Education/student-diagnosis
Student answer → diagnosis card
AI grades reasoning level with four evidence rules
01
Problem
On the same reading question, most answers say “agree” or “partly agree.” Asking a model to grade them yields a score and a comment. Neither says where each answer's reasoning stops, so there is no basis for what to teach next.
02
Generation only
Points per key idea. The result is a score; the reasoning leaves no record.
A paragraph of feedback. The standard drifts with the prompt, and nothing ties the verdict to a rule or a quote.
03
With structure
Evidence reasoning is written as four checkable rules, each naming the signals it needs to see. The model only finds signals and quotes them verbatim; the rules decide the level.
| Rule | Level | Name | Signals required |
|---|---|---|---|
| R-01 | L1 | Restates claim | no citation + circular claim + no causal challenge |
| R-02 | L2 | Personal example | personal example + no citation + limits not seen |
| R-03 | L3 | Structural account | cites text + sees limits + logic incomplete |
| R-04 | L4 | Transfer analysis | cites text + causal analysis + names the fallacy |
04
Result
Pick an answer to see the full reasoning and diagnosis card once the AI runs on OCG.
// 1. The AI receives a taskconst input = { domain: "language_reading", ability: "EvidenceReasoning", question: "Someone says, “Xiaohua gets good g…", answer: "I don't think the claim is accurat…",};