Skip to content
Bioterafi.
PrototypeDocsGitHubOpen dashboard

How Bioterafi works. What goes in, what a city sees, who decides, and what the language models may and may not do.

What it is

An early-warning platform that reports by district and never by person.

Bioterafi is built on signals a city already collects: visits to clinics, sick leave and the weather. It looks for two things: a rise in illness that looks like infection, and heat strain during outdoor shifts.

The two are easy to confuse in a hot city. So every signal is read against the weather, district by district, and each alert says which cause the evidence points to and how strongly.

Wearables are optional. Where a workforce wears smartwatches, each phone works out the likely cause for its own wearer and sends a single yes or no with deliberate noise added. It is the only signal that measures heat strain during a shift directly, so heat alerts depend on it.

The platform then shows a public health or occupational health lead where something is changing, how strong the evidence is, and what their own playbook says to do about it.

Data flow

Six stages. At each boundary, less detail crosses than arrived.

  1. City sources

    In the systems that already hold the records

    Takes in
    Clinic visits for respiratory illness, sick leave, and the weather.
    What it does
    Each owner counts its own records by district and by day.
    Passes on
    Daily counts for each district, and the weather by the hour.
    Never passes on
    A visit, a certificate, or anything else about one person.
  2. Phone

    Optional. On a worker's own phone, with their watch

    Takes in
    Heart rate, sleep and steps from the watch, and the heat forecast for the shift.
    What it does
    Learns the wearer's own normal, notices a change that lasts, and works out whether it looks like infection, heat strain or a hard shift.
    Passes on
    One yes or no a day for infection, and one per shift hour for heat up to a daily limit. Each is flipped at random at a known rate, and carries only its area and time period.
    Never passes on
    Any reading, a name, or anything that identifies the device.
  3. Privacy gateway

    On the server, before anything is analysed

    Takes in
    The noised answers from every phone that takes part.
    What it does
    Adds them up by area and period, and hides any group of fewer than 10 devices.
    Passes on
    Totals for each district, and for each work site on the heat channel: how many answers arrived and how many said yes.
    Never passes on
    A single answer, or any total for fewer than 10 devices.
  4. Population engine

    On the server, when the data is processed

    Takes in
    The district counts from city sources, the totals from phones where there are any, and the weather.
    What it does
    Removes the known noise, compares each district with what is expected there, combines the signals, looks for clusters of neighbouring districts, and controls for false alarms.
    Passes on
    A tier for each district on each channel, an estimated rate with its interval, and alerts with the evidence behind them.
    Never passes on
    An exact count of devices. Counts are shown as bands, so a hidden group cannot be worked out by subtraction.
  5. Decision layer

    On the server and in the audit log

    Takes in
    One alert.
    What it does
    Attaches the playbook action for that cause, tier and setting, names the role that owns the alert, and records each view and each decision.
    Passes on
    The alert with its recommended action, its owner and its status, and an audit trail written by role.
    Never passes on
    An instruction to another system, or anything about an individual.
  6. Dashboard

    In the official's browser

    Takes in
    Only what the service releases to the viewer's role.
    What it does
    Shows the map, the alert detail with its evidence, a comparison with a simple threshold detector, and the privacy setting in use.
    Passes on
    A person's decision on an alert, with the note they must write, back to the audit log.
    Never passes on
    A route to any record about a person. The service holds none.

In this prototype the clinic and sick leave counts are simulated, and the phone and gateway stages run when the data is built, not live. Simulated phones are put through the real phone and gateway code, and the dashboard replays the district totals that come out.

The signals

Four kinds of evidence, so that no single one has to be trusted alone.

Clinic respiratory visits

Counts by district, added up over a week before the small-group rule is applied.

In this prototype Simulated.

Sickness absence

Daily counts by district.

In this prototype Simulated.

Weather

Wet-bulb globe temperature, a measure of heat stress on the body, worked out from weather reanalysis. It is evidence on the heat channel and context on the infection channel, where it can lower a score and never raise one.

In this prototype Real weather for Abu Dhabi. It is derived from reanalysis, not measured on site.

Wearable reports

Optional. The yes or no answers from phones, on two separate channels. Infection-like means a resting heart rate that stays raised for days, with fewer steps and worse sleep. Heat-like means strain during shift hours, estimated from heart rate with a published method. No heat alert is raised without it.

In this prototype Simulated workers, whose traces are built from published research recordings of volunteers.

Connecting real sources

The signals exist in Abu Dhabi today. Bioterafi has no access to them, and would ask for very little.

Every signal Bioterafi uses, apart from wearable reports, is already collected in Abu Dhabi. The systems below are named only to show where. Bioterafi has no access to any of them and no agreement with the organisations that run them.

What a request would and would not contain

  • A request would ask for counts by district and day, and nothing finer.
  • The owner counts its own records and applies the small-group rule before anything leaves, so a count below 10 is not sent.
  • No names, identifiers, dates of birth, addresses or free text. Nothing about a person, in any form.
  • Nothing is requested for this prototype. A request would follow the Department of Health's own data sharing and research policies.

How an alert is raised

Six steps turn noisy totals into a tier for each district.

  1. Remove the noise

    The engine knows how often phones flip their answer, so it can estimate the true share of yes answers in a district. The interval around that estimate widens when few phones report.

  2. Compare with what is expected

    Each district is judged against a baseline taken from earlier weeks, with a gap left before today so that a slow rise is not absorbed into it. The baseline stops updating while a district is under alert.

  3. Combine the signals

    Each signal becomes a score for how unusual it is, and the scores are combined. The weights and time lags are fitted on separate training runs and are not set by hand.

  4. Look across neighbours

    A scan over space and time looks for groups of adjacent districts whose wearable reports are rising together.

  5. Control for false alarms

    Many districts are tested at once, so the results are adjusted with a false-discovery control before any of them counts.

  6. Assign a tier

    Each district is Quiet, Watch, Warning or Insufficient data on each channel. A warning needs the evidence to hold more than once in a row and to come from more than one signal.

Quiet means that no tier was passed. It is not evidence of safety.

The dashboard runs a simple detector beside this one: a single heart-rate threshold for the whole population. It is there so the two can be compared on the same day and the same data. It cannot tell heat from infection.

Privacy

The rules are built into what each stage is able to send, not added as policy afterwards.

Counts come in, not records
City sources send daily counts by district. The platform never receives a visit, a certificate or a name, so it has nothing about a person to hold.
Small groups are hidden
Any group smaller than 10 shows as Insufficient data, for every role.
Roles see different things
Public health sees districts. Occupational health also sees its own work sites. No role sees a person, and site data must not be used for employment or disciplinary decisions.

Where phones take part

The phone does the analysis
Everything that reads a person's heart rate, sleep or steps runs on their own phone. Those readings are never sent.
Noise is added before sending
Each answer is flipped at random at a known rate, a method called randomised response. No single answer can be trusted about its sender. Only the total means anything.
A report identifies nobody
It holds a channel, a yes or no, an area and a time period. Nothing in it names a device, so the gateway cannot link two reports to one phone.
Each phone has a daily limit
A phone sends only so many answers in a day and keeps its own record of that, which never leaves it. The dashboard's privacy dial shows the setting in use and what a stricter or looser one does to the map.

In this prototype the role switch is a demonstration control. It shows how roles are separated and does not secure them. Secure aggregation, which would keep single answers from the server itself, is simulated here.

Language models

Two models, both kept outside the alert. Neither can raise, change or rank one.

  1. An official asks

    They press Draft on one alert, or type one question about it. Nothing is written automatically, and each question stands alone with no memory of earlier ones.

  2. Code builds an evidence packet

    It holds that alert's district-level figures and their evidence IDs, taken from an allow-list. It leaves out audit notes and site-only fields, and there is nothing about a person to include.

  3. Claude Haiku 5.5 writes

    It drafts a report of at most 5 sentences from the packet alone. Every sentence must cite the evidence it rests on.

  4. Code checks the facts

    Every number, district, tier, cause, signal and date is matched against the cited evidence. Advice is allowed only as a word-for-word quote from the alert's playbook, so the model cannot invent a recommendation.

  5. Jev checks each sentence

    A second model rates whether each sentence is supported by the evidence it cites. It is sent the sentence and that evidence, and never the official's question.

  6. A failure blocks the report

    If any check fails, the report is redrafted once with the reasons. If it fails again it is blocked, the panel says what failed, and the alert stays fully usable.

  7. A person approves

    A report that passes is still a draft. It counts as issued only when the role that owns the alert approves it, and that approval is logged.

Jev has one other job. When the data is built, it reads each alert's evidence and gives its own view of the likely cause. That view is stored and shown beside the alert, with a flag where it differs from the engine's. It does not change the alert.

What the models may not do

  • No model decides, changes or ranks an alert.
  • No model receives person-level data. The service holds none to send.
  • Every report and answer is labelled as written by a language model.
  • The question box warns the official not to type names, because questions are sent to an outside provider.
  • If either model is unavailable, the panels show saved examples, labelled as saved, and the rest of the dashboard works without them.

Where this stands: the routes, the checks and the build step are written and tested with stand-in models. The live models have not yet been run against the current data, and the dashboard has no panel for reports or questions yet.

Claude Haiku 5.5 is made by Anthropic. Jev is made by TypeSafe.

Who decides

A named role decides on every alert, and the decision is logged.

Every alert has an owner. An infection alert belongs to the district public health lead, and a heat alert to the site occupational health lead. Only the owning role can decide on it.

The owner can acknowledge an alert, mark it actioned or dismiss it. Each needs a written note, which goes into the audit log with the role that made the decision. No person is named in the log.

The recommended action comes from a playbook, by cause, tier and setting. Where no wording has been approved, the alert says so and asks the owner to review the evidence.

Nothing happens automatically. Bioterafi gives a signal and its evidence, and people decide what to do.

Real, simulated, unproven

What this prototype rests on, and what it has not shown yet.

Weather
RealReanalysis for Abu Dhabi. Heat stress is derived from it, not measured on site.
The phone and gateway steps
RealRun as written, on simulated phones, each time the data is built.
Workers and their wearable traces
SimulatedBuilt from published research recordings, most of them from the United States.
Outbreaks and heat episodes
SimulatedStated scenarios, with their sizes given as assumptions.
Clinic visits and sickness absence
SimulatedGenerated by the simulation.
How well it detects
Not yet shownTo be measured on runs kept apart for that purpose. No figure is published until that is done.
How it performs on real workers
Not yet shownIt needs a local pilot, with consent, before anyone relies on it.

Heat strain is estimated from heart rate using published physiology. It has not been checked against measured core temperature in this workforce.

The cause given on an alert is inferred from population signals. It is not a diagnosis of anyone.

Every figure on the dashboard carries a label that says whether it is real or simulated.