Sentinel reads the same public radio signals anyone with a thirty-dollar antenna can read. Then it does the work of paying attention.

The signals

Sentinel watches four things, and only four things. Each is computed from public data, each has a formal definition, and each has known blind spots. The four signals are deliberately independent: a real event tends to move several of them at once; a data artifact tends to move only one.

Sudden surge

Surge measures deviation from the baseline take-off rate for a region, per hour-of-week. The question it answers is narrow: are meaningfully more private aircraft leaving this region right now than this region's own history says is normal for this hour, this weekday, this season? The baseline window is 90 days.¹

Known limits. Surge is weakest where baselines are sparse: small regions with few flights per day produce noisy baselines, and airports that recently opened or changed role have no history at all. In both cases the gauntlet's data-trust gate suppresses the signal rather than guessing.

Tight clustering

Clustering measures whether the departures that make up a surge are bunched in time more tightly than coincidence allows. The exposed window is 47 minutes of lookback — not a round number, because it was not chosen by hand. It is the 99.6th percentile of inter-departure gaps for same-cohort aircraft in the 90-day archive.²

Known limits. Large hubs at peak hour bunch departures for boring reasons — slot scheduling, weather windows, curfew edges. Clustering alone is therefore never sufficient; it contributes 25% of the composite and cannot cross an alert threshold on its own.

Going everywhere

Scatter measures destination entropy: whether the cohort is dispersing in many directions rather than converging on one place. Convergence is almost always explainable — a conference, a funeral, a final. Dispersal from a common origin, by unrelated parties, inside a tight window, is the rarer and more informative shape.³

Known limits. Small cohorts saturate quickly — with three aircraft, "everywhere" and "coincidence" are hard to distinguish, which is one reason the cohort floor for any published receipt is higher than three.

Independent decisions

Independence asks whether the cohort's members actually decided separately. Seven jets owned by one charter operator is one decision; seven jets owned by seven unrelated principals is seven. We resolve aircraft to beneficial-owner classes through registries and state business filings, unwinding shell entities where the paper trail is public.⁴

Known limits. Recently restructured ownership trees lag the public record by days to weeks; an aircraft sold last Tuesday may still resolve to its old owner. Trust-heavy jurisdictions (Delaware, South Dakota, Nevada, Wyoming) resolve more slowly and sometimes only to the trust, not through it. When resolution confidence is low, independence is scored conservatively downward, never upward.

The sources

Every input to Sentinel is listed below — there are no others. The licensing column is written for the person whose job is to check it. Sortable by column.

The full source registry, with rate limits and ToS links, is also available as JSON at sentinel.app/method/sources.json.

The model

Composite scoring

The four signals combine into a single Sentinel Score on [0, 1]:

In plain English: a surge that is also tightly bunched matters most; where the cohort went and whether its members are unrelated each refine the picture. Below 0.40 the system stays quiet; 0.40–0.85 is a Heads Up; at or above 0.85, and only after the gauntlet and a human review, a Red Alert.⁵

The weights are not aesthetic. They were calibrated by maximizing precision at fixed recall on a manually labeled historical set of n = 240 events spanning 2023-Q4 through 2026-Q2, each labeled by two reviewers with disagreements adjudicated by a third. The labeled set, with identifying details removed, accompanies the classifier repository.⁶

The gauntlet

A score is not an alert. Every candidate above threshold runs through five rejection gates, in fixed order. In the last 90 days, the gauntlet rejected 99.87% of candidate alerts. The five gates rejected, in order of frequency: known-event (62%), data-trust (18%), same-company (11%), known-cause (7%), weekend (2%).⁷

Cross-modal corroboration

Aircraft movement is cross-checked against vessel movement in the window T−12h to T+1h: a cohort scatter accompanied by unusual yacht departures from the same region is materially stronger evidence than either alone. Both are then checked against news lag in the window T+0 to T+24h — if the press never catches up, the event is either small or was never real, and that outcome is recorded against the receipt either way.

The false-positive rate with and without vessel corroboration on the labeled set: [placeholder — research pending; published with the Q4 2026 calibration]. We will not state this number until the labeled set is large enough that it means something.

The Pulse layer

Pulse, the daily divergence surface, consumes the same four signals plus a news-heat layer and classifies every region and sector into exactly four states: red (news ≥ 0.55 and movement ≥ 0.55), white (news ≥ 0.55, movement < 0.35), sentinel (news < 0.35, movement ≥ 0.55), and quiet (everything else, including the deliberate dead zones between thresholds).⁸

Divergence states are coarser than the underlying signals, and that is deliberate. Pulse is for situational awareness, not precision; the receipt — with its full signal values, gauntlet record, and hash — remains the only artifact Sentinel asks anyone to rely on.

Calibration

Weights, gate rules, the clustering window, and corroboration windows are recalibrated quarterly against the growing labeled set, reviewed by the research team named at the end of this document, and frozen for the quarter. Every change — including changes that made the system look worse — is logged publicly at /method/calibration-log. A receipt is always evaluated under the calibration that was live when it was published; that is what makes old receipts reproducible.

Reproducibility

The worked example below reproduces receipt SF-2026-06-08-0731 — the seven-aircraft Bay Area scatter of June 8, 2026 — from public records alone.

The disclosures

What Sentinel does not claim:

1. Sentinel does not predict the future. It reports what has happened, after it has happened, with provenance.
2. Sentinel does not provide investment advice. Receipts describe public observable events; what a recipient does with that information is their decision and their responsibility.
3. Sentinel does not identify individuals. We track cohorts of ≥ 10 members; no receipt, public page, or alert names a natural person.
4. Sentinel does not access private data. Every signal we use is broadcast under federal mandate or sourced from public records.
5. Sentinel does not surveil its users. We do not sell user data, embed third-party tracking, or build a per-user behavior model.
6. Sentinel does not guarantee delivery latency. Our published median news-lead time is a historical average, not a service-level commitment.
7. Sentinel does not claim 100% recall. Some real events will not generate receipts; some receipts will, in retrospect, be false positives. The published precision and recall are both quarterly-recalibrated.
8. Sentinel does not operate in classified domains. We accept no classified data and produce no classified output.
9. Sentinel does not take money from any institution it observes. Family offices, hedge funds, and corporate entities may subscribe to receive receipts; none may subscribe to influence which receipts are produced.
10. Sentinel does not negotiate the contents of its receipts. If a receipt is published, it has passed the gauntlet; we will not unpublish a passing receipt under pressure.
11. Sentinel does not promise the model is static. Signal weights, gauntlet rules, and corroboration windows change quarterly with calibration; all changes are logged.
12. Sentinel does not exempt itself from these disclosures. If you ever see Sentinel violate one, email accountability@sentinel.app.

The errata

Every published correction or retraction since launch, served live from the same archive the receipts come from.

Bibliography & prior art

• RTCA. DO-260B: Minimum Operational Performance Standards for 1090 MHz Extended Squitter ADS-B and TIS-B. RTCA, 2009. rtca.org
• Schäfer, M., Strohmeier, M., Lenders, V., Martinovic, I., Wilhelm, M. "Bringing Up OpenSky: A Large-scale ADS-B Sensor Network for Research." IPSN 2014. opensky-network.org
• Olive, X. "traffic, a toolbox for processing and analysing air traffic data." Journal of Open Source Software 4(39), 2019. doi:10.21105/joss.01518
• Olive, X., Basora, L. "Detection and identification of significant events in historical aircraft trajectory data." Transportation Research Part C 119, 2020. doi:10.1016/j.trc.2020.102737
• Strohmeier, M., Smith, M., Lenders, V., Martinovic, I. "The Real First Class? Inferring Confidential Corporate Mergers and Government Relations from Air Traffic Communication." IEEE EuroS&P 2018. ieeexplore.ieee.org
• Yermack, D. "Flights of fancy: Corporate jets, CEO perquisites, and inferior shareholder returns." Journal of Financial Economics 80(1), 2006. doi:10.1016/j.jfineco.2005.05.002
• FAA. Limiting Aircraft Data Displayed (LADD) program documentation. faa.gov/pilots/ladd
• FAA. Privacy ICAO Aircraft Address (PIA) program. faa.gov
• Silverman, C. (ed.). Verification Handbook: A Definitive Guide to Verifying Digital Content for Emergency Coverage. European Journalism Centre, 2014. verificationhandbook.com
• Bellingcat. Bellingcat's Online Investigation Toolkit. bellingcat.com/resources
• Leetaru, K., Schrodt, P. "GDELT: Global Data on Events, Location, and Tone, 1979–2012." ISA Annual Convention, 2013. gdeltproject.org
• Common Crawl Foundation. Common Crawl news corpus. commoncrawl.org
• GLEIF. Legal Entity Identifier (LEI) system documentation. gleif.org
• ITU. Recommendation ITU-R M.1371-5: Technical characteristics for an automatic identification system (AIS). 2014. itu.int
• Denev, A., Amen, S. The Book of Alternative Data. Wiley, 2020. wiley.com

Authors and reviewers

External review board: this document has been reviewed by [placeholder — academic affiliations, journalism organizations, and policy reviewers are named at sentinel.app/method/reviewers when confirmed].