ai-provenance-tracker 1.0.1

Detect AI-generated content, trace origins, verify authenticity

AI Provenance Tracker - Backend

FastAPI backend for detecting AI-generated content.

Quick Start

# Create virtual environment
python -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -e ".[dev]"

# Run the server
uvicorn app.main:app --reload

API Endpoints

• POST /api/v1/detect/text - Detect AI-generated text
• POST /api/v1/detect/image - Detect AI-generated images
• POST /api/v1/detect/audio - Detect AI-generated audio (WAV)
• POST /api/v1/detect/video - Detect AI-generated video (MVP)
• POST /api/v1/batch/text - Batch text detection
• POST /api/v1/intel/x/collect - Collect X data into trust-and-safety input schema
• POST /api/v1/intel/x/collect/estimate - Estimate X request cost without external calls
• POST /api/v1/intel/x/report - Generate trust-and-safety report from normalized input
• POST /api/v1/intel/x/drilldown - Build cluster/claim drill-down + alerts dataset
• GET /api/v1/intel/x/scheduler/status - Check recurring job status
• POST /api/v1/intel/x/scheduler/run - Trigger one immediate scheduled run
• GET /api/v1/analyze/dashboard - Dashboard-ready analytics metrics
• GET /api/v1/analyze/evaluation - Calibration precision/recall trend for dashboard
• GET /api/v1/analyze/audit-events - Audit log events (HTTP + detection)
• GET /health - Health check

X Intelligence Collection

Set X_BEARER_TOKEN in .env, then either call API:

curl -X POST "http://localhost:8000/api/v1/intel/x/collect" \
  -H "Content-Type: application/json" \
  -d '{"target_handle":"@example","window_days":30,"max_posts":300,"query":"anthropic OR claudecode"}'

or use CLI utility:

python scripts/collect_x_input.py --handle @example --window-days 30 --max-posts 300 --query "anthropic OR claudecode" --output ./x_intel_input.json --show-request-estimate

Low-cost run (tight request budget):

X_MAX_PAGES=1 X_MAX_REQUESTS_PER_RUN=4 python scripts/collect_x_input.py --handle @example --window-days 7 --max-posts 60 --output ./x_intel_input.json --show-request-estimate

Cost precheck endpoint (no external X calls):

curl -X POST "http://localhost:8000/api/v1/intel/x/collect/estimate" \
  -H "Content-Type: application/json" \
  -d '{"window_days":7,"max_posts":60,"max_pages":1}'

Batch text detection:

curl -X POST "http://localhost:8000/api/v1/batch/text" \
  -H "Content-Type: application/json" \
  -d '{"items":[{"item_id":"a","text":"Sample text one..."},{"item_id":"b","text":"Sample text two..."}]}'

Dashboard metrics:

curl "http://localhost:8000/api/v1/analyze/dashboard?days=30"

Audit events:

curl "http://localhost:8000/api/v1/analyze/audit-events?limit=50"

Dashboard drill-down from normalized input:

curl -X POST "http://localhost:8000/api/v1/intel/x/drilldown" \
  -H "Content-Type: application/json" \
  --data-binary @./x_intel_input.json

Trust Report, Benchmark, Evidence Pack

Generate trust report:

python scripts/generate_x_trust_report.py --input ./x_intel_input.json --output ./x_trust_report.json

Benchmark (optional labels file):

python scripts/benchmark_x_intel.py --report ./x_trust_report.json --labels ./evidence/labels_template.json --output ./x_trust_benchmark.json

Build talent-visa evidence pack:

python scripts/build_talent_visa_evidence_pack.py --reports-glob "./x_trust_report*.json" --benchmarks-glob "./x_trust_benchmark*.json" --output-dir ./evidence

Run full pipeline:

python scripts/run_talent_visa_pipeline.py --handle @example --window-days 90 --max-posts 600 --query "anthropic OR claudecode OR claudeai OR usagelimits"

Run pipeline from pre-collected input JSON (offline mode):

python scripts/run_talent_visa_pipeline.py --input-json ./x_intel_input.json --output-dir ./evidence/runs/manual_input --run-id run_snapshot

Compare two run directories:

python scripts/compare_talent_visa_runs.py --base-run-dir ./evidence/runs/run_a --candidate-run-dir ./evidence/runs/run_b --output-json ./evidence/runs/comparisons/run_a_vs_run_b.json --output-md ./evidence/runs/comparisons/run_a_vs_run_b.md

Evaluate confidence-threshold calibration on labeled data:

python scripts/evaluate_detection_calibration.py --input ./labels_text.jsonl --content-type text --output ./calibration_text.json --register
python scripts/evaluate_detection_calibration.py --input ./labels_audio.jsonl --content-type audio --output ./calibration_audio.json --register
python scripts/evaluate_detection_calibration.py --input ./labels_video.jsonl --content-type video --output ./calibration_video.json --register

Audio/video JSONL templates: ./evidence/samples/audio_labeled_template.jsonl, ./evidence/samples/video_labeled_template.jsonl

Weekly pipeline cycle with automatic run comparison:

python scripts/run_weekly_talent_visa_cycle.py --handle @example --window-days 7 --max-posts 60 --output-dir ./evidence/runs/weekly --comparisons-dir ./evidence/runs/comparisons --summary-output ./evidence/runs/weekly/latest_summary.json

Production smoke test for all detect endpoints:

python scripts/smoke_detect_prod.py --base-url https://your-api-domain --output ./evidence/smoke/prod_detect_smoke.json

Run background worker process (scheduler + webhook retry queue):

python -m app.worker.main

Trigger a scheduler run manually:

curl -X POST "http://localhost:8000/api/v1/intel/x/scheduler/run?handle=@example"

Check scheduler status:

curl "http://localhost:8000/api/v1/intel/x/scheduler/status"

Persistence and Migrations

Runtime analysis history is persisted in analysis_records (SQLite by default). Audit events are persisted in audit_events.

alembic upgrade head

Security and Spend Controls

Configure optional API key enforcement and endpoint spend controls in .env:

• REQUIRE_API_KEY
• API_KEYS
• DAILY_SPEND_CAP_POINTS
• RATE_LIMIT_MEDIA_REQUESTS
• RATE_LIMIT_BATCH_REQUESTS
• RATE_LIMIT_INTEL_REQUESTS
• X_COST_GUARD_ENABLED
• X_MAX_REQUESTS_PER_RUN
• CONSENSUS_ENABLED
• COPYLEAKS_API_KEY
• REALITY_DEFENDER_API_KEY
• SCHEDULER_ENABLED
• SCHEDULER_HANDLES
• SCHEDULER_MONTHLY_REQUEST_CAP
• SCHEDULER_KILL_SWITCH_ON_CAP
• SCHEDULER_USAGE_FILE
• RUN_SCHEDULER_IN_API
• WORKER_ENABLE_SCHEDULER
• WORKER_DRAIN_WEBHOOK_QUEUE
• WORKER_TICK_SECONDS
• WEBHOOK_URLS
• WEBHOOK_RETRY_ATTEMPTS
• WEBHOOK_RETRY_BACKOFF_SECONDS
• WEBHOOK_QUEUE_FILE
• WEBHOOK_DEAD_LETTER_FILE
• AUDIT_EVENTS_ENABLED
• AUDIT_LOG_HTTP_REQUESTS
• AUDIT_ACTOR_HEADER

Documentation

• Swagger UI: http://localhost:8000/docs
• ReDoc: http://localhost:8000/redoc