> ## Documentation Index
> Fetch the complete documentation index at: https://docs.abigail.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Glass Box Transparency

> Every analysis includes metadata showing how AI decisions were made.

Every response from the `/analyze` endpoint includes a `glass_box` object that provides
full transparency into the AI decision-making process.

## Why Glass Box?

Patent prosecution requires trust. Attorneys need to verify AI-generated analysis before
filing with the USPTO. Glass Box metadata lets you:

* See which AI expert modules participated in the analysis
* Verify that prior art citations were checked against source documents
* Assess confidence levels per expert
* Track model versions for reproducibility
* Measure processing time

## Structure

```json theme={null}
{
  "glass_box": {
    "experts_invoked": ["claim_analyst", "prior_art_mapper", "examiner_profiler"],
    "citations_verified": true,
    "confidence_scores": {
      "claim_analyst": 0.92,
      "prior_art_mapper": 0.87
    },
    "model_versions": {
      "claim_analyst": "claude-sonnet-4-20250514",
      "prior_art_mapper": "claude-sonnet-4-20250514"
    },
    "elapsed_seconds": 12.4
  }
}
```

## Fields

| Field                | Type      | Description                                                   |
| -------------------- | --------- | ------------------------------------------------------------- |
| `experts_invoked`    | string\[] | AI expert modules that ran during analysis                    |
| `citations_verified` | boolean   | Whether prior art citations were verified against source text |
| `confidence_scores`  | object    | Per-expert confidence (0-1 scale)                             |
| `model_versions`     | object    | LLM model version used by each expert                         |
| `elapsed_seconds`    | number    | Total processing time                                         |

## Expert modules

| Expert                 | Role                                                                        |
| ---------------------- | --------------------------------------------------------------------------- |
| `claim_analyst`        | Parses claims, identifies rejected claims, maps rejection bases             |
| `prior_art_mapper`     | Maps prior art references to specific claim limitations                     |
| `examiner_profiler`    | Analyzes examiner behavior patterns (when `include_examiner_intel` is true) |
| `strategy_recommender` | Recommends per-claim response strategies                                    |

## Using Glass Box in your agent

AI agents should check Glass Box data to calibrate trust:

```python theme={null}
result = analyze_office_action(...)
glass_box = result["glass_box"]

# Only proceed with high-confidence analysis
if glass_box["confidence_scores"].get("claim_analyst", 0) < 0.7:
    # Flag for human review
    notify_user("Low confidence analysis -- recommend manual review")

# Verify citations were checked
if not glass_box["citations_verified"]:
    # Citations may be hallucinated
    notify_user("Citations not verified against source documents")
```