Communication, records and authority

Understand the limits.
Extend what is possible.

How can people retain meaningful control over intelligent infrastructure and rely on it when conditions are imperfect?

01 / The work

Two lines.
One shared requirement.

Communication and records are not enough if people lose the ability to decide about them. That is the relationship the programme seeks to understand.

Line of work

Offline
communication.

Studying how to transmit messages without the internet, and how people coordinate when technical tools fail.

Radio, transmission, local storage and coordination practices. The capabilities of each implementation need to be tested.

Line of work

Records
and integrity.

Studying what makes a record inspectable and trustworthy in a disconnected setting, how it is preserved and who has the authority to keep it.

Integrity, detectable changes, synchronisation and local custody. Retaining data does not guarantee knowledge of everything that happened.

These lines are described in the public programme. Field observation and community listening bring questions that a technical solution alone can overlook. Sources and context ↗

A record is not
the whole event.

“Sent”, “received” and “attended to” are different claims. They should not become interchangeable simply because an interface can display them.

In Pacchanta, the team received a message and heard a bell. They also encountered difficulties in knowing which case had been attended to and who was taking responsibility. A later inspection identified another limit: the node’s clock did not provide trustworthy timestamps. [1]

The question that remains

What evidence supports each claim, what conditions does that evidence require, and what must remain unknown?

This question helps frame the programme.

Return to the field evidence

03 / An approach to research

Ask precisely.
Answer with evidence.

  1. Define what is to be established.

    State the question, the conditions and what would count as an answer. A programme aspiration is not yet a tested hypothesis.

  2. Separate what is known from what is proposed.

    Read the existing work, make assumptions explicit and do not mistake a change in vocabulary for a new contribution.

  3. Choose a test that could contradict us.

    Depending on the question, the work may require formal analysis, computational experiments, technical tests or field research with the appropriate authorisations.

  4. Leave a result that others can examine.

    Separate data, interpretation and limitations. Document failures too, and recognise the actual contributions of the people who did the work.

Where the work stands

The programme has deposited documents, a documented field demonstration and a mathematical draft in preparation. Each work needs its own review; these do not amount to validating a service. The next section explains what the model studies and what remains unresolved.

What the evidence
lets us conclude.

A record can leave parts of an event unresolved yet still answer a specific question: did the service work? Within a model, the answer is determined when every situation compatible with the evidence has the same outcome.

The mathematical draft studies this sufficiency when a cause can damage both components and their records. It presents proofs about how to exhibit ambiguity: two situations with the same evidence and different outcomes.

A small explanation. A larger question.

Under the model’s assumptions, ambiguity can be exhibited by adding a single cause to a prior situation. With further restrictions, the draft bounds how many prior causes are needed. This does not tell us how many measurements are needed or guarantee that the explanation can be found quickly.

The next investigation asks how much computation is needed to check whether a fixed set of reliable measurements is sufficient across all possible records of a compactly described system.

The assumptions and the technical question

Damage accumulates by union, and every combination of causes is allowed. Each observer reports truthfully or is erased. Reliable measurements refer to the same configuration and are fixed in advance; they are not selected adaptively.

The local witness preserves both the record and those measurements when a cause is added that changes the outcome. If the service is monotone — adding damage cannot restore it — and each cause affects at most two entities, including observers, the draft obtains a witness with at most w + 1 prior causes, plus the decisive cause. The parameter w is the largest size of a minimal failure certificate: a set of components whose failure makes the service fail, while no proper subset suffices.

For services described by AND/OR trees with no repeated variables, the available procedure is XP: its exponent grows with w. The question remaining in this work is whether there is an FPT algorithm, whose exponent is independent of w, or a parameterised complexity barrier. We do not claim that the literature recognises this question as open.

The bound compresses a witness of insufficiency. It does not reconstruct what happened in Pacchanta, establish the physical reliability of measurements or transfer authority to act. Fieldwork motivates the model; applying it would require separate justification of these assumptions.

See the distinction in an example

The same evidence.
Two possible outcomes.

Imagine a service with two stages: the first works with A or B; the second needs D. We know that B failed and D works. We do not know what happened to A.

The rule in this exampleThe service works if (A or B) and D work.

Situations compatible with what we know
ABDService
WorksFailedWorksWorks
FailedFailedWorksFails

Two compatible situations

Unresolved outcome

The same evidence allows the service to work or to fail. We cannot yet tell which.

Here, one additional measurement resolves the uncertainty. The research asks more: whether measurements chosen in advance suffice in every case allowed by the model. Knowing the outcome does not determine who has authority to act.

The limits are
part of the research.

Read the publications for their purpose and scope, not as a list of equivalent results.

Read the work