AJT Grounded Extract

Judgment-first grounded extraction engine. Returns ACCEPT with evidence or STOP with proof. Nothing in between.

---

STOP Is Not Failure

STOP is a judgment. STOP is an audit artifact. STOP is how this system succeeds when evidence is insufficient.

Most systems explain answers. This one explains why it stopped.

---

Status

v2.1.0 — Audit-ready | Constitution: Frozen | Attack Tests: 10/10 blocked

---

Installation

pip install ajt-grounded-extract

Zero dependencies. Pure Python stdlib.

---

Core Principle

Extract structured data only when it can be proven; otherwise stop—and prove that you stopped.

Most systems explain answers. This one explains why it stopped.

---

Philosophy: STOP-first

• This project does not aim to extract everything.
• Extraction occurs only when evidence is sufficient.
• When evidence is insufficient, the system stops and proves why.
• Evidence Integrity > Recall: Only extract values with verifiable document evidence
• Default: STOP: When evidence is insufficient, conflicting, or missing → stop extraction
• Negative Proof: Every STOP includes explicit reason + preserved artifacts
• No Fine-tuning: Rule-based + LLM extraction without training pipelines
• Local Execution: Runs entirely on local machine

---

What This Is NOT

This system is blocked-by-design, not secure-by-claim.

• ❌ Multi-domain rule engine
• ❌ Enterprise extraction with thresholds
• ❌ Training/fine-tuning pipeline
• ❌ High-recall extraction system
• ❌ "Secure" or "safe" (we demonstrate how attacks are blocked, not claim safety)

What we guarantee:

• ✅ Stoppability (DEFAULT: STOP)
• ✅ Traceability (decision_maker required)
• ✅ Audit trail (write-once logs)

---

Architecture

Document → Ingest → Extract → Ground → Judge → Archive
           ↓        ↓         ↓        ↓        ↓
           Hash     Candidates Evidence STOP?   Artifacts

Pipeline Stages

1. Ingest: Load document, compute hash, build line index
2. Extract: Find candidate values (rule-based or LLM)
3. Ground: Map each value to exact document span (quote + offsets)
4. Judge: STOP-first decision: ACCEPT | STOP | NEED_REVIEW
5. Archive: Write-once artifacts with timestamps + integrity hashes

Decision Taxonomy

• ACCEPT: Evidence found, confidence sufficient, integrity verified
• STOP: No candidates, conflict, low confidence, or integrity failure
• NEED_REVIEW: Edge cases requiring human judgment

---

Quick Start

Run Extraction

# ACCEPT case (has clear "Effective Date: 01/15/2025")
python run.py examples/accept_example.txt

# STOP case (no explicit effective date)
python run.py examples/stop_example.txt

View Results

Open generated HTML viewer:

open viewer/accept_example_viewer.html
open viewer/stop_example_viewer.html

---

Output Format

JSON Result

{
  "field_name": "effective_date",
  "decision": "ACCEPT",
  "value": "01/15/2025",
  "evidence": {
    "quote": "01/15/2025",
    "start": 245,
    "end": 255,
    "line": 12,
    "context": "...Effective Date: 01/15/2025..."
  },
  "confidence": 0.9
}

STOP Event

{
  "field_name": "effective_date",
  "decision": "STOP",
  "value": null,
  "stop_reason": "no_candidates_found",
  "stop_proof": {
    "searched": true,
    "candidates_found": 0
  }
}

---

HTML Viewer Features

• Evidence Highlighting: Green (ACCEPT) / Red (STOP)
• Navigation Sidebar: Jump to extracted fields
• "Why Stopped" Panel: Explicit reasons with proof artifacts
• Offset Mapping: Click evidence span → see exact document location

---

Directory Structure

ajt-grounded-extract/
├── schema/              # Field definitions
├── engine/              # Core extraction modules
│   ├── ingest.py
│   ├── extract.py
│   ├── ground.py
│   ├── judge.py
│   └── archive.py
├── viewer/              # HTML viewer generator
├── evidence/            # Write-once artifacts (JSONL + manifests)
├── examples/            # Demo documents
└── run.py               # CLI entry point

---

Evidence Requirements

All extractions must satisfy:

• ✅ require_exact_quote: Value must appear verbatim in document
• ✅ require_offset_mapping: Quote mapped to byte offsets
• ✅ stop_on_conflict: Multiple conflicting values → STOP
• ✅ min_confidence: Below threshold → STOP

---

Acceptance Criteria

• Demo shows at least one ACCEPT and one STOP
• STOP includes explicit reason and preserved artifacts
• Viewer navigates evidence spans correctly
• Non-goals stated explicitly

---

Regulatory Mapping & Review

This system includes industry-specific regulatory risk mappings for:

• Financial Services — Authorization scope, customer isolation, advisory vs execution separation
• Healthcare — Patient data isolation, complete clinical evidence requirements, clinician traceability
• Legal Practice — Attorney responsibility, client-matter isolation, conflict-of-interest prevention

Navigation: See REGULATORY_REVIEW_GUIDE.md for audience-specific entry points.

Key documents:

• REGULATORY_META_MAP.md — Cross-industry risk-control mappings
• docs/REG_MAP_FINANCE.md — Financial services mapping
• docs/REG_MAP_HEALTHCARE.md — Healthcare mapping
• docs/REG_MAP_LEGAL.md — Legal practice mapping
• COMPLIANCE_GUIDE.md — Audit artifact generation
• ATTACK_TEST.md — Adversarial verification results

Principle: This project demonstrates how specified risks are blocked. It does not claim regulatory compliance.

---

Reference

Normative Specification

This implementation follows the AJT (Adjudicative Judgment Trace) constitutional framework:

• Spec Repository: ajt-spec — Normative rules and judgment structure
• Reference Implementation: This repository (ajt-grounded-extract) — Executable proof of concept

Relationship:

• ajt-spec: Constitutional rules (what must be proven)
• ajt-grounded-extract: Execution + case law (how it's proven in practice)

---

Motivation

Motivated by ajt-negative-proof-sim (sealed reference).

Core principle: Prove extraction succeeded OR prove why you stopped.

---

License

MIT