tylergarlick/abraxas/scribe

Scribe System - Source Provenance & Citation Integrity

Traditional: Record-keeper, scribe

Purpose: Track where claims originated, validate citation integrity, and monitor source reliability over time. Detects when sources are retracted, corrected, or downgraded in reliability.

Overview

Scribe provides citation integrity for epistemic claims by:

• Capturing source metadata at claim creation
• Scoring source reliability dynamically
• Checking links for rot/retractions
• Alerting when sources degrade

Architecture (L1-L4 Pattern)

S1: Source Capture Engine

File: source_capture.py

Captures and normalizes source metadata at claim creation.

Features:

• URL normalization (remove tracking params)
• Source type classification (academic/news/blog/social/official)
• Initial reliability scoring
• Persistent storage

Source Types:

Key Classes:

• SourceMetadata — Source data structure
• SourceCapture — Capture engine

S2: Reliability Scorer Engine

File: reliability_scorer.py

Dynamically adjusts source reliability based on track record.

Reliability Factors:

Track Record Tracking:

• Records accuracy outcomes per claim
• Calculates accuracy rate over time
• Recalculates scores periodically

Key Classes:

• ReliabilityUpdate — Score change record
• ReliabilityScorer — Scoring engine

S3: Link Checker Engine

File: link_checker.py

Monitors sources for link rot and retraction notices.

Link Statuses:

Detection Patterns:

• Retraction: "retracted", "withdrawn", "expression of concern"
• Correction: "corrected", "amended", "erratum", "updated"

Scheduling:

• Active sources: check every 30 days
• Corrected/retracted: check every 1 day
• Broken links: check every 7 days

Key Classes:

• LinkCheckResult — Check outcome
• LinkChecker — Async checking engine

S4: Provenance Tracker Engine

File: provenance_tracker.py

Maintains provenance graphs and generates alerts.

Alert Types:

Features:

• Citation chain tracking (claim → sources)
• Impact analysis (affected claims)
• Alert callbacks for real-time notification
• Alert acknowledgment workflow

Key Classes:

• ProvenanceAlert — Alert record
• CitationChain — Claim-source linkage
• ProvenanceTracker — Tracking engine

Commands

/scribe capture {url} [--title {title}] [--author {author}]
    Capture source metadata from URL
 
/scribe reliability {source_id}
    Get reliability info for a source
 
/scribe check {source_id|--all}
    Check source link status
 
/scribe alerts [--unacknowledged]
    Show provenance alerts
 
/scribe acknowledge {alert_id}
    Acknowledge an alert
 
/scribe provenance {claim_id}
    Show provenance for a claim
 
/scribe affected {source_id}
    Show claims affected by a source
 
/scribe stats
    Show system statistics

Integration with Existing Systems

Chronos Integration

Chronos tracks claims over time; Scribe tracks sources over time. Together they provide full temporal provenance coverage.

Logos Integration

Logos decomposes claims into atomic propositions; Scribe can track sources for each atom independently.

Pheme Integration

Pheme verifies claims against sources; Scribe provides the source infrastructure for Pheme to use.

Janus Integration

Janus labels claims; Scribe adds source provenance metadata to those labels.

Usage Examples

Example 1: Capture Sources

from scribe import SourceCapture
capture = SourceCapture()
source_id = capture.capture_source(
    url="https://arxiv.org/abs/2026.12345",
    title="Example Paper",
    author="John Doe"
)

Example 2: Check Reliability

from scribe import ReliabilityScorer
scorer = ReliabilityScorer(capture)
info = scorer.get_source_reliability(source_id)
print(f"Score: {info['current_score']}")
print(f"Accuracy rate: {info['accuracy_rate']}")

Example 3: Check Links

import asyncio
from scribe import LinkChecker
checker = LinkChecker(capture)
result = asyncio.run(checker.check_link(source_id))
print(f"Status: {result.status}")
print(f"Retracted: {result.retraction_detected}")

Example 4: Track Provenance

from scribe import ProvenanceTracker
tracker = ProvenanceTracker(capture, scorer, checker)
# Register claim with sources
chain = tracker.register_claim_sources(
    claim_id="claim_001",
    source_ids=[source_id]
)
# Run checks and get alerts
alerts = tracker.run_all_checks()
for alert in alerts:
    print(f"Alert: {alert.alert_type} - {alert.severity}")

Example 5: Alert Callback

def handle_alert(alert):
    print(f"ALERT: {alert.message}")
    print(f"Affected claims: {alert.affected_claims}")
tracker.register_alert_callback(handle_alert)

Testing

Run tests:

cd /abraxas/systems/scribe
python -m pytest test_scribe.py -v

Test Coverage:

• Source capture (URL normalization, classification)
• Reliability scoring (updates, track record)
• Link checking (async, status detection)
• Provenance tracking (alerts, chains)

Storage

Scribe persists data to ~/.abraxas/scribe/:

• sources.json — Source metadata
• reliability_data.json — Score history & track records
• link_checks.json — Check history & schedules
• provenance.json — Citation chains & alerts

Performance Considerations

• Link checking is async (aiohttp)
• Batch checks for multiple sources
• Scheduled checks avoid redundant polling
• For production: consider background job queue

Future Enhancements

1. Retraction API integration — Crossref, PubMed retraction APIs
2. Wayback Machine fallback — Archive.org for rotted links
3. Domain reputation database — External reputation scoring
4. Citation graph visualization — Interactive provenance graphs
5. Integration with retrieval MCP — Auto-capture during retrieval

Version

1.0.0 — Initial implementation (Phase 2, 2026-03-19)

MCP tools

ingest_fragment, source_capture, source_get, source_reliability_update, scribe_link_check (Bun + TypeScript server, mcp/src/tools/scribe.ts). The Python modules and pytest instructions above are superseded by mcp/tests/scribe.test.ts.