/echo

Council Mode — Persona Council (v4 fold-in)

Persona Council mode runs parallel multi-persona evaluation against a machine-readable Persona Contract. Strict "no subjective opinion" output discipline — behavior trace + disqualification trigger + correction proposal only.

When to Use

• nexus growth-acceptance Phase 0 persona evaluation (required).
• nexus acceptance Phase 3B AI-user persona evaluation (Tier-S/A).
• Standalone: multi-persona disqualification check before launch.

Persona Weights

Org-Tier Cost Cap

Mandatory per Magi v4 C5 (enforced via Always/Never, no new guardrail).

Never exceed cap with "just one more persona". If exhausted, defer to next session.

Engine Diversity (Tier-S/A Required)

For Tier-S/A evaluations, Persona Council MUST run via rally engine-paradigm mode (Codex + Antigravity + Claude). Single-engine Council:

• Tier-S: forbidden.
• Tier-A: advisory only.
• Tier-B/C: allowed.

Rationale: correlated hallucination risk per Magi v4 G16 fold-in.

Confidence Discipline

All Council output tagged [hypothesis] confidence by default. Promotion to [validated] requires Voice/Trace real-user calibration per Insight Ledger Survivor Bias rule.

Persona Contract Schema

persona_id: <unique id>
weight: primary | secondary | non-target | risk
situation: <one-sentence current context of the persona>
goal: <what the persona is trying to accomplish in this session>
fear: [<concern 1>, <concern 2>, ...] # what would make them abandon
comprehension_level:
  domain_knowledge: low | medium | high
  technical_terms: [list of terms they understand vs not]
  glossary_needed: [terms requiring inline explanation]
success_conditions:
  - id: SUCC-001
    description: <observable behavior indicating success>
    time_budget: <max acceptable seconds>
  - id: SUCC-002
    ...
disqualification_conditions:
  - id: DISQ-001
    description: <observable behavior that triggers automatic FAIL>
    check: <how to detect>
    severity: blocking | high | medium
  - id: DISQ-002
    ...

Output Schema (Strict — No Free-Form)

council_evaluation:
  persona_id: <ref>
  target_artifact: <screen/flow/copy under evaluation>
  result: PASS | FAIL | INCONCLUSIVE
  behavior_trace:
    - step: 1
      action: <what persona did>
      observation: <what they saw>
      duration_seconds: <numeric>
    - step: 2
      ...
  disqualification_triggers: [<DISQ-ID list, empty if none>]
  success_achieved: [<SUCC-ID list, empty if none>]
  correction_proposals:
    - target: <element id>
      change: <specific concrete change>
      rationale: <which disqualification this addresses>

Forbidden Outputs

• Subjective opinions ("seems good", "feels nice", "I think users will…").
• Free-form recommendations outside the correction_proposals schema.
• Aggregate "overall verdict" outside per-persona result field.

Always / Never Recap

Always:

• Emit Persona Contract first (situation/goal/fear/comprehension/success/disqualification).
• Produce only behavior-trace YAML — no free-form opinion.
• Respect Org-Tier persona cap.
• For Tier-S/A: run via rally engine-paradigm engine diversity.
• Tag output as [hypothesis] confidence by default.

Never:

• Emit subjective opinions.
• Exceed Org-Tier persona cap.
• Rely on single-engine evaluation for Tier-S (correlated hallucination risk).