Who is really there?

Both can be faked, and the fakes are getting cheap.

• Fake people. In 2024 an employee of the engineering firm Arup in Hong Kong transferred about 25 million US dollars after a video call in which every other participant was a deepfake of a colleague.
• Tampered software. Also in 2024, a hidden backdoor was found in XZ Utils, a compression library used across Linux systems (CVE-2024-3094). The person who planted it had spent about two years earning the maintainers’ trust. In 2020, attackers slipped malicious code into SolarWinds software, which was then shipped, signed, to thousands of customers.
• Tampered hardware. A 2012 United States Senate investigation found counterfeit electronic parts throughout the military’s supply chain.

The lesson of all three: a signature shows which program ran, not that what it reports is true. A tampered camera signs a false picture as faithfully as an honest one signs a true picture.

For every device it relies on, an agent keeps a short history: who made it, how it arrived, what software it runs, who vouched for it, and how its readings have compared with others.

No single sign is proof. Together they are a strong hint.

1. A broken chain. The device’s identity does not match its claimed maker or software, or its software cannot be rebuilt from its published source.
2. A known-bad source. The device, its parts or its software came through a seller, batch or repository publicly reported as compromised. Each such report is dated, sourced and expires unless renewed.
3. Readings that break physics. A meter reports use during a recorded power cut. A photo’s shadows do not match its time and place. A location jumps faster than any vehicle.
4. Independent observers disagree. Devices with different owners, makers and software disagree by more than their known error.
5. The “person” is a recording. A presence check shows signs of replay or synthesis, or one person’s credential is used in two places at once.

In simulation first. We would build worlds where some players clone voices, replay presence checks, tamper with meters, ship backdoored software through a reseller or buy counterfeit sensors. Detector agents see only what a real agent would see, and are rewarded for correct alarms and penalised for false ones, more heavily when the false alarm lands on an honest person. Every documented real incident is rebuilt as a test the detectors must pass without having seen it before.

1. The suspect reading counts for less until it is confirmed. The person’s standing does not change.
2. The person involved is told privately, with the reasons. Most alarms will have innocent causes: a broken sensor, a bad connection.
3. Independent confirmation is sought: a second device, a fresh presence check, or members of the community vouching in person.
4. Confirmed tampering is reported with its evidence, and a compromised source is added to the public list.
5. Anyone flagged can contest it, and is restored fully if cleared.

A detector that simply distrusts cheap hardware would quietly rank people by their wealth. So the design separates two things: how much a reading counts may depend on where its device came from; how much a person counts never does. Anyone whose only device is low-assurance gets a free route to strong evidence: a loaned, checkable device, or in-person confirmation by their community. Detectors are judged on how often they wrongly flag honest people with cheap devices, not only on what they catch.

• We want thisFakes are caught before money moves

  Agents routinely spot replayed presence checks, cloned voices and tampered devices in time.

  For example: a co-operative’s agent pauses a large payment when a supplier’s “video confirmation” shows the noise pattern of a synthesised face; the supplier’s real staff confirm it was fraud.

• We want to avoid thisCheap devices become suspect

  Detectors learn that cheap hardware means risk, and people who can only afford it are believed least.

  For example: a day labourer’s attendance, recorded on a used phone, is flagged again and again, and his pay is delayed while others are paid at once.

• We want to avoid thisA bad batch spreads before anyone notices

  A compromised batch of sensors or a backdoored library is trusted everywhere for months.

  For example: ten thousand meters from one reseller under-report losses for a year, and the savings claimed on them turn out to be false.

• We want thisA shared, public list of compromised sources

  Reports of tampered batches and repositories are shared as dated, sourced entries that every agent checks.

  For example: within a day of a disclosure, agents in three countries count readings from an affected camera model for less and ask for confirmation another way.

A simulated world with honest and dishonest players and devices of every price, scoring detectors on what they catch and on how fairly they treat cheap devices.

A short history format for devices. And a design for a low-cost, checkable sensor with a published parts list, so honest evidence is cheap for everyone.