# Evolution 81 — Specialist Competency Report

## Method

- 30 disciplines, 3 distinct shapes each (Analyst, Engineer, Reviewer) = 90 profiles.
- Each profile was evaluated on a **3-item, discipline-neutral probe** (arithmetic, instruction
  following, verification) with the real local Qwen2.5-0.5B-Instruct model.
- Level bands: L5 ≥ 0.90, L4 ≥ 0.75, L3 ≥ 0.50, L2 ≥ 0.25, L1 > 0, L0 = unmeasured or 0.
- Competency is **measured**, never self-declared. An unevaluated profile is L0.

## Result

| level | profiles |
|---|---|
| L0 | 30 |
| L1 | 0 |
| L2 | 30 |
| L3 | 30 |
| L4 | 0 |
| L5 | 0 |

Mean probe score: **33.3%** across
90 profiles.

## Per discipline

| discipline | profiles | mean probe score |
|---|---|---|
| ai_research | 3 | 33.3% |
| autonomous_operations | 3 | 33.3% |
| biotech_literature | 3 | 33.3% |
| business_strategy | 3 | 33.3% |
| creative_design | 3 | 33.3% |
| cybersecurity | 3 | 33.3% |
| data_science | 3 | 33.3% |
| devops | 3 | 33.3% |
| distributed_systems | 3 | 33.3% |
| economics | 3 | 33.3% |
| education | 3 | 33.3% |
| energy_systems | 3 | 33.3% |
| environmental_science | 3 | 33.3% |
| financial_analysis | 3 | 33.3% |
| formal_verification | 3 | 33.3% |
| healthcare_research | 3 | 33.3% |
| infrastructure | 3 | 33.3% |
| legal_research | 3 | 33.3% |
| market_intelligence | 3 | 33.3% |
| mathematics | 3 | 33.3% |
| ml_engineering | 3 | 33.3% |
| physical_ai | 3 | 33.3% |
| product_development | 3 | 33.3% |
| research_synthesis | 3 | 33.3% |
| robotics | 3 | 33.3% |
| scientific_research | 3 | 33.3% |
| software_engineering | 3 | 33.3% |
| statistical_analysis | 3 | 33.3% |
| supply_chain | 3 | 33.3% |
| technical_writing | 3 | 33.3% |

## Honest interpretation

This is a **reliability floor**, not domain expertise. A 0.5B CPU model cannot be shown to
perform professional work in law, medicine, economics or formal verification by a 3-item probe.
The registry supports real domain benchmarks (the `record_evaluation` API accepts any benchmark,
dataset hash and held-out flag), but none were run. G11 is therefore **PARTIAL**.

## What is real

- 90 durable, versioned profiles with tools, memory scopes, risk ceilings and audit history.
- Composition is authority-attenuating and tested (SEC-056/SEC-057).
- Selecting a specialist never grants authority; the Gate still decides every call.
