/darwin

Official Fitness Criteria Reference

Sources:

• "The Complete Guide to Building Skills for Claude" (Anthropic, 2025) — Progressive Disclosure / frontmatter / instruction structure baseline.
• "New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration" (Anthropic, 2026-05-06) — adds Outcomes rubrics as the new official success-criteria primitive. URL: https://claude.com/blog/new-in-claude-managed-agents.

Official-standards-based fitness evaluation reference used by Darwin in the ASSESS / EVOLVE phases.

2026-05 update: Anthropic's Managed Agents now ship a first-class Outcomes mechanism in public beta — agent operators write a rubric describing what success looks like and the agent works toward it. Internal Anthropic testing measured up to +10 percentage points task-success lift over a plain prompting loop (+8.4 pp on docx generation, +10.1 pp on pptx generation). Darwin treats the presence of an explicit Outcomes rubric (or local equivalent) as a positive OSC signal for the Instruction Structure and Error Handling criteria below. Conversely, an agent stuck on a plain prompting loop with no rubric — when running under Claude Managed Agents — is now an OSC penalty (≥ −10 toward the Instruction Structure dimension), because the +10 pp ceiling lift is freely available.

1. EFS Official Spec Conformance Dimension

Integrate Official Spec Conformance (OSC) as a supplementary metric into the existing EFS 5-dimension evaluation.

OSC Scoring (0-100)

OSC Grades

How It Integrates with EFS

Rather than being a 6th EFS dimension, OSC functions as a quality correction factor on the existing 5 dimensions:

Adjusted_EFS = EFS × (0.8 + 0.2 × (OSC / 100))

• OSC = 100 → EFS × 1.0 (no change)
• OSC = 50 → EFS × 0.9 (10% reduction)
• OSC = 0 → EFS × 0.8 (20% reduction)

A conservative integration that doesn't break the existing EFS calculation. Higher OSC preserves EFS; lower OSC applies only a mild penalty.

2. Strengthening RS Evaluation with Official Quality Signals

Incorporating Official Metrics into the Agent Relevance Score (RS)

Add the following official quality signals as supplementary inputs to the existing RS calculation:

3. Integrating Official Standards into Evolution Triggers

New Trigger: ET-09 — Official Spec Deviation

Linking Official Standards to Existing Triggers

4. Applying Official Standards by Lifecycle Phase

Expected OSC by Phase

Application in the ASSESS Phase

1. Detect the lifecycle phase (existing logic)
2. Look up the Minimum OSC for that phase
3. Calculate the current OSC for each agent
4. Add agents where Current OSC < Minimum OSC to the improvement candidate list
5. Include an OSC section in the DARWIN_REPORT

5. Ecosystem Coverage Analysis by the 3 Official Use-Case Categories

Agent Mapping by Category

Coverage Gap Detection

Evaluate the following in the ASSESS phase:

• Whether each category has sufficient agents
• Whether the balance across categories is appropriate
• Whether uncovered categories are prioritized when proposing new agents

6. Adding an Official Standards Section to DARWIN_REPORT

Output Format Extension

## Official Spec Conformance (OSC)

| Agent | OSC Score | Grade | Min Required | Status |
|-------|-----------|-------|-------------|--------|
| [name] | [0-100] | [S-F] | [phase-based] | [PASS/BELOW] |

### OSC Summary
- Ecosystem Average: [score]
- Agents Below Minimum: [count]
- Top Improvement Candidates: [list]

### Use Case Coverage
- Document & Asset Creation: [agent count] agents
- Workflow Automation: [agent count] agents
- MCP Enhancement: [agent count] agents