Parse the Critic's structured verdict

Background

The Critic emits machine-readable text so downstream code can branch on it:

GAP: <a sub-question with no useful data>
UNSUPPORTED: <a claim not backed by any finding>
VERDICT: PASS
VERDICT: FAIL

CriticVerdict parses this into three fields with plain string ops — no regex, no JSON.

Problem statement

Implement parse_verdict(raw) returning {"passed", "gaps", "unsupported"}.

Input

• raw — the Critic's raw output string.

Output

Returns a dict:

• "passed" → True iff "VERDICT: PASS" appears in raw,
• "gaps" → list of texts from lines starting with "GAP:" (prefix stripped),
• "unsupported" → list of texts from lines starting with "UNSUPPORTED:" (prefix stripped).

Examples

Example 1 — clean pass

Input:  "VERDICT: PASS"
Output: {"passed": True, "gaps": [], "unsupported": []}

Example 2 — a gap and an unsupported claim

Input:  "GAP: impact had no data\nUNSUPPORTED: StarCoder adoption\nVERDICT: FAIL"
Output: {"passed": False, "gaps": ["impact had no data"],
         "unsupported": ["StarCoder adoption"]}

Constraints

• passed is "VERDICT: PASS" in raw.
• gaps: each line starting "GAP:", text after the prefix, stripped.
• unsupported: each line starting "UNSUPPORTED:", text after the prefix, stripped.

Notes

• Note "GAP:" is 4 characters and "UNSUPPORTED:" is 12 — slice off the right prefix length, then strip.