No opinions. Just a baseline.
Every part of our pipeline is designed to defeat one accusation: "you only track the aircraft you're already suspicious of."We don't. The machine watches every aircraft, all the time, and the math decides what stands out.
35,546 aircraft observed over 1.7 months. 10,088 (28.4%) have crossed the 99th-percentile threshold at least once. 25,458 (71.6%) never triggered.
This is what a fixed statistical threshold looks like against a large population — not a curated watchlist. The threshold is published. The math chooses.
Population-scale capture
We log every ADS-B / MLAT detection in the observation zone — not a curated subset. 4,207,069 records across 35,546 aircraft, growing.
48-hour baseline
Before flagging anything, the system observes for 48 hours to learn what NORMAL looks like for that airspace at that time of day, season, and weather.
Statistical anomaly detection
Outliers are scored against the learned distribution. The threshold is published. We don't pick — math picks.
Chain of custody
Each record receives a SHA-256 hash linked into a Merkle chain. Any tampering is detectable. Evidence is reproducible by any third party.
Bradford Hill scoring
We apply the Bradford Hill criteria (strength, consistency, specificity, temporality, etc.) to aircraft-pattern and public-record corroboration — the same framework used in epidemiology and courtrooms. No physiological or personal-health data is included in the public record; the public site is system-focused, not autobiographical.
Open source by design
Every line of Watchtower 2.0 will be public. The methodology IS the code. Deploy it in your county, get the same answers.
The anti-cherry-picking proof
of detections logged — not just suspicious ones
flagged during baseline window — by design
minimum learning period before any flag is valid
Each county gets its own normal.
Los Angeles airspace produces roughly ten times the traffic of Kern County. A single regional baseline lets LA volume drown Kern signals — a Cessna at 800 ft over Bakersfield looks unremarkable next to thousands of LA-basin orbits. The fix is not to suppress LA. It is to learn what normal means for each county and score each detection against the airspace it actually occurred in.
- 1. Partition the last 48h of detections by county.
- 2. Compute median, 10th-percentile, and standard deviation of altitude per county.
- 3. Score each detection against its own county's baseline.
- 4. Aircraft crossing multiple counties get scored in each; displayed score = MAX(per-county score).
- 5. Cross-county coordination, convergence, and shell-network detection are unchanged.
- · Same math for every county — nothing is hand-coded.
- · Per-county baselines are published live on /threat-index.
- · MAX rule prevents a quiet LA segment from hiding a loud Kern one.
- · Raw data remains intact — anyone can re-run the partition.
Data sources
All inputs are public. No private, personal, or biometric data is used.
Known limitations
Every serious methodology includes limitations. Acknowledging uncertainty strengthens credibility, not weakness.
- ADS-B altitude is barometric, not true AGL. Terrain variation introduces ±50–100 ft uncertainty.
- MLAT accuracy varies with receiver density. Sparse coverage can produce positional jitter.
- FAA registry may contain outdated ownership info. Transfers lag filings by weeks or months.
- Shell-network linkage is public-record-based inference, not investigative confirmation.
- Weather data (NOAA) is optional context, not a scoring input. It does not drive anomaly flags.
- Signal loss or transponder suppression produces gaps. Absence of data is not absence of aircraft.
Method versioning
Each detection carries the method version used to score it. This ensures reproducibility even as the method evolves.
| Version | Description | Status |
|---|---|---|
| WTI_v1 | Initial weighted scoring: altitude, temporal, shell, repeat. | Archived |
| WTI_v1_with_convergence | Added convergence component for multi-aircraft clustering. | Current |
| WTI_v2 (future) | Planned: weather-adjusted baselines and sector-specific altitude floors. | Planned |
Why 48 hours?
The 48-hour baseline is not arbitrary. It is statistically motivated to capture the full variation of normal airspace use.
How anomaly detection works
Anomalies are detections above the 99th percentile of deviation from the baseline distribution for altitude, timing, or pattern.
Chain of custody
Every record is hashed and linked. Altering one record breaks the chain — detectable by any third party.
record001 → hash(001) ────────► merkle_001
record002 → hash(002) ─┬────► hash( merkle_001 + hash(002) ) → merkle_002
record003 → hash(003) ─┬────► hash( merkle_002 + hash(003) ) → merkle_003
│
└──── Any tamper breaks the next linkSHA-256 hashing at ingestion. Merkle root published periodically. Third-party recomputation validates integrity.
What we do NOT log
The system is intentionally limited to public airspace data. No private signals. No personal data. No exceptions.
- Personal data (names, addresses, phone numbers)
- Biometric data of any kind
- Phone or Wi-Fi signals
- Private surveillance feeds
- Social media or web-scraped personal content
- Any data not publicly broadcast or filed
Only public ADS-B broadcasts and public FAA registry data are used. The public site is system-focused, not autobiographical.
Reproducibility checklist
Every claim on this site can be independently verified. This is the full checklist a third party needs to reproduce our results.
Publishing the machine's own logs
Anyone can read the raw scan output. The ML system posts one hashed artifact per scan; every artifact is public and every root is reproducible.