same evidence as proponents
         →  DA may consult external sources; produces written attack list

CHALLENGE →  DA presents 3-7 strongest objections in priority order
          →  proponents may not interrupt; one round of clarifying questions only

REBUT    →  proponents respond to each objection; DA may counter once per objection
         →  facilitator scores: addressed / partially addressed / unaddressed

DECIDE   →  unaddressed objections become explicit risks in the risk register
         →  Magi VOTE proceeds with stress-test results integrated; verdict is final

Role Charter

RAND Tradition (Method of Multiple Hypotheses)

Intellectual Honesty Rules

• No strawmanning: DA must engage the proposal as proponents actually argue it.
• No ad hominem: attack the argument, not the person; institutional DA roles forbid personality contests.
• No tactical opposition: DA must believe the objection is genuinely strong, not raise weak ones to fill time.
• Disclose weak objections: DA explicitly labels low-confidence objections so they're not weighted equally.
• Concede when bested: if proponents convincingly rebut, DA acknowledges. The role is to test, not to win.
• Single-issue depth over breadth: 3 strong objections beat 10 weak ones; weak objections dilute the signal.

When to Invoke

Backfire Effects (Mitigation)

Tie-In With Magi's Logos / Pathos / Sophia

• DA's objections naturally cluster by perspective:
  • Logos-attack: technical infeasibility, evidence gaps, logic flaws, calibration issues
  • Pathos-attack: human cost, team capacity, ethics, stakeholder harm
  • Sophia-attack: opportunity cost, ROI miscalculation, market timing, sunk cost masking
• After DA challenge: each Magi perspective re-votes with DA's objections explicitly considered. Confidence deflation is expected and welcome.
• DA is not a fourth vote — it's a stress test feeding the three votes. Running DA + 3-perspective vote yields stronger calibration than 4 independent voters.
• For 3-0 verdicts: post-DA, if confidence drops below 70, downgrade to "approve with reservations" and elevate unaddressed objections to risk register.

Anti-Patterns

• Same person plays DA repeatedly — rotates identity-based dismissal into structural one. Rotate; track rotation log.
• DA briefed only on weaknesses ("find what's wrong") — produces tactical opposition. Brief on the full proposal and let DA find the strongest flaws.
• Skipping the rebut phase — DA's objections sit unanswered, decisions ship with hidden risks. Always force concrete rebuttal per objection.
• Treating DA as veto — DA's role is to test, not to block. Unaddressed objections become risks, not stop signs (unless severity warrants).
• Running DA only when convenient — DA on easy decisions and skipping on hard ones inverts the value. Triggers should be threshold-based, not mood-based.
• Over-running DA on reversible low-stakes — overhead destroys throughput; reserve for irreversibility, high-stakes, or unanimity.
• Conflating DA with Black Hat — Black Hat is a 3-5 min mode within Six Hats; DA is a 30-60 min role-charter protocol with independent prep, presentation, rebuttal, and audit trail.
• Authoring proposal and playing DA — the same mind cannot honestly attack its own work; assign DA to someone else, full stop.

Handoff

• To Magi VOTE: DA report (objection list with addressed/partial/unaddressed status) becomes input evidence; perspectives re-evaluate with explicit consideration of DA's strongest unaddressed point.
• To Omen: unaddressed Logos-attacks with concrete failure mechanisms escalate to Omen for FMEA-grade quantification.
• To Flux: if DA's strongest objection is "wrong frame," halt VOTE and route to Flux for reframing before resuming.
• To human: if DA produces an unaddressable objection on an irreversible decision, escalate to human regardless of consensus pattern.
• To audit trail: DA name, objections, rebuttal scores, and rotation log preserved; future Magi sessions reference DA quality history for selection.