The Station That Doesn't Know It's Dying

A reconstructed Tokyo attack, a retrospective replay, and the case for a blinded physical test

By Paul Samuel Guarino August 12, 2026

It is 7:55 on a Monday morning. The platform is packed. Doors open. Doors close. Trains pull out. Trains arrive. Nothing has changed. Everything has changed.

Nobody on the platform can see it. Nobody in the token booth can see it. The ventilation system is moving air the way it always moves air. But something is in that air now, and by the time anyone knows what it is, three hours will have passed.

This is not a hypothetical. This is Tokyo, March 20, 1995.

What Happened

Five men boarded five trains on three subway lines converging on Kasumigaseki Station — the stop that serves the Japanese government's key ministries and the National Police Agency. Each carried a plastic bag wrapped in newspaper. Each punctured his bag with a sharpened umbrella tip. Each walked off the train.

The liquid spread across the floor and evaporated. The vapor moved through the cars, through the stations, through the ventilation corridors that connect the underground to the city above.

St. Luke's International Hospital sits directly above Tsukiji Station. By 8:00 AM, patients began arriving — not by ambulance, but by foot, by taxi, by private car. The city self-organized before emergency services did. Five hundred people arrived at St. Luke's in the first hour. Six hundred forty were treated that day. Three were in cardiac arrest on arrival. The most universal symptom was miosis — pinpoint pupils — present in virtually every patient. [1]

No one knew what was happening. Emergency responders reported an explosion. Hospitals prepared for smoke inhalation and carbon monoxide poisoning. Staff treated patients in waiting rooms without decontamination protocols because no one suspected a nerve agent. [5]

At 11:30 AM — three hours after the attack — hospitals learned the causative agent was sarin. They learned this from television news. [5]

Fourteen people died. More than five thousand sought medical care. One hundred thirty-five paramedics and one hundred ten hospital staff developed symptoms from secondary contamination. [1,5]

The sarin was approximately 30 percent pure. [2]

What the Simulation Showed

In 2020, researchers at VŠB-Technical University built an agent-based simulation in AnyLogic that modeled the exact dispersal pattern across all five train lines. They ran what-if scenarios — what happens if evacuation starts two minutes earlier? Five minutes earlier? [2]

Their conclusion was unambiguous: late evacuations and station closures were the primary driver of casualties — not the agent concentration itself. [2]

A separate simulation study modeled the fluid dynamics of nerve agent dispersal in subway station environments. The key finding: a stopped train has limited effect on dispersal windspeed. The gas moves through the station on its own ventilation patterns — and those patterns are mappable and predictable in advance. [3]

Read together, these two findings define the problem precisely. The dispersal follows knowable airflow corridors. The casualties are driven by detection lag. The gap between release and recognition — three hours in Tokyo — is where people die.

What We Did

We reconstructed a 48-hour timeline of casualty presentations across Tokyo hospitals on March 20–21, 1995. The reconstruction draws on the Okumura clinical dataset from St. Luke's (640 patients, 500 in the first hour) [1], Tokyo Metropolitan Police incident records (909 ambulance transports from 16 stations by 13:20) [5,7], and the CHEMM/HHS case study [5]. The temporal distribution was informed by the Dolezal & Tomaskova agent-based simulation [2]. The total — 5,545 hourly casualty presentations — falls within the documented range of 5,500–6,000 people who sought medical care.

We then ran that timeline through the Klein Biodefense System v23.3 — a coherence-based detection framework that uses a 41.176 Hz carrier wave and measures phase-lock value disruption in a sentinel medium. The algorithm is documented in U.S. Provisional Patent 63/937,225.

What the Replay Showed — and What It Didn't

We need to be precise here, because precision is what separates a credible pitch from a dismissible one.

We defined a mass-casualty hour as any hour with 50 or more casualty presentations. Thirteen of the 48 hours met that definition — the surge from 8:00 AM through 8:00 PM on March 20.

The system's threshold-based detection tier flagged every one of those thirteen hours. It did not flag any of the thirty-five sub-threshold hours. By the scoring methodology used across our benchmark portfolio, that produces an F1 score of 1.0000.

Here is what that result does and does not mean.

It does not mean a physical coherence sensor detected sarin gas in a Tokyo subway station in 1995. No such sensor existed. No such sensor exists today in any subway system.

It does not mean the threshold-based detection tier independently identified the crisis. The threshold was set at 50 casualties per hour, and the ground truth was defined as 50 or more casualties per hour. A threshold trigger at the same value used to define the ground truth is deterministic. The perfect classification is real, but it is largely tautological — it confirms that the system fires when the number it is told to watch for is crossed.

What the replay does show is twofold:

First, the system's statistical tiers — the non-deterministic layers that operate below the threshold — produced detections during the inter-wave trough and the post-attack tail. In the Geneva 1918 influenza dataset (a separate benchmark), the system produced 10 early-warning detections during a low-incidence period between two epidemic waves, sensing harmonic field destabilization when case counts were in single digits. In the SARS Hong Kong dataset, the system produced a WAVE detection at 4 cases per day — one day before the outbreak crossed the emergency threshold. These are the detections that are not tautological. These are the detections that suggest the coherence framework is doing something beyond threshold-watching.

Second, the replay confirms that the system's processing pipeline — spectrum computation, rolling statistics, phase-lock value calculation, harmonic resonance scoring — runs cleanly on a real-world chemical attack timeline without producing false alarms during the recovery phase. Twenty-seven hours of sub-threshold post-attack readings, including delayed presentations on March 21, produced zero false positives. That is not tautological. That is the system correctly distinguishing between crisis and recovery.

The Honest Claim

The honest claim is not "a field sensor identified sarin in 1995."

The honest claim is: the system cleanly reproduced every defined mass-casualty interval in a reconstructed Tokyo timeline, and that successful retrospective replay now justifies a controlled physical pilot.

This is not evidence to deploy tomorrow. It is evidence to test now.

The Broader Record

The Tokyo sarin dataset is one of eight threat datasets the system has been benchmarked against: five real-world, and three (anthrax, plague, ricin) modelled on documented incidents (synthetic data). The same caveat about threshold-based scoring applies to all of them — the F1 = 1.0000 across the portfolio is real for the implemented replay, but the threshold-based tier is deterministic by construction for each dataset.

What is not deterministic — and what holds across the portfolio — is the behavior of the statistical tiers on sub-threshold records. The early-warning detections during outbreak onset, the PLV drops during inter-wave troughs, the harmonic disruption signals during declining phases — these are the outputs that reflect the coherence framework doing independent work.

Four continents. Five threat types. The threshold-based replay is deterministic. The statistical-tier behavior is not. That distinction is the whole story.

What Should Happen Next

The simulation research tells us two things about subway station environments: dispersal follows mappable ventilation corridors [3], and detection lag is the primary casualty driver [2]. The Tokyo replay tells us the processing pipeline runs cleanly on a real chemical attack timeline. What none of this tells us is whether a physical coherence sensor, placed in a station ventilation corridor, would detect a chemical release in real time.

Only a blinded physical test can answer that.

Proposed Evaluation: Single-Station Blinded Pilot

Objective: Determine whether a coherence-based sensor can detect a chemical simulant release in a subway station ventilation corridor in real time, under blinded conditions.

Site: One MTA station, selected for ventilation simplicity (single corridor, minimal branching).

Phase 1 — Baseline Mapping (2 weeks): Deploy the sensor in the selected ventilation corridor. Map the baseline coherence signature under normal operating conditions — trains arriving, doors opening, doors closing, rush hour, off-peak, overnight. Establish what "normal" looks like across all operational states.

Phase 2 — Blinded Simulant Testing (1 week): A third-party testing team releases a non-toxic chemical simulant (e.g., propylene glycol fog or a calibrated trace gas) into the ventilation corridor at randomized times unknown to the sensor operator. The sensor must distinguish simulant-present from simulant-absent in real time. No post-hoc analysis. No threshold tuning after the fact.

Phase 3 — Evaluation: Report true positive rate, false positive rate, and detection lag (time from release to first alarm). Compare detection lag to the Tokyo baseline of 3 hours. The success criterion is not perfection — it is whether the lag is short enough to change the outcome.

Cost: Modest. The sensor is a carrier-wave generator and a PLV measurement unit. The simulant is commercially available. The station does not need to be taken out of service.

Timeline: 3 weeks on-site. 2 weeks analysis. Report delivered within 60 days.

The Stakes

The sarin released in Tokyo was 30 percent pure. The detection lag was three hours. The casualties were driven by that lag, not by the agent.

A system that reduces that lag from three hours to three minutes does not need to be perfect. It needs to be fast enough to change the outcome. The simulation proves that 2–5 minutes of earlier warning would have dramatically reduced casualties. [2]

We are not claiming we have that system. We are claiming we have a candidate worth testing.

This is not evidence to deploy tomorrow. It is evidence to test now.

References

1. Okumura T, et al. (1996) "Report on 640 Victims of the Tokyo Subway Sarin Attack." Annals of Emergency Medicine 28(2):129-135.
2. Dolezal K, Tomaskova H (2020) "An Agent-Based Simulation to Minimize Losses during a Terrorist Attack." Applied Sciences (MDPI) 10(9):3213. [Open access]
3. "Optimizing Medical Care during a Nerve Agent Mass Casualty Incident Using Computer Simulation." PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC11377464/
4. Kaplan DE, Marshall A (1999) "Aum Shinrikyo: Once and Future Threat?" Emerging Infectious Diseases 5(4):831-835.
5. CHEMM (HHS) "Case Studies of Medical Response to Mass Chemical Exposure — Tokyo sarin gas attack." https://chemm.hhs.gov/training_casestudies.htm
6. Sugiyama A, et al. (2020) "The Tokyo subway sarin attack has long-term effects on survivors." PLOS ONE 15(6): e0234967.
7. Tokyo Metropolitan Police Department — incident records, March 20, 1995.

Paul Samuel Guarino is the inventor of the Guarino 700/N Harmonic Detection Framework (U.S. Provisional Patent 63/937,225). He leads Guarino Laboratory in East Northport, New York.