Opus Convergence Analysis

Internal Reference — This is a meta-analysis document, not user documentation.

Date: 2026-01-16 Analysis: Patterns to converge between Marianne Opus (v20, 20 cycles) and RLF Opus (v15, 15 cycles)

---

Executive Summary

Two independent self-improving opus scores have evolved in parallel: - Marianne Opus (v20): Orchestration framework evolution, 9-movement structure - RLF Opus (v15): HTTP API wiring evolution, 5-sheet structure

Both have independently discovered similar patterns, validating their utility. This document identifies patterns to converge, creating stronger scores on both sides.

---

Patterns FROM RLF → Marianne

1. Proactive Discovery Mode ⭐ HIGH VALUE

RLF Pattern:

proactive_mode:
  evaluation_cycle_threshold: 3  # After 3 consecutive eval cycles
  graduated_model:
    cycle_3: "LIGHT PROACTIVE"   # ONE small improvement
    cycle_4_plus: "FULL PROACTIVE"  # Multiple improvements

  light_proactive_options:
    A: "Error Message Review - audit 3-5 error paths"
    B: "Performance Spot Check - profile ONE endpoint"
    C: "Test Gap Assessment - identify ONE untested edge case"
    D: "Documentation Audit - review ONE public API doc"

Why Marianne needs this: - Marianne has no mechanism to prevent idle cycles - When no high-CV candidates emerge, Marianne should still improve - Proactive discovery provides bounded, safe improvement work

Integration approach: - Add to Sheet 1 (External Discovery) phase - Check for consecutive evaluation cycles in header - If threshold reached, mandate ONE proactive task from options

---

2. Variance Tier Model ⭐ HIGH VALUE

RLF Pattern:

variance_tiers:
  TIGHT:
    expected_variance: "5-50%"
    task_types: ["pure delegation", "simple wiring"]

  TIGHT_MODERATE:
    expected_variance: "10-50%"
    task_types: ["light transformation", "DTO reuse"]

  MODERATE:
    expected_variance: "25-100%"
    task_types: ["manual parsing", "new DTOs"]

  LOOSE:
    expected_variance: "50-200%"
    task_types: ["embedded logic", "auth/branching"]

Why Marianne needs this: - Marianne tracks LOC accuracy but doesn't set expectations by task type - Creates unrealistic pressure for high accuracy on complex tasks - Tier-based approach acknowledges irreducible variance

Integration approach: - Add variance tier classification to Sheet 5 (Specification) - Use tier to set pass/fail criteria in Sheet 7 (Validation) - Track variance within tier as success metric (not raw accuracy)

---

3. Task Type CV Modifiers ⭐ MEDIUM VALUE

RLF Pattern:

task_type_cv_modifiers:
  wiring: +0.05        # Infrastructure exists
  bug_fix: +0.05       # Regression prevention
  refactoring: +0.03   # No new features
  documentation: +0.10 # No code changes
  proactive_discovery: +0.03  # Polish work
  true_integration: -0.05     # Cascading changes

Why Marianne needs this: - Marianne calculates CV without task type adjustment - Documentation improvements have artificially low CV - Maintenance work gets undervalued relative to features

Integration approach: - Add to CV calculation in Sheet 3 (TDF Synthesis) - Apply AFTER domain/boundary average calculation - Document modifier applied in synthesis output

---

4. Task Type LOC Multipliers ⭐ MEDIUM VALUE

RLF Pattern:

task_type_loc_multipliers:
  wiring:
    impl: "× 0.8"  # Infrastructure exists
    test: "× 1.0"

  proactive_discovery:
    impl: "× 0.8"  # Polish work is small
    test: "× 0.5"  # May not need extensive tests

  green_field:
    impl: "× 1.35"
    test: "× 1.5"

  true_integration:
    impl: "× 1.5"
    test: "× 2.0"

Why Marianne needs this: - Marianne applies multipliers inconsistently - Proactive/maintenance work is systematically overestimated - Clear multipliers enable accurate planning

Integration approach: - Add to Sheet 5 (Specification) LOC formulas - Apply based on task classification from Sheet 3

---

5. Stable Deferral Threshold ⭐ MEDIUM VALUE

RLF Pattern:

stable_deferral:
  threshold_cycles: 10  # After 10 cycles deferred
  behavior: "STOP RE-EVALUATING"
  rationale: |
    If something has been deferred 10+ cycles without user friction,
    it's not actually needed. Stop wasting discovery cycles on it.

  reactivation_triggers:
    - explicit_user_request
    - new_user_friction_evidence
    - product_decision_made

Why Marianne needs this: - Marianne re-evaluates deferred candidates every cycle - Wastes discovery effort on perpetually low-priority items - Stable deferrals should be marked as CLOSED

Integration approach: - Add deferral cycle counter to research candidates - At 10+ cycles, mark as STABLE_DEFERRAL - Only re-evaluate if explicit trigger

---

6. Test Coverage Multiplier for Enum Complexity ⭐ MEDIUM VALUE

RLF Pattern:

test_coverage_multiplier:
  simple_types: 1.0
    # Single struct request/response

  single_enum: 1.5
    # One enum type in request/response

  multi_variant_enum: 2.5
    # Enum with 4+ variants - each needs test

  nested_enum: 3.0
    # Recursive/nested enum variants

Why Marianne needs this: - Marianne's test LOC formulas miss enum complexity - Multi-variant types need per-variant test coverage - Systematic underestimation when enums involved

Integration approach: - Add enum detection to Sheet 5 (Specification) - Apply multiplier to test LOC calculation - Document detected enum complexity

---

7. LOC Accuracy Trend Tracking ⭐ LOW VALUE (already partial)

RLF Pattern:

# Tracked in opus header
LOC_ACCURACY_TREND:
  v4: "23% variance (baseline)"
  v5: "52% variance (handler type + patterns) <- BIG WIN"
  v6: "52% variance (plateau)"
  v7: "37.8% variance (REGRESSION - soft LOC noise)"
  v8: "6.6% variance (2-step model validated)"
  # ...

Why Marianne might benefit: - Visualizes LOC model evolution over time - Identifies regression points in estimation accuracy - Marianne tracks accuracy but doesn't plot trend

Integration approach: - Add trend section to opus header - Update each cycle with variance achieved - Flag regressions for formula adjustment

---

Patterns FROM Marianne → RLF

1. Pattern Trust Scoring ⭐ HIGH VALUE

Marianne Pattern:

pattern_trust:
  formula: "base + quarantine_penalty + validation_bonus + age_factor + effectiveness_modifier"
  range: [0, 1]

  relevance_adjustments:
    quarantined: "-0.3 score penalty"
    high_trust: "+0.1 to +0.2 bonus"
    validated: "+0.05 bonus"

Why RLF could benefit: - RLF doesn't track pattern effectiveness over time - Patterns discovered in early cycles may become stale - Trust scoring enables selective pattern application

---

2. Synergy Pair Validation ⭐ HIGH VALUE

Marianne Pattern:

synergy_pair:
  criteria: "Both candidates address same problem space"
  implementation_order: "Foundation first, dependent second"
  combined_cv: "Average of pair CVs"

  example:
    pair: ["Pattern Quarantine", "Pattern Trust"]
    shared_space: "Safe autonomous learning"
    order: "Quarantine first (provides status), Trust second (uses status)"

Why RLF could benefit: - RLF implements single candidates per cycle - Related features could be paired for coherent evolution - Reduces context switching between cycles

---

3. TDF Quadruplet Analysis ⭐ MEDIUM VALUE

Marianne Pattern: - Marianne uses full TDF (COMP, SCI, CULT, EXP, META) - Boundary analysis between all domain pairs - Recognition level assessment (P0-P5)

Why RLF could benefit: - RLF uses simplified TDF without META domain - No explicit recognition level tracking - Full TDF could improve candidate selection

---

Shared Patterns (Independently Discovered)

Both opus scores independently discovered these patterns, validating their importance:

1. Coverage Gates (Marianne: ≥80% new code, RLF: >70% coverage)
2. Code Review During Implementation (both track early catch ratio)
3. Evaluation Cycles (both allow cycles without implementation)
4. Multi-Tier LOC Estimation (both use multipliers for complexity)
5. Fixture Catalog Tracking (both maintain test fixture inventory)

---

Integration Priority

Pattern	From	To	Priority	Complexity	Value
Proactive Discovery Mode	RLF	Marianne	1	Medium	High
Variance Tier Model	RLF	Marianne	2	Low	High
Task Type CV Modifiers	RLF	Marianne	3	Low	Medium
Task Type LOC Multipliers	RLF	Marianne	4	Low	Medium
Stable Deferral Threshold	RLF	Marianne	5	Low	Medium
Test Coverage (Enum)	RLF	Marianne	6	Medium	Medium
Pattern Trust Scoring	Marianne	RLF	7	High	High
Synergy Pair Validation	Marianne	RLF	8	Medium	High
TDF Quadruplet	Marianne	RLF	9	High	Medium

---

Recommended Convergence Strategy

Phase 1: Quick Wins (v21)

• Add Task Type CV Modifiers to Marianne
• Add Variance Tier Model to Marianne
• Add Stable Deferral Threshold to Marianne

Phase 2: Structural (v22)

• Add Proactive Discovery Mode to Marianne
• Add Task Type LOC Multipliers to Marianne
• Add Test Coverage (Enum) Multiplier to Marianne

Phase 3: Cross-Pollination (v23+)

• Port Pattern Trust Scoring to RLF
• Port Synergy Pair Validation to RLF
• Consider unified opus template

---

Next Steps

1. Create marianne-opus-convergence.yaml score to implement convergence
2. Create score-creation-skill.md documenting patterns from both
3. Update example scores with lessons learned