Research on one question: when an automated system is wrong, does the evidence reach someone who can change the rule? Essays, scored cases, a working paper.
Results so far
The results so far, a proposal, and the argument they test. Each says what it rests on.
Casebook · 5 cases
Each case was established by a court, an inquiry, or a regulator, and scored on six safeguards. The casebook calls this one standing: the right of the people a system decides about to challenge a decision and be answered. Correction, the means to halt or reverse the system, held in one.
Read the casebook →Frontier doctrine scan · Sep 2026
The public doctrine of ten frontier AI labs, scored 0 to 4 against the twelve laws. The law that a system able to do something is not thereby allowed to do it averages 2.9, and human oversight 2.3. Dependence on a system averages 0.7, standing 0.8, and keeping decisions open to challenge 0.8. No score has a source attached yet.
Read the scan →Working paper · v0.2
A benchmark proposal: do AI agents meet their targets by using up human effort their metrics do not count? It defines the measure and states falsifiable hypotheses. No experiments have been run.
Read the working paper →Theory · 23 essays
The argument behind the laws. Start with insulation: how an institution keeps the harm a decision causes from reaching anyone who can change the rule.
Read the theory →v1.0 was released for public review. It stays a draft until it has two independent implementation reports and a decision recorded through the public request-for-comments process.
View STD-01 →A review of typed decision models, which return a fixed answer type with a probability per option, led to three STD-08 clauses. Text a system reads cannot grant it permission. A confidence cut-off is a policy with its own record. The threshold that decides which cases reach a person is part of the written terms of that person's oversight.
View STD-08 →Search matches titles, summaries, and tags. The section filter narrows to one part of the page; the tag filter finds related publications and artifacts.
Showing 15 entries
No entries match that search or filter. Try another keyword or browse the sections below.
Last updated Oct 2026. Each change is dated in the changelog at the foot of the page.
How questions, methods, and results are published here.
Each focus area states the questions it is trying to answer. A finding is published only with the method and data behind it.
Agenda items, focus areas, and publications link to the glossary terms they use, so a reader can check each definition.
Research that changes a requirement lands as a dated clause in a standard, recorded in that standard's publication history.
Bridge artifacts
Dated research notes and the theory essays behind the laws.
Dated research note
What the framework supplies to scholars who already hold the values, to governance mechanisms it does not replace, and to the counterpositions that would make it unnecessary overhead, with six worked examples.
Dated research note
Ten frontier labs scored 0 to 4 against the twelve laws on public doctrine and product architecture, not safety performance.
Essay series
The essays that argue for the laws, most of them on why the evidence that a system is wrong fails to reach the rule that produced it. The essays on insulation and exception learning are good places to start.
Essay archive
Essays on what happens when an institution's system is wrong and the people it decides about cannot get it corrected. Several draw on casebook cases: Robodebt, Post Office Horizon, and the Dutch childcare benefits scandal.
Research program
A related research page on the human effort that keeps an institution's reported performance up. It splits observed performance into what the design delivers and what people make up, reports field observations, and sets out how the hypothesis would be tested.
Two questions the results so far leave open, the work each needs, and the glossary terms it uses. Then the three areas they sit in: two of the six safeguards, standing and correction, and the cost they keep with the institution, the work of correcting the system.
Open question
Glossary: Exception absorption , Exception learning
Open question
Glossary: Burden transfer event , Compensated performance
Who can make a running system answer: a challenge route, a date by which it must be answered, and someone who can force a halt.
Glossary: Corrective standing , Standing mechanism
Whether an upheld challenge reaches the rule that produced the decision, and how long that takes.
Glossary: Exception learning , Correction obligation
Who does the work of correcting the system, and whether the institution counts that work as a cost.
Glossary: Burden transfer event , Principle of non-expropriation of resilience , Manufactured virtue , Epistemic erasure cycle , Compensatory inputs
Publications
Each entry states its question, method, and status. A planned study has not been run and reports no results.
report · planned
Planned pilot to test which operational signals the workload modeler should weight most.
Structured abstract
Data & artifacts
Ethics & consent
Glossary: Burden index , Signal credibility
protocol · published
Working paper v0.2, The Green Dashboard: a proposed 52-week simulation of a hospital department that tests whether AI agents meet their targets by using up staff effort their metrics do not count. It names that failure, sets out an experimental design with falsifiable hypotheses, and lists seven gaps in current evaluation practice. No experiments have been run.
Structured abstract
Data & artifacts
Ethics & consent
Glossary: Reciprocal accommodation , Extractive cannibalism , Green dashboard trap , Dual-ledger evaluation , Systemic refusal