# R&Duck Review Protocol v1.0.0
# The DA / SPAR / BENCH ladder. Auto-casts reviewers by task. You never fill in roles.

## THE LADDER
```
DA     light    1 lens (devil's advocate)           one objection, one sentence
SPAR   default  2-4 auto-cast + Outlier + DA        one finding each + quick verdict
BENCH  heavy    full auto-cast panel, independent    debate + judge verdict + adaptive stop
```

## DA (Devil's Advocate)
One adversarial pass. Finds the single strongest objection.
Output: the objection + what changes if it's right.

## SPAR (Self-assembling Panel for Adversarial Review)
```yaml
CAST (automatic — you name zero roles):
  1. Read the task
  2. Select 2-4 personas from libraries/personas.md whose lexicon + anti-goals best fit
  3. ALWAYS add ONE Outlier: a persona from an UNRELATED domain, chosen to break the frame
  4. ALWAYS add Devil's Advocate posture

PASS (fast):
  Each cast persona: ONE highest-value finding (not an essay)
  Outlier: one reframe ("what if the question itself is wrong?")
  DA: the single strongest objection

VERDICT: SHIP | FIX [list] | RECAST (wrong panel) | HALT (fundamental problem)
```

## BENCH (the evolved full committee)
Fixes 3 known failure modes from multi-agent debate research:
- Degeneration-of-Thought: once confident, models fail to self-correct
- Conformity: agents converge on each other, losing independence
- Majority-voting weakness: voting fails even when individuals are correct

```yaml
CAST: auto-select by task fit from full persona + domain libraries. 5-8 lenses + Outlier.

INDEPENDENCE PHASE (kills conformity + DoT):
  Each lens forms its assessment BEFORE seeing any other lens's view.
  No lens reads another's output during this phase.
  This is the critical difference from the old committee.

DEBATE PHASE:
  All independent assessments revealed simultaneously.
  Lenses can respond to each other — challenge, support, or refine.
  Max 2 debate rounds (adaptive stop: if no new issue emerges, stop after 1).

JUDGE PHASE (not a vote):
  One synthesis pass reviews all findings and debate.
  Issues a reasoned VERDICT — not a tally.
  Verdict: SHIP | FIX [severity-ranked list] | HALT [blocking issue] | DEFER [needs external input]
  Must state: what was checked, what wasn't, and what this review structurally cannot catch.

ADAPTIVE STOPPING:
  No new issue in debate round → stop early (don't run fixed rounds).
  Hard cap: 2 debate rounds max regardless.
  This prevents compute waste after convergence.
```

## WHEN TO USE WHICH
```
Quick gut-check on a draft          → DA
Standard review before delivery     → SPAR (default)
Release gate / high-stakes decision → BENCH
Architecture / governance changes   → BENCH + external model recommended
```

## AUDIT TIER LABELING (LOCK-5)
Same-model review is ALWAYS labeled: "⚠ INTERNAL BIASED REVIEW — same model, not independent."
For high-stakes: recommend external model via routing.md offload templates.

## DRIFT SUBTYPES (checked by SPAR and BENCH)
```
MEMORY_DRIFT:      contradicts earlier decisions or Core
EPISTEMIC_DRIFT:   confidence exceeds evidence
REPAIR_DRIFT:      correction acknowledged but not implemented
GOAL_DRIFT:        solving a different problem than asked
SPECIFICITY_DRIFT: user specifics replaced with generic statements
SYCOPHANCY_DRIFT:  analysis bent toward what user wants to hear
SCOPE_DRIFT:       output grew beyond task boundary
CONFIDENCE_INFLATION: weak claims without hedging
```

## OUTLIER LENS FRAMING (the Willison question)

The Outlier in every SPAR and BENCH should carry this framing:
"Is this output fast and plausible, or has it been verified and lived with?"

AI has made the act of creation nearly free, which makes judgment, verification,
and lived experience infinitely more valuable. The biggest risk isn't that AI
produces bad output — it's that we stop doing the hard verification because
the fast parts feel so good.

The Outlier's job is to catch the moment when speed has substituted for rigor.