Start with the people affected.
Name who can be helped, harmed, left out, or unable to contest the system before choosing a technical metric.
A care-centered method
Safety is a relationship, not a number.
The score is a prompt for a better conversation. It does not certify a model or replace affected people, domain experts, or independent review.
Make power visible.
Autonomy, reversibility, oversight, impact, and voice form a relationship map. No factor is allowed to hide inside an average.
Leave a route back.
A good case ends with a control a person can use: pause, appeal, revise, roll back, or escalate to someone accountable.
A tiny protocol
Start with the person, not the benchmark.
Move the two controls that matter most: how much authority the system has, and how much power affected people can exercise. Then run the same explainable engine used by the public ledger.
Your factor breakdown will appear here.
Read the room
Research is part of the interface.
The ledger links to public safety and responsible-AI research. Treat the feed as context, not as a substitute for local knowledge or lived experience.
Reading public signals…
Evidence with a boundary
Scores are deterministic and inspectable, but they inherit the assumptions of the case maker. The UI keeps those assumptions in view.
Care as infrastructure
Consent, appeal, explanation, and human contact are treated as controls with operational consequences.
Portable by default
Every case is available through REST and MCP, so an agent can help maintain the record without becoming a hidden gatekeeper.