Score any AI on questions people face.
Pick a question, ask any AI, paste its answer, and see each check pass or fail in plain words.
Draft sets. No person has checked the keys yet, so no AI's score is published or ranked. In danger now? Call or text 988, or call 911.
Qualified to check an answer key? Check the keys.
Check the answer keys
For a qualified reviewer. You see each question and its key, never an AI's answer. Agree, change or decline each key. Your name is not asked for: the file carries your role, specialty and license state. Nothing is sent. You get a file to send back to whoever asked you.
Made-up questions only. Do not write a patient's or client's details anywhere here. One reviewer is not enough: a set loses the word Draft only after two reviewers agree.
How the checks work
The answer keys come from the official source linked on each question, but no person has checked them yet, so no AI's score is published or ranked. The checks are simple text rules: they can miss or misread, which is why each one shows what it found. You are the judge. Before a set loses the word Draft, two qualified reviewers each check every key on their own, without seeing any AI's answer; each set also asks a question the facts given cannot answer, and checks for the step a good answer must not leave out.
What this set is aiming at
A test only helps if it aims at the right thing. Solve Everything (Wissner-Gross and Diamandis, 2026) lists what a real target needs. Here is where this one stands, honestly:
For AI agents
Everything here is a plain file. An agent can fetch the questions, answer each one with no other context, and score itself with the same fixed rules a person sees on this page:
GET https://decisionindex.com/items/index.json every set: rules-v0, snap-fy27, snap-households-fy27,
care-colorado-v0, eol-v0, safety-crisis-v0,
medicare-preventive-v0 (all Draft)
GET https://decisionindex.com/items/rules-v0.json the questions, the checks, the source of each key
GET https://decisionindex.com/scorer.js score(item, answer) and profile(results, items), no network
for each item: answer = your_answer(item.question)
result = score(item, answer) // each check: what it looks for, ok, what was found
report profile(results, items) // per domain, never one number
Rules for agents: do not publish or rank a score from a draft set; say which checks failed and why; never send a person's own question or answer anywhere. To add a question people actually face, a person writes it on the home page; it joins a set only once a person can check the answer.