|
|
| import json |
| from lamm.verify import ProofAuditor |
| from lamm.metrics import summarize |
|
|
| tests = [ |
| ("algebra_correct", ["STEP 1: 2*x + 3 = 7", "STEP 2: 2*x = 4", "STEP 3: x = 2"], "2"), |
| ("algebra_wrong", ["STEP 1: 2*x + 3 = 7", "STEP 2: 2*x = 5", "STEP 3: x = 2.5"], "2"), |
| ("derivative_correct", ["STEP 1: f(x)=x^2", "STEP 2: f'(x)=2*x"], "2*x"), |
| ("derivative_wrong", ["STEP 1: f(x)=x^2", "STEP 2: f'(x)=x^3"], "2*x"), |
| ("integral", ["STEP 1: integral x^2 dx = x^3/3 + C"], "x^3/3 + C"), |
| ] |
| auditor = ProofAuditor() |
| rows=[] |
| for name, steps, target in tests: |
| a = auditor.audit(steps,target) |
| rows.append((name,a)) |
| print(name, a.label, a.score, "ACS=",a.acs_consequence,"terminal=",a.terminal_ok) |
| print(json.dumps(summarize([a for _,a in rows]), indent=2)) |
|
|
