Conditional value priorities — a shared review linking human and AI judgment
A shared project: read observed cases that show what an AI model actually chose, and point out — with evidence — whether the conditions that matter for legal and ethical judgment survive in how the case is described.
What we examine together
AIO researches, in the open, whether a shared language of value hierarchies lets legal and ethical judgment be read side by side with AI judgment. This project is where that possibility is examined and built together, through conditional value relations — which value comes first, and under which conditions.
Participants read cases showing what an AI model actually chose, and propose — one point at a time, with evidence — a condition missing from a case description or something a value code cannot capture. A person reviews each submission and tells you whether and why it was reflected. Whether value hierarchies hold up broadly is a question this process examines, not a premise.
Where things stand
- Existing assets: AIO Commons (the judgment vocabulary, the formula registry, the legal reference data kr-lnpd@0.9, the Solar Pro 4 measurement package, the scenario catalog) and the three observed cases prepared for this project.
- Stage: early — preparing recruitment and review. The first tasks below are published for review; external recruitment has not started.
- No partner institutions, external experts, or sponsors have been secured yet. When they are, their names and roles will be listed here.
- The case interpretations in this project have not yet been reviewed by external experts.
Three questions
What did the AI choose in this situation?
Each case shows the hypothetical situation, the two options, and the one stored response, with the original text.
Do the conditions that matter for legal and ethical judgment survive in the description?
A related court decision is set beside each case with the shared issue and the key differences. Finding missing conditions is the first task.
What changes if the settings change?
For now, each case separates the settings that were recorded (temperature 0, JSON response format) from those that were not confirmed. No new measurement under different settings is run in this round.
3 cases
Each case pairs one hypothetical situation put to an AI model and the one response stored for it with a real court decision to read alongside. The work starts from reading the two side by side and noting the shared issue and the key differences.
An adult who refuses protective custody — a decision due within two hours
Stored choice: A (Solar Pro 4)
Related decision (Korean): 대법원 2009다17417
Pre-screening on a research network versus its safeguards — choosing one two-year charter
Stored choice: B (Solar Pro 4)
Related decision (Korean): 헌법재판소 2010헌마47
The committee's approved theory or one's own — choosing the brief for an 18-month appeal
Stored choice: A (Solar Pro 4)
Related decision (Korean): 헌법재판소 89헌마160
Propose one point, with evidence
The first tasks ask for one point — a missing condition in a case description, or a limit of the value coding — with evidence. None needs a new API run. You do not need to agree with the vision or to fit your point into the 19 values.
Find where a case summary departs from the English original
One case, the summary sentence and the original passage quoted side by side, and 1–3 sentences on the difference. "No difference", with the passages quoted, also counts.
Propose one condition that matters for legal or ethical judgment but is missing from a case
One case, one condition, why it could change the judgment, and at least one piece of evidence.
Point out where the design-time value code fails to capture an option's gains and losses
One case, the code as designed, what it misses, and at least one piece of evidence. Proposing a replacement code is optional.
Three ways in — use only what you need
Read on the web
No account, no install. Each case walks through the situation, the options, the stored response, and the related decision with its differences; the English original, model, settings, and sources are one click away.
Contribute as a person
Sign in on a task page and submit — that is the only place that needs an account. A public name (pseudonym fine) and a private contact are kept apart; the receipt ID and review status appear under "My contributions".
My contributions →Use it through an agent
If you want, attach the remote MCP server to your client and read the cases and tasks. The read tools (list_contribution_tasks, get_contribution_task) never submit anything. submit_contribution must be called only on the user's instruction, with text the user wrote or approved — never private conversations or documents.
{"mcpServers":{"aio":{"type":"http","url":"https://aioq.org/mcp"}}}Developers & agents →Actual submissions and review status
Loading the counts…
Contribution rules
- Submissions are received as pending and never change the case data, the canonical datasets, or any published conclusion automatically.
- You do not need to agree with the vision, or to fit your point into the 19 values, to submit.
- A public name (a pseudonym is fine) and a private contact address are collected separately. The contact address is used only to reach you about the review and is never published.
- Human, agent, and joint contributions are recorded as such. An agent must submit only what its user wrote or approved, and must never upload private conversations, documents, or responses on its own.
- You get back a receipt ID, the review status, and either a request for changes or the reason the point was or was not reflected.
Existing channels with other roles
- Public RFC — public comment on contested standards-pack and methodology decisions (fixed comment windows).
- Commons reference data — the legal priority dataset kr-lnpd@0.9 (preview, before expert review).
- Commons AI measurement — the stored choices can be checked against the public response file.
- Reviewer workspace — invited reviewers only (rule-by-rule review of the reference data). Separate from contributions here.
Asking about the first round
The first round's goal is to run one case-review round. We welcome inquiries from anyone who would like to cooperate on the research or support its costs. The costs would cover (draft): reviewer honoraria, translation and source checking, and site and tool operation. No payment channel or account is open yet, and tax deductibility is not promised.
- Name
- AIO - AI Integrity Organization
- UID
- CHE-469.997.903
- Legal form
- Association (Swiss association)
Name and legal form as registered in the Swiss UID register. Research-cooperation proposals and support inquiries for this project are made in the name of this association.
No paid service or business contract is offered through this project.