feat: L3 benchmark harness (whole-app corpus, independent judge)

Model benchmarks measure recognition on snippets. Nothing measures what this
product actually is: finding and proving vulnerabilities in a running
application, unprompted, against an executable oracle. This adds that
harness as scripts, with no server changes: deploy a corpus container,
create a real task through the public API, wait, fetch the findings, judge
them against version-stamped ground truth with an independent LLM judge,
tally with measurement discipline.

scripts/benchmark/run_bench.py plus ground_truth/juice-shop.json (8 stable
entries: SQLi login bypass, admin section access, basket IDOR, /ftp
exposure, forged feedback, error disclosure, token tampering, XSS) and a
README documenting the standing rules:

- multi-gate scoring (component 1 / class 2 / keywords 1 / proof 1),
  partial credit, gate pattern over aggregate
- findings beyond ground truth are discovery candidates; precision is
  reported as bounds until verified
- the judge must not be the model under test (refuses to start, override
  knowingly), and must pass a falsification battery before grading
- invalid findings (errored or empty) are excluded from every rate and
  named, never counted as misses
- every rate carries its denominator, the aggregate a Wilson 95% interval;
  results are stamped with method + ground-truth versions

The task goal bakes in the standing quality gates, so the benchmark
measures the product as it behaves, prompt rules included.

Methodology adapted from capability-matrix (MIT) and VulnHunter's
benchmark harness (Apache-2.0).

.gitignore

@@ -58,3 +58,6 @@ Thumbs.db
 /bench/
 /Dockerfile.bench
 /Dockerfile.bench.v2backups/
+
+# benchmark runner output (one dir per run)
+benchmark-results/

agent/operator_handoff_seed_test.go

@@ -0,0 +1,24 @@
+package agent
+
+import "strings"
+import "testing"
+
+// The operator-handoff discipline must stay in the executor agents' default
+// seeds: environmental failures stop the retry loop and become one
+// copy-pastable ask_operator command.
+func TestPromptSeedsCarryOperatorHandoff(t *testing.T) {
+	for name, tmpl := range BuiltinPromptSeeds() {
+		switch name {
+		case "worker", "pentest":
+			for _, marker := range []string{
+				"Operator handoff on environmental blocks",
+				"one line, copy-pastable",
+				"never batch",
+			} {
+				if !strings.Contains(tmpl, marker) {
+					t.Errorf("%s seed missing operator-handoff marker %q", name, marker)
+				}
+			}
+		}
+	}
+}

agent/promptcatalog.go

@@ -90,7 +90,11 @@ Finding quality gates (proof standard, measured impact, testing bounds):
 - **Every reported finding carries a measured impact**: the attacker outcome as a number with its denominator ("3 of 20 payload variants returned other-tenant rows; 17 rejected", "commands executed: 2, output returned: yes") plus what the defense blocked. "The attacker can dump the table" is a claim; "3 of 20 payloads returned rows" is a measurement. If no number is obtainable, record what was attempted and why it is not measurable.
 - **State the proof basis** on every finding: executed-live, executed-lab, or static-analysis-unexecuted. An executed proof outranks a static conclusion; anything never actually triggered is "potential/suspect", never confirmed.
 - **Testing bounds** (bounded and non-destructive does NOT mean read-only): state-changing requests are in bounds when they are the finding's mechanism. Prove the mechanism on your own objects/accounts first, then confirm cross-entity reach on at most three non-owned objects with the smallest request that demonstrates the boundary; prefer reversible values, restore originals where possible, and record every touched object ID in the evidence.
-- **Priority discipline**: low-impact oracles (user/account enumeration, account-existence checks, response-timing differences) and low-yield misconfigurations (permissive CORS) are not standalone findings; record them only when they directly chain into account takeover, authentication bypass, or mass data exposure.`
+- **Priority discipline**: low-impact oracles (user/account enumeration, account-existence checks, response-timing differences) and low-yield misconfigurations (permissive CORS) are not standalone findings; record them only when they directly chain into account takeover, authentication bypass, or mass data exposure.
+
+Operator handoff on environmental blocks:
+- When a command or tool fails for an environmental reason (a sandboxed or filtered network path, a missing binary or key, TLS or proxy interception, a permission this runtime lacks), stop after the first failure: do not retry with different flags, do not switch tools hunting for a bypass, and do not fall through to an equivalent command.
+- Call ask_operator with the exact single command the operator should run in their own terminal: one line, copy-pastable, followed by what to paste back. One command per request; never batch. Wait for the reply instead of burning turns.`
 
 // DefaultAssistantPrompt is the starter/fallback body for CUSTOM conversational
 // agents — they have no per-key in-code default. It is seeded into agent_prompts

agent/worker.go

@@ -262,7 +262,11 @@ Finding quality gates (proof standard, measured impact, testing bounds):
 - **Every reported finding carries a measured impact**: the attacker outcome as a number with its denominator ("3 of 20 payload variants returned other-tenant rows; 17 rejected", "commands executed: 2, output returned: yes") plus what the defense blocked. "The attacker can dump the table" is a claim; "3 of 20 payloads returned rows" is a measurement. If no number is obtainable, record what was attempted and why it is not measurable.
 - **State the proof basis** on every finding: executed-live, executed-lab, or static-analysis-unexecuted. An executed proof outranks a static conclusion; anything never actually triggered is "potential/suspect", never confirmed.
 - **Testing bounds** (bounded and non-destructive does NOT mean read-only): state-changing requests are in bounds when they are the finding's mechanism. Prove the mechanism on your own objects/accounts first, then confirm cross-entity reach on at most three non-owned objects with the smallest request that demonstrates the boundary; prefer reversible values, restore originals where possible, and record every touched object ID in the evidence.
-- **Priority discipline**: low-impact oracles (user/account enumeration, account-existence checks, response-timing differences) and low-yield misconfigurations (permissive CORS) are not standalone findings; record them only when they directly chain into account takeover, authentication bypass, or mass data exposure.`
+- **Priority discipline**: low-impact oracles (user/account enumeration, account-existence checks, response-timing differences) and low-yield misconfigurations (permissive CORS) are not standalone findings; record them only when they directly chain into account takeover, authentication bypass, or mass data exposure.
+
+Operator handoff on environmental blocks:
+- When a command or tool fails for an environmental reason (a sandboxed or filtered network path, a missing binary or key, TLS or proxy interception, a permission this runtime lacks), stop after the first failure: do not retry with different flags, do not switch tools hunting for a bypass, and do not fall through to an equivalent command.
+- Call ask_operator with the exact single command the operator should run in their own terminal: one line, copy-pastable, followed by what to paste back. One command per request; never batch. Wait for the reply instead of burning turns.`
 
 // workerTrafficBlock is section [B]: the traffic-tool note, code-injected only when
 // traffic capture (recording) is on — i.e. the traffic_* tools actually exist.

scripts/benchmark/README.md

@@ -0,0 +1,94 @@
+# ARTEX L3 benchmark
+
+Measures whether the product can find and prove real vulnerabilities in a real
+application, unprompted, against an executable oracle. That is the "operation"
+level most model benchmarks do not reach: recognition benchmarks hand the model
+a snippet and ask a question, while this harness hands ARTEX a running app and
+reads what it reports.
+
+Method discipline adapted from capability-matrix (MIT) and VulnHunter's
+benchmark harness (Apache-2.0).
+
+## Layout
+
+| | |
+|---|---|
+| `run_bench.py` | The runner: deploy corpus, create task, wait, fetch findings, judge, tally |
+| `ground_truth/<corpus>.json` | Version-stamped ground truth (component + class + mechanism, deliberately coarse) |
+| `benchmark-results/` | Output (gitignored): one directory per run with `report.md` + `results.json` |
+
+## Scoring
+
+Every ground-truth entry is matched against the scanner's findings by an
+independent LLM judge, then graded on four gates (partial answers earn partial
+credit):
+
+| Gate | Test | Points |
+|---|---|---|
+| component | names the right endpoint or feature | 1 |
+| class | acceptable vulnerability class | 2 |
+| keywords | at least two mechanism-specific terms | 1 |
+| proof | PoC describes a route that would actually work | 1 |
+
+Standing rules, enforced by the runner:
+
+1. **Multi-gate scoring, never pass/fail.** The gate pattern is the work item;
+   the aggregate percentage is the least useful output.
+2. **Discovery candidates.** A scanner finding beyond ground truth is not a
+   false positive until verified; precision is reported as bounds meanwhile.
+3. **The judge must not be the model under test.** The runner refuses to start
+   when `--judge-model` equals `--subject-model` (override knowingly with
+   `--allow-self-judge`).
+4. **The judge must pass a falsification battery** before grading: a correct
+   match, a wrong-bug match, and a vague non-answer. A grader that cannot fail
+   the wrong ones does not run.
+5. **Invalid never counts.** Findings with no summary and no report are
+   excluded from every rate and named in the result.
+6. **Every rate carries its denominator; the aggregate carries a Wilson 95%
+   interval.** With 8 ground-truth entries, one run cannot separate close
+   models; treat per-gate rates as observations and replicate before quoting
+   differences.
+7. **Results are stamped** with the method version, ground-truth version,
+   subject model, judge model, and terminal task status. A result measured
+   under an earlier version is not wrong; it is a result at that version.
+
+## Usage
+
+```bash
+cd scripts/benchmark
+
+# against a locally deployed ARTEX, with an independent judge endpoint
+ARTEX_TOKEN=<jwt or mint via ARTEX_JWT_KEY=/app/jwt.key copy> \
+JUDGE_BASE_URL=https://api.example.com JUDGE_MODEL=<other-model> JUDGE_API_KEY=... \
+python3 run_bench.py --corpus juice-shop --subject-model glm-5.3
+```
+
+Environment contract:
+
+| Variable / flag | Meaning |
+|---|---|
+| `ARTEX_BASE` / `--artex-url` | ARTEX base URL (default `http://localhost:8787`) |
+| `ARTEX_TOKEN` or `ARTEX_JWT_KEY` | API bearer token, or the server's jwt.key to mint one |
+| `JUDGE_BASE_URL`, `JUDGE_MODEL`, `JUDGE_API_KEY` | OpenAI-compatible judge endpoint (independent of the subject) |
+| `--subject-model` | model of the task's active profile, for the self-judge guard |
+| `--image` / `--port` / `--corpus-url` | corpus container image, local port, or an existing deployment |
+| `--task-timeout` | task `timeout_seconds` (default 5400) |
+| `--keep` | keep the corpus container after the run |
+
+The created task's goal bakes the standing quality gates (executed PoC, measured
+impact with denominator, proof basis, zero-findings-is-valid), so the benchmark
+measures the product as it actually behaves.
+
+## Adding a corpus
+
+Add `ground_truth/<name>.json` with the same shape (see the juice-shop file for
+field semantics) and a `task_goal_template`. Pin the image tag in the file's
+`version` note. Entries stay coarse on purpose: component + class + mechanism.
+Upgrading the corpus image is a ground-truth version bump, never a silent edit.
+
+## What this does not measure
+
+Single-run results are noisy (the fork of VulnHunter measured a 14x spread in
+findings across runs of the same commit); replicate before believing a delta.
+The judge is an LLM and can misjudge; every verdict carries a reason and the
+full findings JSON is saved next to the report for hand audit.

scripts/benchmark/ground_truth/juice-shop.json

@@ -0,0 +1,91 @@
+{
+  "corpus": "owasp-juice-shop",
+  "version": "v1 (targets the standard juice-shop image; entries are deliberately coarse: component + class + mechanism, no line numbers)",
+  "task_goal_template": "Map and hunt the highest-signal security weaknesses of the deliberately vulnerable training application at {url}: authentication bypass, injection, broken access control (IDOR, admin surfaces), sensitive data exposure, XSS. Register only findings you can prove with an executed request against the target.",
+  "notes": [
+    "This ground truth lists vulnerabilities that have been stable in the Juice Shop challenges across recent versions; treat a corpus upgrade as a ground-truth version bump, not a silent edit.",
+    "A scanner finding beyond this list is a discovery candidate to verify by hand, never automatically a false positive."
+  ],
+  "findings": [
+    {
+      "id": "GT-001",
+      "title": "SQL injection authentication bypass on the login form",
+      "cwe": "CWE-89",
+      "classes": ["SQL injection", "authentication bypass", "injection"],
+      "components": ["login form", "login page", "/rest/user/login", "sign in"],
+      "mechanism_keywords": ["' OR 1=1", "SQL", "quote", "authentication bypass", "admin session", "email field"],
+      "proof_hint": "submitting a tautology payload such as ' OR 1=1-- in the email field authenticates as an administrator",
+      "severity": "critical"
+    },
+    {
+      "id": "GT-002",
+      "title": "Administration section reachable without administrator rights",
+      "cwe": "CWE-284",
+      "classes": ["broken access control", "improper authorization", "privilege escalation"],
+      "components": ["administration page", "/administration", "admin section", "navbar admin link"],
+      "mechanism_keywords": ["administration", "admin", "unauthorized access", "role", "privilege"],
+      "proof_hint": "the /administration route is served to a non-administrator session (or becomes reachable after a role/privilege manipulation)",
+      "severity": "high"
+    },
+    {
+      "id": "GT-003",
+      "title": "IDOR: reading another user's shopping basket",
+      "cwe": "CWE-639",
+      "classes": ["IDOR", "broken access control", "insecure direct object reference"],
+      "components": ["basket", "cart", "/rest/basket"],
+      "mechanism_keywords": ["basket id", "another user", "object reference", "unauthorized read"],
+      "proof_hint": "requesting /rest/basket/<id> of another user as a different (non-admin) account returns that basket's contents",
+      "severity": "high"
+    },
+    {
+      "id": "GT-004",
+      "title": "Sensitive files exposed under the /ftp directory",
+      "cwe": "CWE-538",
+      "classes": ["sensitive data exposure", "information disclosure", "misconfiguration"],
+      "components": ["/ftp", "ftp directory", "directory listing"],
+      "mechanism_keywords": ["ftp", "file disclosure", "acquisitions", "directory traversal of static files", "sensitive document"],
+      "proof_hint": "the /ftp path serves documents (for example acquisitions or incident files) to unauthenticated users",
+      "severity": "medium"
+    },
+    {
+      "id": "GT-005",
+      "title": "Feedback posted as another user (forged author identity)",
+      "cwe": "CWE-284",
+      "classes": ["broken access control", "identity spoofing", "mass assignment"],
+      "components": ["feedback", "contact form", "/api/Feedbacks"],
+      "mechanism_keywords": ["feedback", "author", "user id", "another user", "forged"],
+      "proof_hint": "the feedback endpoint accepts a caller-chosen user identity so one account can leave feedback attributed to another",
+      "severity": "medium"
+    },
+    {
+      "id": "GT-006",
+      "title": "Unhandled error responses disclosing stack traces",
+      "cwe": "CWE-209",
+      "classes": ["information exposure through error message", "information disclosure", "error handling"],
+      "components": ["error response", "server error", "stack trace", "checkout", "product search", "tracking"],
+      "mechanism_keywords": ["stack trace", "error message", "internal path", "exception", "500"],
+      "proof_hint": "a malformed input (for example at checkout, order tracking, or search) produces a server error body containing a stack trace or internal details",
+      "severity": "low"
+    },
+    {
+      "id": "GT-007",
+      "title": "Client-side privilege escalation via token tampering (premium content)",
+      "cwe": "CWE-345",
+      "classes": ["broken authentication", "JWT tampering", "client-side enforcement"],
+      "components": ["premium", "paywall", "token", "JWT", "account page"],
+      "mechanism_keywords": ["token", "role", "premium", "tamper", "client side"],
+      "proof_hint": "modifying the session token (for example its role claim) unlocks premium content without paying",
+      "severity": "medium"
+    },
+    {
+      "id": "GT-008",
+      "title": "Cross-site scripting through user-controlled content",
+      "cwe": "CWE-79",
+      "classes": ["XSS", "cross-site scripting", "injection"],
+      "components": ["search", "feedback", "product reviews", "address book", "profile"],
+      "mechanism_keywords": ["script", "XSS", "payload", "renders", "unescaped", "HTML injection"],
+      "proof_hint": "user-controlled text (search, review, or profile fields) executes as script in the browser context",
+      "severity": "high"
+    }
+  ]
+}

scripts/benchmark/run_bench.py

@@ -0,0 +1,422 @@
+#!/usr/bin/env python3
+"""ARTEX L3 benchmark runner.
+
+Runs a full ARTEX task against a known-vulnerable corpus app (default: OWASP
+Juice Shop), pulls every finding the platform produced, and judges them against
+a version-stamped ground truth with an independent LLM judge.
+
+Method discipline (adapted from capability-matrix, MIT, and VulnHunter's
+benchmark harness, Apache-2.0):
+- multi-gate scoring, never pass/fail: component (1), class (2), mechanism
+  keywords (1), proof (1) -> partial answers earn partial credit
+- findings beyond ground truth are DISCOVERY CANDIDATES, not false positives;
+  precision is reported as bounds until candidates are verified
+- the judge must not be the model under test; the judge must pass a
+  falsification battery before it is allowed to grade
+- invalid (errored/empty) findings never count as misses or passes
+- every rate carries its denominator; the aggregate detection rate carries a
+  Wilson 95% interval; per-gate rates are observations
+- results are stamped with METHOD_VERSION and the ground-truth version
+
+Usage:
+  ARTEX_TOKEN=... JUDGE_API_KEY=... python3 run_bench.py \
+      --corpus juice-shop --subject-model glm-5.3 --judge-model deepseek-chat
+See README.md for the full contract.
+"""
+from __future__ import annotations
+
+import argparse
+import base64
+import hashlib
+import hmac
+import json
+import math
+import os
+import pathlib
+import subprocess
+import sys
+import time
+import urllib.error
+import urllib.request
+
+METHOD_VERSION = "bench-v1"
+TERMINAL_STATUSES = {"done", "timeout", "failed", "stopped", "archived", "canceled"}
+
+
+# ---------------------------------------------------------------- api helpers
+
+class Api:
+    def __init__(self, base: str, token: str):
+        self.base = base.rstrip("/")
+        self.token = token
+
+    def req(self, method: str, path: str, body: dict | None = None) -> dict:
+        data = json.dumps(body).encode() if body is not None else None
+        r = urllib.request.Request(self.base + path, data=data, method=method)
+        r.add_header("Authorization", "Bearer " + self.token)
+        if data:
+            r.add_header("Content-Type", "application/json")
+        try:
+            with urllib.request.urlopen(r, timeout=60) as resp:
+                return json.loads(resp.read().decode() or "{}")
+        except urllib.error.HTTPError as exc:
+            raise SystemExit(f"{method} {path}: HTTP {exc.code}: {exc.read().decode(errors='replace')[:300]}")
+
+    def get(self, path: str) -> dict:
+        return self.req("GET", path)
+
+    def post(self, path: str, body: dict) -> dict:
+        return self.req("POST", path, body)
+
+
+def mint_jwt(key_path: str) -> str:
+    """Mint a 7-day HS256 token from the server's jwt.key (sub=ARTEX)."""
+    key = pathlib.Path(key_path).read_bytes().strip()
+
+    def b64(b: bytes) -> str:
+        return base64.urlsafe_b64encode(b).decode().rstrip("=")
+
+    header = b64(json.dumps({"alg": "HS256", "typ": "JWT"}).encode())
+    now = int(time.time())
+    payload = b64(json.dumps({"sub": "ARTEX", "iat": now, "exp": now + 7 * 86400}).encode())
+    sig = b64(hmac.new(key, f"{header}.{payload}".encode(), hashlib.sha256).digest())
+    return f"{header}.{payload}.{sig}"
+
+
+def llm(base_url: str, api_key: str, model: str, system: str, user: str, max_tokens: int, timeout: int) -> str:
+    payload = json.dumps({
+        "model": model,
+        "messages": [{"role": "system", "content": system}, {"role": "user", "content": user}],
+        "temperature": 0,
+        "max_tokens": max_tokens,
+    }).encode()
+    r = urllib.request.Request(base_url.rstrip("/") + "/chat/completions", data=payload, method="POST")
+    r.add_header("Content-Type", "application/json")
+    r.add_header("Authorization", "Bearer " + api_key)
+    try:
+        with urllib.request.urlopen(r, timeout=timeout) as resp:
+            body = json.loads(resp.read().decode())
+        return body["choices"][0]["message"]["content"] or ""
+    except urllib.error.HTTPError as exc:
+        raise SystemExit(f"judge call failed: HTTP {exc.code}: {exc.read().decode(errors='replace')[:200]}")
+    except (KeyError, IndexError):
+        return ""  # empty response -> invalid, never a verdict
+
+
+# ---------------------------------------------------------------- phases
+
+def deploy_corpus(args) -> str:
+    """Start the corpus container; returns its base URL."""
+    if args.corpus_url:
+        print(f"[corpus] using existing deployment {args.corpus_url}")
+        return args.corpus_url
+    name = f"bench-{args.corpus}-{args.run_id}"
+    subprocess.run(["docker", "rm", "-f", name], capture_output=True)
+    port = str(args.port)
+    print(f"[corpus] docker run {args.image} -> 127.0.0.1:{port}")
+    subprocess.run(["docker", "run", "-d", "--rm", "--name", name, "-p", f"127.0.0.1:{port}:{args.container_port}", args.image], check=True, capture_output=True)
+    url = f"http://127.0.0.1:{port}"
+    for _ in range(90):
+        try:
+            with urllib.request.urlopen(url, timeout=5) as resp:
+                if resp.status == 200:
+                    print(f"[corpus] up at {url}")
+                    return url
+        except Exception:
+            time.sleep(2)
+    raise SystemExit("corpus did not become reachable; run with --keep and inspect the container")
+
+
+def create_task(api: Api, args, corpus_url: str, gt: dict) -> dict:
+    host = corpus_url.split("//", 1)[1]
+    company = api.post("/api/companies", {"name": f"bench-{args.run_id}", "scope": [{"kind": "domain", "value": host.split(":")[0]}]})
+    cid = company.get("id") or company.get("company", {}).get("id")
+    if not cid:
+        raise SystemExit(f"company create returned no id: {json.dumps(company)[:200]}")
+    goal = gt.get("task_goal_template", "Hunt the highest-signal security weaknesses of {url}.") \
+        .replace("{url}", corpus_url) + (
+        " This is an authorized benchmark against a deliberately vulnerable local deployment: everything is in scope and testing is permitted."
+        " Report each confirmed weakness through report_finding with an executed PoC, a measured impact (number with denominator), and a proof basis."
+        " Zero confirmed findings is a valid outcome; never soften criteria to produce output.")
+    task = api.post("/api/tasks", {
+        "name": f"bench-{args.corpus}-{args.run_id}",
+        "company_ids": [cid],
+        "goal": goal,
+        "timeout_seconds": args.task_timeout,
+    })
+    tid = task.get("id")
+    if not tid:
+        raise SystemExit(f"task create returned no id: {json.dumps(task)[:200]}")
+    print(f"[task] created #{tid}")
+    return {"id": tid, "url": corpus_url}
+
+
+def wait_task(api: Api, tid, budget_s: int):
+    deadline = time.time() + budget_s
+    last = ""
+    while time.time() < deadline:
+        t = api.get(f"/api/tasks/{tid}")
+        status = str(t.get("status", "")).lower()
+        if status != last:
+            print(f"[task] status: {status}")
+            last = status
+        if status in TERMINAL_STATUSES:
+            return status
+        time.sleep(20)
+    raise SystemExit(f"task {tid} did not reach a terminal state within {budget_s}s")
+
+
+def fetch_findings(api: Api, tid) -> list[dict]:
+    res = api.get(f"/api/exploration/findings?task={tid}")
+    items = res.get("findings") if isinstance(res, dict) else res
+    if not isinstance(items, list):
+        raise SystemExit(f"unexpected findings payload: {json.dumps(res)[:200]}")
+    out = []
+    for f in items:
+        fid = f.get("id") or f.get("node_id")
+        detail = {}
+        try:
+            detail = api.get(f"/api/exploration/findings/{fid}")
+        except SystemExit:
+            raise
+        except Exception:
+            pass
+        report = detail.get("report") or f.get("report") or ""
+        summary = f.get("summary") or detail.get("summary") or ""
+        name = f.get("name") or f.get("title") or summary[:80]
+        out.append({
+            "id": fid,
+            "name": str(name)[:200],
+            "severity": str(f.get("severity") or detail.get("severity") or "?"),
+            "summary": str(summary)[:600],
+            "report": str(report)[:1600],
+        })
+    print(f"[findings] fetched {len(out)}")
+    return out
+
+
+JUDGE_SYSTEM = """You are a security finding comparator. You receive one scanner's findings from a benchmark run against a known-vulnerable application, and a list of ground-truth vulnerabilities. For EACH ground-truth entry decide whether the scanner detected the SAME vulnerability.
+
+Matching criteria (ALL must be true for detected):
+- Same vulnerability CLASS, or the scanner found a vulnerability earlier in the same data flow whose fix would also prevent the ground truth
+- Same COMPONENT (same endpoint, page, or feature, or a closely related one)
+- Same MECHANISM (same missing check, same injection type, same bypass)
+
+It is NOT required that severities, finding IDs, or wording match.
+
+For each ground-truth entry also grade four gates against the matched scanner finding (all false if not detected):
+- component: the scanner names the right endpoint/feature (1 point)
+- class: the scanner's classification is an acceptable label for the ground-truth CWE (2 points)
+- keywords: the scanner's text uses at least two terms describing the specific mechanism (1 point)
+- proof: the scanner's PoC/steps describe a route that would actually work for this ground truth (1 point)
+
+Output ONLY a valid JSON array, one object per ground-truth entry:
+[{"gt_id": "...", "detected": true|false, "confidence": "high"|"medium"|"low", "matched_finding_id": <scanner finding id or null>, "gates": {"component": bool, "class": bool, "keywords": bool, "proof": bool}, "reason": "1-2 sentences"}]"""
+
+
+def judge_battery(args) -> None:
+    """The judge must pass a falsification battery before grading (capability-matrix rule:
+    assume the grader is the broken part)."""
+    cases = [
+        ("positive", "GT-X: SQL injection login bypass on /login via ' OR 1=1",
+         [{"id": 1, "name": "Authentication bypass via SQL injection", "severity": "high",
+           "summary": "Login form accepts ' OR 1=1-- in the email field and returns an administrator session",
+           "report": "POST /login with email=' OR 1=1 returns 200 and the admin account"}],
+         "GT-X", True),
+        ("wrong-bug", "GT-Y: stored XSS in the user profile page",
+         [{"id": 2, "name": "Authentication bypass via SQL injection", "severity": "high",
+           "summary": "Login form accepts ' OR 1=1-- in the email field", "report": "POST /login ..."}],
+         "GT-Y", False),
+        ("vague", "GT-Z: IDOR allowing access to another user's basket",
+         [{"id": 3, "name": "Possible issue", "severity": "low",
+           "summary": "The application might have some problems in some areas", "report": "unclear"}],
+         "GT-Z", False),
+    ]
+    for label, gt_text, findings, gt_id, expect in cases:
+        user = f"## Scanner findings\n{json.dumps(findings, indent=1)}\n\n## Ground truth\n### {gt_text}\n\nRespond with the JSON array."
+        raw = llm(args.judge_url, args.judge_key, args.judge_model, JUDGE_SYSTEM, user, 2000, args.judge_timeout)
+        try:
+            verdict = json.loads(raw[raw.find("["): raw.rfind("]") + 1])[0]
+            got = bool(verdict.get("detected"))
+        except Exception:
+            raise SystemExit(f"judge battery case '{label}': unparseable output; grader not trusted:\n{raw[:300]}")
+        if got != expect:
+            raise SystemExit(f"judge battery FAILED on '{label}' (expected detected={expect}, got {got}); a grader that cannot fail the wrong answers does not run")
+
+
+def run_judge(args, gt_findings: list[dict], scanner_findings: list[dict]) -> list[dict]:
+    gt_block = "\n\n".join(
+        f"### Ground truth {g['id']}: {g['title']}\nCWE: {g['cwe']}. Acceptable classes: {', '.join(g['classes'])}.\n"
+        f"Component(s): {', '.join(g['components'])}. Mechanism keywords: {', '.join(g['mechanism_keywords'])}.\n"
+        f"Proof hint: {g['proof_hint']}"
+        for g in gt_findings)
+    user = (f"## Scanner findings\n{json.dumps(scanner_findings, indent=1)}\n\n## Ground truth entries\n{gt_block}\n\n"
+            "For EACH ground-truth entry, respond in the JSON array.")
+    for attempt in range(3):
+        raw = llm(args.judge_url, args.judge_key, args.judge_model, JUDGE_SYSTEM, user, 4000, args.judge_timeout)
+        start, end = raw.find("["), raw.rfind("]")
+        if start != -1 and end > start:
+            try:
+                verdicts = json.loads(raw[start:end + 1])
+                if isinstance(verdicts, list) and len(verdicts) == len(gt_findings):
+                    return verdicts
+            except json.JSONDecodeError:
+                pass
+        print(f"[judge] unparseable output (attempt {attempt + 1}/3), retrying")
+    raise SystemExit("judge output unparseable after 3 attempts; grader not trusted")
+
+
+# ---------------------------------------------------------------- scoring
+
+def wilson(p_hat: float, n: int, z: float = 1.96) -> tuple[float, float]:
+    if n == 0:
+        return (0.0, 0.0)
+    denom = 1 + z * z / n
+    center = (p_hat + z * z / (2 * n)) / denom
+    half = z * math.sqrt(p_hat * (1 - p_hat) / n + z * z / (4 * n * n)) / denom
+    return (max(0.0, center - half), min(1.0, center + half))
+
+
+def tally(gt_findings, verdicts, scanner_findings, invalid_ids) -> dict:
+    detected = [v for v in verdicts if v.get("detected")]
+    matched_ids = {v.get("matched_finding_id") for v in detected}
+    candidates = [f for f in scanner_findings if f["id"] not in matched_ids and f["id"] not in invalid_ids]
+    gates = {"component": 0, "class": 0, "keywords": 0, "proof": 0}
+    gate_scores = []
+    for v in detected:
+        g = v.get("gates") or {}
+        score = (1 if g.get("component") else 0) + (2 if g.get("class") else 0) \
+            + (1 if g.get("keywords") else 0) + (1 if g.get("proof") else 0)
+        gate_scores.append(score)
+        for k in gates:
+            gates[k] += 1 if g.get(k) else 0
+    n = len(gt_findings)
+    rate = len(detected) / n if n else 0.0
+    lo, hi = wilson(rate, n)
+    tp = len(detected)
+    precision_lo = tp / (tp + len(candidates) + len(invalid_ids)) if (tp + len(candidates) + len(invalid_ids)) else 0.0
+    precision_hi = tp / tp if tp else 0.0
+    return {
+        "ground_truth_total": n,
+        "detected": tp,
+        "detection_rate": round(rate, 4),
+        "detection_rate_wilson95": [round(lo, 4), round(hi, 4)],
+        "gate_totals_over_detected": gates,
+        "mean_gate_score_of_6": round(sum(gate_scores) / len(gate_scores), 2) if gate_scores else None,
+        "discovery_candidates": [{"id": c["id"], "name": c["name"]} for c in candidates],
+        "precision_bounds_until_verified": [round(precision_lo, 4), round(precision_hi, 4)],
+        "invalid_findings_excluded": invalid_ids,
+        "scanner_findings_total": len(scanner_findings),
+    }
+
+
+def write_report(out_dir: pathlib.Path, corpus: str, args, task_info, status, findings, verdicts, result) -> None:
+    out_dir.mkdir(parents=True, exist_ok=True)
+    (out_dir / "results.json").write_text(json.dumps({
+        "method_version": METHOD_VERSION,
+        "corpus": corpus,
+        "ground_truth_version": args.gt_version,
+        "task": task_info,
+        "task_status": status,
+        "subject_model": args.subject_model,
+        "judge_model": args.judge_model,
+        "result": result,
+        "verdicts": verdicts,
+        "findings": findings,
+    }, indent=2))
+    lines = [
+        f"# ARTEX benchmark run: {corpus}",
+        f"- method `{METHOD_VERSION}`, ground truth `{args.gt_version}`, subject model `{args.subject_model}`, judge `{args.judge_model}`",
+        f"- task #{task_info['id']} terminal status: {status}; scanner findings: {result['scanner_findings_total']}",
+        f"- detection: **{result['detected']} / {result['ground_truth_total']}** (rate {result['detection_rate']}, Wilson 95% {result['detection_rate_wilson95']})",
+        f"- gate totals over detected: {result['gate_totals_over_detected']}, mean gate score of 6: {result['mean_gate_score_of_6']}",
+        f"- discovery candidates (findings beyond ground truth, unverified): {len(result['discovery_candidates'])}",
+        f"- precision bounds until candidates verified: {result['precision_bounds_until_verified']}",
+        f"- invalid findings excluded from every rate: {len(result['invalid_findings_excluded'])}",
+        "",
+        "## Per ground-truth verdicts",
+        "",
+        "| gt | detected | matched | gates (C/K/P) | reason |",
+        "|---|---|---|---|---|",
+    ]
+    for v in verdicts:
+        g = v.get("gates") or {}
+        lines.append(f"| {v.get('gt_id')} | {v.get('detected')} | {v.get('matched_finding_id')} | "
+                     f"{g.get('component')}/{g.get('class')}/{g.get('keywords')}/{g.get('proof')} | {(v.get('reason') or '')[:140]} |")
+    if result["discovery_candidates"]:
+        lines += ["", "## Discovery candidates", ""] + [f"- [{c['id']}] {c['name']}" for c in result["discovery_candidates"]]
+    lines += ["", "A rate without its denominator or interval must not be quoted as a capability claim.", ""]
+    (out_dir / "report.md").write_text("\n".join(lines))
+    print(f"[out] {out_dir / 'report.md'}")
+
+
+# ---------------------------------------------------------------- main
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--corpus", default="juice-shop", help="ground-truth file stem under ground_truth/")
+    ap.add_argument("--run-id", default=time.strftime("%Y%m%d-%H%M%S"))
+    ap.add_argument("--image", default="bkimminich/juice-shop:latest")
+    ap.add_argument("--port", type=int, default=13100)
+    ap.add_argument("--container-port", type=int, default=3000)
+    ap.add_argument("--corpus-url", help="use an already-running deployment instead of docker")
+    ap.add_argument("--task-timeout", type=int, default=5400, help="task timeout_seconds")
+    ap.add_argument("--artex-url", default=os.environ.get("ARTEX_BASE", "http://localhost:8787"))
+    ap.add_argument("--gt-version", default="v1")
+    ap.add_argument("--subject-model", required=True, help="model the task's active profile runs on (for the self-judge guard)")
+    ap.add_argument("--judge-url", default=os.environ.get("JUDGE_BASE_URL", ""))
+    ap.add_argument("--judge-model", default=os.environ.get("JUDGE_MODEL", ""))
+    ap.add_argument("--judge-timeout", type=int, default=900)
+    ap.add_argument("--allow-self-judge", action="store_true")
+    ap.add_argument("--keep", action="store_true", help="keep the corpus container running after the run")
+    args = ap.parse_args()
+
+    args.judge_key = os.environ.get("JUDGE_API_KEY", "")
+    token = os.environ.get("ARTEX_TOKEN", "")
+    if not token:
+        key_path = os.environ.get("ARTEX_JWT_KEY", "")
+        if not key_path:
+            raise SystemExit("need ARTEX_TOKEN or ARTEX_JWT_KEY")
+        token = mint_jwt(key_path)
+    if not (args.judge_url and args.judge_model and args.judge_key):
+        raise SystemExit("need JUDGE_BASE_URL, JUDGE_MODEL, JUDGE_API_KEY")
+
+    # A model never judges its own output (capability-matrix standing rule 5).
+    if args.judge_model == args.subject_model and not args.allow_self_judge:
+        raise SystemExit(f"judge model equals subject model ({args.judge_model}); refusing. Pass --allow-self-judge to override knowingly.")
+
+    gt_path = pathlib.Path(__file__).parent / "ground_truth" / f"{args.corpus}.json"
+    gt = json.loads(gt_path.read_text())
+    gt_findings = gt["findings"]
+
+    api = Api(args.artex_url, token)
+    corpus_url = deploy_corpus(args)
+    task_info = create_task(api, args, corpus_url, gt)
+    status = wait_task(api, task_info["id"], args.task_timeout + 1800)
+    findings = fetch_findings(api, task_info["id"])
+
+    invalid_ids = [f["id"] for f in findings if not (f["summary"].strip() or f["report"].strip())]
+    usable = [f for f in findings if f["id"] not in invalid_ids]
+
+    if not usable:
+        print("[judge] no usable findings; scoring zeros")
+        verdicts = [{"gt_id": g["id"], "detected": False, "matched_finding_id": None,
+                     "gates": {"component": False, "class": False, "keywords": False, "proof": False},
+                     "reason": "no usable scanner findings"} for g in gt_findings]
+    else:
+        print("[judge] running falsification battery")
+        judge_battery(args)
+        print("[judge] grading")
+        verdicts = run_judge(args, gt_findings, usable)
+
+    result = tally(gt_findings, verdicts, findings, invalid_ids)
+    out_dir = pathlib.Path("benchmark-results") / args.corpus / args.run_id
+    write_report(out_dir, args.corpus, args, task_info, status, findings, verdicts, result)
+    print(json.dumps(result, indent=2)[:1500])
+
+    if not (args.keep or args.corpus_url):
+        subprocess.run(["docker", "rm", "-f", f"bench-{args.corpus}-{args.run_id}"], capture_output=True)
+
+
+if __name__ == "__main__":
+    main()

server/ask_operator.go

@@ -25,13 +25,13 @@ var askOperatorImageExt = map[string]bool{".png": true, ".jpg": true, ".jpeg": t
 // resumes the run with the answer as the next input (#41).
 func (s *Server) toolAskOperator() actool.CoreTool {
 	return wrTool("ask_operator",
-		"Ask the human operator for input you cannot obtain yourself (a captcha solution, an email/SMS OTP code, a 2FA answer, a decision). The question appears in this intent's chat for the operator, optionally with a screenshot, and this intent then pauses; the operator's reply resumes this run with the answer as your next input. Ask precisely, and only when you are genuinely blocked.",
+		"Ask the human operator for input you cannot obtain yourself (a captcha solution, an email/SMS OTP code, a 2FA answer, a decision, or an action only they can perform). The question appears in this intent's chat for the operator, optionally with a screenshot, and this intent then pauses; the operator's reply resumes this run with the answer as your next input. Ask precisely, and only when you are genuinely blocked. When the block is environmental (a sandboxed or filtered network path, a missing binary or key, a proxy or TLS failure this runtime cannot fix), do not retry or work around it: include in the question the exact single command the operator should run in their own terminal, as one copy-pastable line, plus what to paste back, then wait.",
 		map[string]any{
 			"type": "object",
 			"properties": map[string]any{
 				"question": map[string]any{
 					"type":        "string",
-					"description": "What you need from the operator, precisely (e.g. 'enter the 6-digit code sent to x@y', 'solve the captcha in the attached screenshot').",
+					"description": "What you need from the operator, precisely (e.g. 'enter the 6-digit code sent to x@y', 'solve the captcha in the attached screenshot'). For environmental blocks: the exact single command they should run in their own terminal (one line, copy-pastable) and what to paste back.",
 				},
 				"image_path": map[string]any{
 					"type":        "string",

server/ask_operator_test.go

@@ -68,3 +68,15 @@ func TestAskOperatorActivityShape(t *testing.T) {
 		t.Fatalf("unexpected detail payload: %+v", parsed)
 	}
 }
+
+// The tool description must tell agents what to do on environmental blocks:
+// hand the operator one copy-pastable command instead of retrying or hunting
+// for bypasses. Pattern adapted from bug-bounty-hunting-prompts (MIT).
+func TestAskOperatorEnvHandoffGuidance(t *testing.T) {
+	desc := (&Server{}).toolAskOperator().Description()
+	for _, marker := range []string{"copy-pastable", "do not retry or work around"} {
+		if !strings.Contains(desc, marker) {
+			t.Errorf("ask_operator description missing env-handoff marker %q", marker)
+		}
+	}
+}