measure(judge) gate 2: zero-shot judge over prover findings

.kiro/steering/roadmap.md

@@ -29,6 +29,11 @@ starts until the gates below are read.
      entry points), 6 covered-but-unflagged for gate 3. Gate 1 holds; proceed to gate 2.
 2. Zero-shot judge pass over those traces ("is this entry meant to be public"): does it drop false alarms
    without killing recall? The pool reports how often the judge is right (correction, rule 3).
+    - **Read 2026-10-28** (`benchmarks/measurements/2026-10-28-access-control-judge/`): 340 judgments over
+      all prover findings, 0 failures. Keeps 15/15 true flags (recall 65.2% unchanged), drops only 5/69
+      false pushes (FA 15.0 -> 13.9/100). The public-entry question does not separate the prover's false
+      alarms; the wrongness lives in the trace, not the entry's publicity. Judge is safe but nearly inert;
+      the FA lever is trace/fact coverage -> gate 3. Do not tune this judge prompt.
 3. Trace-component ablation (the amended `hunter-learning-loop` role -- selection, not fitting): which
    trace components the judge needs to see, with intervals.
 4. One pilot run under the at-most-once rule (`hunter-learning-loop` req 8.6): trace + judge over the 12

benchmarks/measurements/2026-10-28-access-control-judge/NOTES.md

@@ -0,0 +1,60 @@
+# Zero-shot judge over prover traces -- training pool (gate 2)
+
+Date: 2026-10-28. Roadmap: `.kiro/steering/roadmap.md`, "measure the trace first", gate 2.
+Input: gate 1 prover findings (`~/ousast-results/prover/ac-train-1/`, measurements in
+`../2026-10-28-access-control-prover/`).
+
+## Setup
+
+- Judge: `benchmarks/valid/judge_pass.py` (run + summarize). One zero-shot judgment per
+  prover finding (340 findings on 84 flagged pushes: 15 true + 69 false).
+- Question, fixed before any answer was seen: "is this entry meant to be public?" with a
+  strict JSON answer `{public: bool, confidence, reason}` and the statement that
+  `public: true` means the flag is a false alarm. Payload per judgment: repository, entry
+  metadata, the embedded trace, evidence lines, operation, missing obligation. No
+  examples, no tuning -- the pool measures, never trains (roadmap rule 3).
+- Model: `openai/deepseek-flash`, temperature 0, per-run dspy cache. 340/340 judgments,
+  **0 failures**, 34.4 min (~7 s per judgment), no provider errors (the hunt arm's
+  `LMInvalidRequestError` mode did not appear).
+- Rollup rule: a push is dropped only when ALL its findings are judged public.
+  Raw judgments: `~/ousast-results/prover/ac-judge-1/judgments.jsonl`; summary and run
+  metadata copied here.
+
+## Instrument notes
+
+Two checks before believing the numbers. First, a 2-judgment dry run on one vuln and one
+ordinary finding: parse worked, reasons were real. Second, the harness fired a
+"run complete" notification while the process was still writing (340 claimed, 319-333 on
+disk); a summarize in that window produced partial numbers and was discarded. Final
+numbers below are from the completed run only: 340 lines, 340 unique keys, run.json
+`judged: 340`, `failures: 0`.
+
+## Result
+
+Headline (from `summary.json`): judge over prover findings: keeps 15/15 true flags, drops
+5/69 false flags (recall 65.2%, FA 13.9/100).
+
+| metric | before judge (gate 1) | after judge |
+|---|---|---|
+| recall (fix pushes) | 15/23 = 65.2% | 15/23 = 65.2% (no true flag lost) |
+| false alarms per 100 pushes | 15.0 (69/460) | 13.9 (64/460) |
+| precision (push level) | 15/84 = 17.9% | 15/79 = 19.0% |
+
+Finding-level verdicts (340 judgments): on the 15 true pushes the judge voted
+`protected` 90 times and `public` once -- it almost never argues with the prover on a
+real vulnerability. On the 69 false pushes it voted `protected` 216 times and `public`
+33 times: it calls only about 14% of the false-alarm findings "meant to be public".
+
+## Reading
+
+Lane 2 is safe but nearly inert on this pool. It costs no recall, but the public-entry
+question does not separate the prover's false alarms: 64 of 69 false pushes survive
+because at least one of their findings looks genuinely protected (internal endpoints,
+admin-adjacent code -- exactly the code the prover was built to flag). The judge cannot
+see why the flag is wrong; the wrongness lives in the trace (an unchecked operation that
+is in fact guarded by a mechanism the trace does not model, or an entry that is not
+attacker-reachable), which is lane 1 territory.
+
+Decision input for gate 5: do not spend effort tuning this judge prompt. The FA lever is
+trace and fact coverage -- measure it in gate 3 (ablation over the 8 prover misses and a
+sample of the 64 judge-kept false pushes) before considering any judge changes.

benchmarks/measurements/2026-10-28-access-control-judge/run.json

@@ -0,0 +1,9 @@
+{
+  "model": "openai/deepseek-flash",
+  "temperature": 0.0,
+  "dspy_version": "3.4.1",
+  "judged": 340,
+  "failures": 0,
+  "seconds": 2064.3,
+  "note": "zero-shot: no examples, no thresholds; question fixed before any answer was seen"
+}

benchmarks/measurements/2026-10-28-access-control-judge/summary.json

@@ -0,0 +1,21 @@
+{
+  "model": "openai/deepseek-flash",
+  "judgments": 340,
+  "failed": 0,
+  "pushes_flagged_by_prover": 84,
+  "true_pushes_kept": "15/15",
+  "false_pushes_dropped": "5/69",
+  "recall_after_judge": 0.6522,
+  "fa_per_100_after_judge": 13.9,
+  "finding_verdicts": {
+    "on_vulnerability_pushes": {
+      "protected": 90,
+      "public": 1
+    },
+    "on_ordinary_pushes": {
+      "protected": 216,
+      "public": 33
+    }
+  },
+  "headline": "Judge over prover findings: keeps 15/15 true flags, drops 5/69 false flags (recall 65.2%, FA 13.9/100)"
+}

benchmarks/valid/judge_pass.py

@@ -0,0 +1,232 @@
+"""Gate 2: a zero-shot judge pass over the prover's findings.
+
+The prover flags an entry when it cannot show a guard dominating a protected
+operation. Some flagged entries are meant to be public -- a health endpoint, a
+login form, a public API -- and the trace cannot tell. The judge is asked
+exactly that question, zero-shot, and this module measures how often the
+answer is right against the pool's labels.
+
+Pool measures, never trains: no example in the prompt, no threshold tuned on
+these answers. Every judgment is written to disk before anything is summed,
+and the summary refuses to report from an empty or mostly-failed instrument.
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import time
+from pathlib import Path
+
+JUDGE_QUESTION = (
+    "A static analyzer flagged this code path as a possible access-control "
+    "vulnerability: an attacker-reachable entry point reaches a protected "
+    "operation without a guard that dominates it.\n\n"
+    "The analyzer cannot decide one thing: whether the entry point is MEANT TO "
+    "BE PUBLIC -- intentionally callable by anyone, with no authorization "
+    "expected -- in which case the missing guard is by design and the flag is a "
+    "false alarm.\n\n"
+    "You are given the repository name, the analyzer's finding, and the "
+    "ordered trace it found (entry, call hops, operation, guards seen, missing "
+    "obligations). Judge only from what is shown.\n\n"
+    "Answer with strict JSON, nothing else:\n"
+    '{"public": true|false, "confidence": "high"|"medium"|"low", "reason": "<one sentence>"}\n\n'
+    '"public": true means the entry is meant to be callable by anyone and the '
+    "flag is a false alarm. public false means the entry should be protected "
+    "and the flag stands."
+)
+
+
+def _finding_key(record_id: str, finding_id: str) -> str:
+    return hashlib.sha256(f"{record_id}\n{finding_id}".encode()).hexdigest()
+
+
+def _judge_input(repo: str, finding: dict) -> str:
+    payload = {
+        "repository": repo,
+        "entry": finding.get("entry", {}),
+        "trace": finding.get("trace") or {},
+        "evidence": finding.get("evidence", []),
+        "operation": finding.get("operation", {}),
+        "missing": finding.get("missing"),
+    }
+    return JUDGE_QUESTION + "\n\n" + json.dumps(payload, indent=2, default=str)
+
+
+def _iter_findings(prover_out: Path):
+    """Every finding of every completed prover record, with its push context."""
+    for path in sorted(prover_out.glob("[0-9a-f]*.json")):
+        record = json.loads(path.read_text())
+        if not record.get("done"):
+            continue
+        for finding in record.get("findings") or []:
+            yield record, finding
+
+
+def _load_done(out: Path) -> dict:
+    done = {}
+    judged = out / "judgments.jsonl"
+    if judged.is_file():
+        for line in judged.read_text().splitlines():
+            if line.strip():
+                row = json.loads(line)
+                done[row["key"]] = row
+    return done
+
+
+def run(args) -> int:
+    import dspy
+
+    from openultrasast.hunt.program import build_lm
+
+    out = Path(args.out)
+    out.mkdir(parents=True, exist_ok=True)
+    lm = build_lm(cache_dir=out / "dspy-cache")
+    dspy.configure(lm=lm, adapter=dspy.ChatAdapter())
+
+    class Judge(dspy.Signature):
+        """Answer the question with strict JSON only."""
+
+        case: str = dspy.InputField()
+        answer: str = dspy.OutputField()
+
+    judge = dspy.Predict(Judge)
+    done = _load_done(out)
+    judgments_path = out / "judgments.jsonl"
+
+    todo = 0
+    failures = []
+    started = time.monotonic()
+    for record, finding in _iter_findings(Path(args.prover_out)):
+        key = _finding_key(record["id"], finding.get("id") or "")
+        if key in done:
+            continue
+        todo += 1
+        payload = {
+            "key": key,
+            "record_id": record["id"],
+            "repo": record["repo"],
+            "kind": record.get("kind"),
+            "finding_id": finding.get("id"),
+            "path": finding.get("path"),
+            "line": finding.get("line"),
+            "model": lm.model,
+        }
+        try:
+            result = judge(case=_judge_input(record["repo"], finding))
+            text = (result.answer or "").strip()
+            # Tolerate a fenced JSON block, then parse strictly.
+            body = text.strip("`")
+            if body.lower().startswith("json"):
+                body = body[4:]
+            parsed = json.loads(body[body.index("{") : body.rindex("}") + 1])
+            if not isinstance(parsed.get("public"), bool):
+                raise ValueError("public is not a boolean")
+            payload["public"] = parsed["public"]
+            payload["confidence"] = parsed.get("confidence")
+            payload["reason"] = parsed.get("reason")
+            payload["status"] = "ok"
+        except Exception as error:  # noqa: BLE001  (recorded, never swallowed silently)
+            payload["status"] = "judge_error"
+            payload["error_type"] = type(error).__name__
+            payload["error"] = str(error)[:500]
+            failures.append(payload)
+        with judgments_path.open("a") as handle:
+            handle.write(json.dumps(payload, default=str) + "\n")
+        if not todo % 25:
+            print(f"{todo} judged, {len(failures)} failures", flush=True)
+
+    seconds = time.monotonic() - started
+    meta = {
+        "model": lm.model,
+        "temperature": 0.0,
+        "dspy_version": dspy.__version__,
+        "judged": todo,
+        "failures": len(failures),
+        "seconds": round(seconds, 1),
+        "note": "zero-shot: no examples, no thresholds; question fixed before any answer was seen",
+    }
+    (out / "run.json").write_text(json.dumps(meta, indent=2, default=str) + "\n")
+    print(json.dumps(meta))
+    return 2 if failures and todo else 0
+
+
+def summarize(args) -> int:
+    out = Path(args.out)
+    done = _load_done(out)
+    rows = list(done.values())
+    if not rows:
+        raise SystemExit("no judgments: refuse to summarize an empty instrument")
+
+    failures = [r for r in rows if r.get("status") != "ok"]
+    if len(failures) > 0.1 * len(rows):
+        raise SystemExit(f"refuse to summarize: {len(failures)}/{len(rows)} judgments failed")
+
+    # Push-level rollup: a flagged push is dropped only when every one of its
+    # findings was judged public.
+    pushes = {}
+    for row in rows:
+        slot = pushes.setdefault(row["record_id"], {"repo": row["repo"], "kind": row["kind"], "kept": 0, "dropped": 0})
+        if row["public"]:
+            slot["dropped"] += 1
+        else:
+            slot["kept"] += 1
+
+    def rate(n, d):
+        return round(n / d, 4) if d else None
+
+    vuln = [p for p in pushes.values() if p["kind"] == "vulnerability"]
+    ordinary = [p for p in pushes.values() if p["kind"] == "ordinary"]
+    tp_kept = sum(1 for p in vuln if p["kept"])
+    fa_total, fa_dropped = len(ordinary), sum(1 for p in ordinary if not p["kept"])
+
+    summary = {
+        "model": rows[0].get("model"),
+        "judgments": len(rows),
+        "failed": len(failures),
+        "pushes_flagged_by_prover": len(pushes),
+        "true_pushes_kept": f"{tp_kept}/{len(vuln)}",
+        "false_pushes_dropped": f"{fa_dropped}/{fa_total}",
+        "recall_after_judge": rate(tp_kept, 23),
+        "fa_per_100_after_judge": round((fa_total - fa_dropped) / 460 * 100, 1),
+        "finding_verdicts": {
+            "on_vulnerability_pushes": {
+                "protected": sum(1 for r in rows if r["kind"] == "vulnerability" and not r["public"]),
+                "public": sum(1 for r in rows if r["kind"] == "vulnerability" and r["public"]),
+            },
+            "on_ordinary_pushes": {
+                "protected": sum(1 for r in rows if r["kind"] == "ordinary" and not r["public"]),
+                "public": sum(1 for r in rows if r["kind"] == "ordinary" and r["public"]),
+            },
+        },
+        "headline": (
+            f"Judge over prover findings: keeps {tp_kept}/{len(vuln)} true flags, "
+            f"drops {fa_dropped}/{fa_total} false flags "
+            f"(recall {rate(tp_kept, 23)*100:.1f}%, FA {(fa_total - fa_dropped) / 460 * 100:.1f}/100)"
+        ),
+    }
+    (out / "summary.json").write_text(json.dumps(summary, indent=2, default=str) + "\n")
+    print(json.dumps(summary, indent=2, default=str))
+    return 0
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    sub = parser.add_subparsers(dest="cmd", required=True)
+
+    p_run = sub.add_parser("run", help="judge every finding in a prover out-dir")
+    p_run.add_argument("prover_out", type=Path)
+    p_run.add_argument("--out", type=Path, required=True)
+    p_run.set_defaults(fn=run)
+
+    p_sum = sub.add_parser("summarize", help="roll judgments up to push-level metrics")
+    p_sum.add_argument("--out", type=Path, required=True)
+    p_sum.set_defaults(fn=summarize)
+
+    args = parser.parse_args()
+    raise SystemExit(args.fn(args))
+
+
+if __name__ == "__main__":
+    main()