CanaryObservatory
 

Living laboratory / research notebook

Experiments

A hypothesis is an invitation to look carefully, not a finding. These protocols record what we need, what might fool us, and what would change our minds.

Historical replay →

Research maturity is not a confidence score. Collector availability, data freshness, baseline readiness, statistical deviation and hypothesis support are separate questions. No stage here measures real-world risk.

The Pizza Hypothesis

Hypothesis

Unusual late-night restaurant activity near major government facilities systematically precedes major geopolitical announcements.

A late dinner is not, by itself, a geopolitical briefing.

Data required

  • Hourly busyness near the Pentagon and other facilities
  • A dated list of major announcements
  • Typical busyness by hour of week

Confounders

  • Sporting events and holidays
  • Restaurant openings and closures
  • Weather
  • Selection bias in remembered anecdotes

Proposed methodology

Compute busyness z-scores by hour of week. Compare the rate of z ≥ 2 nights in the 48 hours before announcements with the rate on all other nights, using only data that was available at the time.

Evidence so far / historical evaluation

This experiment is Proposed. It has produced no evidence.

No data collected. Historical claims are anecdotal and have not been tested here.

Next-stage requirement

Begin timestamped hourly busyness collection, retain availability times, and define the announcement list and hour-of-week comparison protocol before evaluation.

What would disprove it?

In an adequately powered, held-out comparison, elevated busyness is no more frequent before announcements than on matched control nights after accounting for confounders.

These are proposed rejection criteria, not observed results. An underpowered or incomplete test would remain inconclusive.

Conclusion

None. There is not yet sufficient evidence to draw one.

The Exodus Hypothesis

Hypothesis

Unusual private aviation departures from a region precede major regional crises.

Data required

  • Business-jet departures by airport
  • Dated crisis onsets
  • Seasonal travel baselines

Confounders

  • Holidays and conferences
  • Weather diversions
  • Incomplete ADS-B coverage
  • Privacy-filtered aircraft

Proposed methodology

Compare departure anomalies in the 72 hours before crisis onsets with matched control periods.

Evidence so far / historical evaluation

This experiment is Proposed. It has produced no evidence.

No data collected.

Next-stage requirement

Begin storing airport-level business-jet departures and their availability times; define crisis onsets, seasonal controls and coverage exclusions before evaluation.

What would disprove it?

In an adequately powered, held-out comparison, departure anomalies do not precede crisis onsets more often than matched control periods after accounting for coverage and travel patterns.

These are proposed rejection criteria, not observed results. An underpowered or incomplete test would remain inconclusive.

Conclusion

None. There is not yet sufficient evidence to draw one.

The Connectivity Hypothesis

Hypothesis

Particular patterns of national internet disruption provide early evidence of political instability or conflict.

Data required

  • IODA country outage events (collecting)
  • Dated political events (e.g. ACLED, curated list)
  • Power-outage data to separate causes

Confounders

  • Power failures
  • Cable cuts
  • Scheduled exam shutdowns
  • Measurement artefacts

Proposed methodology

Classify outage events by duration and data-source agreement. Measure how often multi-source national outages precede or coincide with recorded events, against a base rate.

Evidence so far / historical evaluation

Outage events are being stored since collection began. No outcome dataset is linked yet, so no evaluation has been run.

Next-stage requirement

Link a dated outcome dataset and power-outage context to stored outage events; freeze classification rules, base rates and temporal matching before moving to Evaluating.

What would disprove it?

In an adequately powered, held-out comparison, the specified multi-source outage patterns are no more associated with recorded political events than the base rate after accounting for alternative causes.

These are proposed rejection criteria, not observed results. An underpowered or incomplete test would remain inconclusive.

Conclusion

None. There is not yet sufficient evidence to draw one.

The Maritime Hypothesis

Hypothesis

Abrupt shipping diversions away from chokepoints anticipate escalating geopolitical tensions.

Data required

  • Daily chokepoint transit counts
  • Dated security incidents
  • Freight-rate and seasonal context

Confounders

  • Freight market cycles
  • Canal maintenance
  • Weather
  • Insurance-driven rerouting

Proposed methodology

Detect structural breaks in transit counts and test their lead time relative to incident dates.

Evidence so far / historical evaluation

This experiment is Proposed. It has produced no evidence.

No data collected.

Next-stage requirement

Begin timestamped chokepoint transit collection and assemble dated incidents, freight rates and seasonal context; specify structural-break and lead-time tests before evaluation.

What would disprove it?

In an adequately powered, held-out comparison, transit breaks do not anticipate security incidents more often than control periods once market, maintenance, weather and insurance effects are considered.

These are proposed rejection criteria, not observed results. An underpowered or incomplete test would remain inconclusive.

Conclusion

None. There is not yet sufficient evidence to draw one.