phase1/scripts/33_test_grounding.py

"""CP-A regression test: the render-from-ledger gate drops hallucinated / un-loaded / fabricated claims.
Imports serve (loads the index/models) but only exercises verify_claims (pure)."""
import serve

cases = [
    {"case_name": "A v B", "neutral_citation": "2020 INSC 1",
     "chunk": "The dying declaration can be the sole basis of conviction if it inspires full confidence of the court."},
    {"case_name": "C v D", "neutral_citation": "2020 INSC 2",
     "chunk": "Corroboration of a dying declaration is only a rule of prudence."},
]
arr = [
    {"claim": "A dying declaration can be the sole basis of conviction.", "n": 1,
     "quote": "sole basis of conviction if it inspires full confidence"},          # VALID — verbatim substring of case 1
    {"claim": "Corroboration is mandatory in every case.", "n": 2,
     "quote": "corroboration is mandatory and always required"},                    # HALLUCINATED quote — not in case 2
    {"claim": "Some invented holding from a case that was never loaded.", "n": 5,
     "quote": "anything"},                                                          # OUT-OF-RANGE [n]
    {"claim": "A fabricated proposition.", "n": 1,
     "quote": "this exact sentence does not appear in the loaded text"},            # FABRICATED quote
]
arr += [
    {"claim": "Bool-n bypass attempt.", "n": True,
     "quote": "sole basis of conviction if it inspires full confidence"},          # BOOL n — isinstance(True,int) is True; must drop
    {"claim": "Trivial-substring bypass.", "n": 1, "quote": "the court"},            # SHORT quote (<4 words) — substring of almost anything; must drop
]
verified, dropped = serve.verify_claims(arr, cases)
print("verified claims:", [v["claim"] for v in verified])
print("dropped:", dropped)
assert len(verified) == 1 and verified[0]["n"] == 1, "FAIL: gate did not keep exactly the one grounded claim"
assert dropped == 5, f"FAIL: expected 5 dropped (hallucinated quote, out-of-range n, fabricated quote, bool-n, short-quote), got {dropped}"
print("PASS: hallucinated quote, out-of-range n, fabricated quote, bool-n, and trivial short-quote ALL dropped; only the verbatim-grounded claim survives.")