# Review Architecture
> [!summary]
> Eval Labs now separates fast employee judgment from senior adjudication so Lucia can learn from structured signal without letting non-expert reviewer language become training truth.
---
## The three-layer review model
Eval Labs review is not one flat annotation step.
It is a layered judgment system:
```text
Layer 1: Employee Review
→ Layer 2: Escalation Routing
→ Layer 3: Senior Adjudication
```
Each layer has a different job.
---
## Behavioral Observatory position
Behavioral Observatory adds a saved behavioral label layer beside the Review Queue workflow.
It does not erase the three-layer review model.
Use this distinction:
```text
Review Queue = score and review the prompt/response item.
Behavioral Observatory = save structured behavioral labels for the conversation.
Registry Diagnostics = inspect derived classification suggestions.
```
Behavioral Observatory labels can become useful behavioral evidence, but they should not be confused with senior adjudication or Gold Standard approval.
---
## Layer 1 — Employee Review
Employee Review captures fast human reaction.
It is designed for reviewers who may not know AI, prompting, ontologies, or model training.
Employees answer guided questions such as:
```text
Did Lucia understand what was needed?
Did Lucia give the right next move?
Did Lucia make the situation feel calmer?
Did anything feel risky, confusing, or wrong?
Should a senior reviewer look at this?
Could this teach Lucia something reusable?
```
This layer should be:
<ul class="eval-status-list">
<li>simple</li>
<li>fast</li>
<li>structured</li>
<li>low-friction</li>
<li>non-technical</li>
<li>psychologically clear</li>
</ul>
Employee reviewers should not be asked to invent labels, taxonomies, intent categories, action classes, or training language.
---
## Layer 2 — Escalation Routing
Escalation Routing converts guided review answers into workflow state.
Examples:
```text
seniorReview = true → reviewState: needs_adjudication
reusableLearning = true → canonCandidate: true
riskOrConfusion = slightly_off → reviewState: needs_review
riskOrConfusion = definitely_wrong → reviewState: needs_review
```
The important doctrine:
```text
The employee reports what they experienced.
The system routes the case.
The senior reviewer interprets meaning.
```
---
## Layer 3 — Senior Adjudication
Adjudication is the senior-review layer where canonical meaning is assigned.
Adjudication may include:
- final human labels
- final intent interpretation
- follow-through decision
- final action type
- emotional read
- owner pressure level
- reason for the final call
- reusable canon/training signal
This layer exists to prevent ontology drift.
Employees should not train Lucia directly with improvised language.
---
## Why this architecture matters
Without this separation, Eval Labs risks collecting inconsistent reviewer opinions as if they were stable training truth.
That creates:
```text
ontology drift
label noise
inconsistent training signal
reviewer fatigue
low-confidence exports
```
With this separation, Eval Labs captures simple human judgment while preserving a high-quality senior interpretation layer.
---
## Canon rule
> [!warning]
> Non-expert reviewers provide signal. Senior adjudication provides meaning.
This is the core protection layer for scalable human evaluation.