convergence-check

name	convergence-check
description	Skill for determining whether the review process has converged (accept), needs another round (iterate), or has reached diminishing returns (exhausted).

Convergence Check

Use this skill when evaluating whether the peer review loop should terminate, iterate, or be declared exhausted.

Convergence Criteria

ACCEPT (Converged)

All of the following must be true:

∀ reviewer: recommendation ∈ {accept, accept_minor}
AND no unresolved items with severity = critical
AND no unresolved items with severity = major
AND mean_overall ≥ venue_threshold
AND no new blocking concerns raised in re-review

REJECT (Terminal)

Any of the following:

mean_overall < venue_threshold - 1.5
OR ≥ 2 reviewers recommend reject
OR any reviewer identifies fatal flaw:
  - incorrect proof that invalidates main theorem
  - fabricated or manipulated data
  - fundamental methodological error
  - plagiarism or duplicate submission

ITERATE (Another Round)

NOT ACCEPT AND NOT REJECT
AND round < max_rounds (default: 3)
AND score_delta(last_round) > 0.3  (improvement detected)
AND at least one reviewer has unresolved major concerns

DISCUSSION (Score Disagreement)

Any dimension has score spread ≥ 4 points across reviewers
→ Trigger discussion round before issuing decision

The discussion round:

1. Unblinds all reviews to all reviewers
2. Each reviewer writes a brief rebuttal to disagreeing points
3. AE moderates and produces updated synthesis
4. Counts as an additional round toward max_rounds
5. After discussion, re-apply decision rules

EXHAUSTED (Terminal — Diminishing Returns)

round ≥ max_rounds
OR (score_delta(last_round) ≤ 0.3 AND score_delta(round_before_last) ≤ 0.3)

When exhausted, produce an honest assessment with:

• Current state of the paper
• Remaining unresolved concerns
• Specific guidance for future improvement
• Recommendation for alternative venues if appropriate

Applying the Check

Step 1: Gather Data

# Read all re-review files for the current round
ROUND=$(grep 'round:' .review/state.yaml | awk '{print $2}')

# Extract recommendations
for r in alpha beta gamma; do
  REC=$(grep 'recommendation:' ".review/rounds/round-$ROUND/re-review-$r.yaml" | awk '{print $2}' | tr -d '"')
  SCORE=$(grep 'updated_score:' ".review/rounds/round-$ROUND/re-review-$r.yaml" | awk '{print $2}')
  echo "$r: score=$SCORE recommendation=$REC"
done

Step 2: Compute Mean and Delta

# Pseudocode
current_mean = mean(alpha_score, beta_score, gamma_score)
previous_mean = mean(previous_alpha, previous_beta, previous_gamma)
delta = current_mean - previous_mean

Step 3: Check Blockers

# Check for unresolved critical/major concerns
for r in alpha beta gamma; do
  grep -A2 'status: "unresolved"' ".review/rounds/round-$ROUND/re-review-$r.yaml"
done

# Check for new critical concerns
for r in alpha beta gamma; do
  grep -A3 'severity: "critical"' ".review/rounds/round-$ROUND/re-review-$r.yaml"
done

# Check for remaining blockers
for r in alpha beta gamma; do
  grep -A2 'remaining_blockers:' ".review/rounds/round-$ROUND/re-review-$r.yaml"
done

Step 4: Apply Decision

convergence_result:
  round: <N>
  mean_score: <float>
  score_delta: <float>
  all_accept: true | false
  unresolved_critical: <count>
  unresolved_major: <count>
  new_blockers: <count>
  verdict: "CONVERGED" | "ITERATE" | "EXHAUSTED" | "REJECT"
  reasoning: "<why this verdict>"

Score Trajectory Visualization

Track across rounds for pattern detection:

Round 1: α=5  β=7  γ=6  mean=6.0  decision=MAJOR_REVISION
Round 2: α=7  β=8  γ=7  mean=7.3  delta=+1.3  → improving
Round 3: α=7  β=8  γ=8  mean=7.7  delta=+0.4  → improving (slower)

Patterns:

• Monotonic improvement (delta > 0 each round) → likely to converge
• Plateau (delta < 0.3 for 2+ rounds) → consider exhaustion
• Oscillation (alternating positive/negative delta) → fundamental issue
• One holdout (2 accept, 1 stuck) → may need discussion round

Discussion Round

When reviewer scores have spread ≥ 4 points on any dimension:

1. Share all reviews with all reviewers (controlled unblinding)
2. Ask each reviewer to provide a brief rebuttal to the disagreeing reviewer
3. AE moderates and produces updated synthesis
4. This counts as an additional round toward max_rounds

Exhaustion Report

When declaring EXHAUSTED, produce:

exhaustion_report:
  rounds_completed: <N>
  final_scores:
    alpha: <score>
    beta: <score>
    gamma: <score>
    mean: <float>
  score_trajectory: [<round1_mean>, <round2_mean>, ...]
  remaining_concerns:
    critical: [<list>]
    major: [<list>]
    minor: [<list>]
  improvement_guidance:
    - "<specific actionable suggestion>"
  venue_recommendation: "<alternative venue if appropriate>"
  honest_assessment: |
    Multi-paragraph honest assessment of the paper's current state,
    what was improved during review, and what still needs work.