An evolving research notebook on human judgment, authorship and human–AI collaboration.
Status note — English is the reference version of this document; the other languages are complete translations for reading.
The question this station keeps
What we decide, what we hand over, and what we gradually stop practising all shift long before anyone announces a change. This station studies where the person sits as the surrounding system grows, and what has to be designed and maintained for that position to hold.
A center holds through continuing work — designed, maintained, and checked.
Agency has four working parts: understanding what happened, a meaningful choice among options, the ability to intervene, and accountability afterwards. The research question is practical — which decisions stay in human hands, and how the seam between person and system is kept visible in the interface and in the record.
reading: meaningful human control · contestability · automation bias
Values arrive plural, local, and often in conflict, and a system that averages them has already made a choice. We study how values get encoded, who can contest that encoding, and what it takes to treat refusal as a first-class outcome with handling of its own.
reading: value pluralism · consent · the right to say no
Complementarity is harder to design than raw capability. A system that retains far more than a person can hold, and a person who understands why it matters, pass work back and forth, each keeping to what it does well. Calibrated trust — judging when to rely on a result and when to check it — is the skill being designed for.
reading: handoff · calibrated trust · explanation as dialogue
The pause before the answer. In this notebook, that deliberation stays with the person.
fig. 02aTwo roads drawn in pencil. Which one gets taken, and who answers for it afterwards.
fig. 02bAuthorship as an act: something made on purpose, by someone, for a reason.
fig. 02cUnderstanding held by a person as well as by the record. A slow skill that fades when it goes unused.
fig. 02dCapability drawn as arcs around the person: reaching further while the pen stays in their hand. The drawing states a design aim.
fig. 02eFive postures, one figure. Everything else on these pages is drawn around them.
Each one begins as a human question and ends as an engineering problem. That descent — from meaning to mechanism — is the whole work.
When anything can be generated, origin becomes the scarce thing. Identity work asks how a person, a work or a claim can carry attribution across systems that were not built to preserve it, while everyday expression stays free of surveillance.
tied to: digital identity · verifiable credentials · provenance chains · consent
A sentence read without its situation is easy to misread. Context is the conditions a record was made in — when, where, under what pressure, by whom, for whom. A century of archival practice shows that meaning erodes as context is stripped away, and models inherit that loss at scale.
tied to: archives & preservation · records continuum · situated data · decay
People, works, places and moments matter because of what they point at. Recording the pointing itself as data keeps a human world navigable by both people and machines; links left as decoration disappear with the platform that displayed them.
tied to: knowledge graphs · linked data · URIs · graph traversal
Formal representation buys inspection: structure you can read, argue with, and correct. Statistical inference buys reach. The open question is the boundary — how much of human knowledge survives being written down as structure, and how to handle the rest honestly.
tied to: knowledge representation · ontologies · Semantic Web · limits of formalisation
Decentralised identifiers, verifiable credentials, signed provenance. Decades of work on proving who said what, while leaving everyone's name out of a single central authority.
RDF, ontologies, shared vocabularies. The long, stubborn attempt to make meaning machine-readable while keeping it human-authored and open to disagreement.
Entities joined by typed edges across independent sources. A model of the world where relationships are explicit and inspectable. Outliving any one system is stated here as a design aim, and remains an aim rather than a demonstrated result.
Appraisal, description, custody, migration. The discipline of keeping things legible over time, including the conditions that made them mean something in the first place.
The failure mode to watch is a person quietly giving up the work in the left column because the right column makes it easy to skip. Designing for this division means keeping that work practised.
Where EarthLinC enters
A protocol is a modest instrument. Deciding what matters stays with people. What it governs is how a decision is recorded, so that when someone decides, the decision stays legible to whoever comes later, including the people who were not in the room. The design status of that aim is set out in the scope note below.
Station open — research continuesStill open