poc-tier

name	poc-tier
description	Classify PoC file quality as Tier 1-4. Run after exploiter completes, before triager-sim. Matches "PoC tier", "PoC quality", "poc quality", "exploit verify"
user-invocable	true
argument-hint	<poc-file-path>
allowed-tools	["Read","Bash","Grep"]

PoC Quality Tier Classification

CRITICAL RULES (NEVER VIOLATE)

1. Tier 3-4 = submission FORBIDDEN — PoC without execution evidence is 100% Informative
2. No real network call + response capture = cannot be Tier 1 — local mock = Bronze at best

Automatically classifies PoC quality as Tier 1-4. Prevents Tier 3-4 submission failures (Veda, Katana, etc. — 4 failures).

Input

• $ARGUMENTS: PoC file path (e.g., evidence/ssrf/poc.py, poc/foundry-test/test/Exploit.t.sol)

Tier Definitions

Tier	Name	Requirements	Outcome
1	Gold	Runtime verified + integration test + evidence capture + UA fingerprint	ACCEPT (high confidence)
2	Silver	Executable script + output capture, no integration test	ACCEPT (moderate confidence)
3	Bronze	Script exists but output is theoretical/mock	DROPPED — submission forbidden
4	Reject	No PoC, pseudocode only, "left as exercise"	DROPPED — submission forbidden

Few-Shot Examples

Tier 1 (Gold) — Real API call + response capture + assertion

resp = requests.get(f"{BASE}/api/users/2", headers={"Authorization": f"Bearer {token_user1}"})
assert resp.status_code == 200
data = resp.json()
assert "email" in data  # user1 can read user2's PII
print(f"[VULN] Cross-user data: {data['email']}")  # captured in output.txt

Tier 3 (Bronze) — Script exists but mock output

# This should work against the production API
url = "https://example.com/api/admin"  # placeholder
# TODO: replace with actual target URL
resp = requests.get(url)  # hypothetical response would contain admin data
print("Exploitation successful")  # no actual evidence

Procedure

Step 1: Read PoC File

Read $ARGUMENTS

Step 2: Positive Signal Detection (+1 tier each)

Network calls present: !grep -cE "requests\.|fetch\(|curl |remote\(|cast send|cast call|axios\.|http\.|urllib" "$ARGUMENTS" 2>/dev/null || echo "0"

Real response capture present: !grep -cE "response\.|status_code|\.json\(\)|recvline|recvuntil|interactive|200 OK|HTTP/" "$ARGUMENTS" 2>/dev/null || echo "0"

Test framework usage: !grep -cE "forge test|npm test|pytest|unittest|assert|vm\.expect|console\.log" "$ARGUMENTS" 2>/dev/null || echo "0"

Step 3: Negative Signal Detection (-1 tier each)

Incomplete markers: !grep -ciE "TODO|FIXME|theoretical|hypothetical|would work|should work|left as exercise|mock|placeholder" "$ARGUMENTS" 2>/dev/null || echo "0"

Hardcoded mock data: !grep -cE "fake_|mock_|dummy_|example\.com|0xdead|placeholder" "$ARGUMENTS" 2>/dev/null || echo "0"

Commented-out core logic: !grep -cE "^#.*exploit|^#.*send|^#.*remote|^//.*attack" "$ARGUMENTS" 2>/dev/null || echo "0"

Step 4: Tier Calculation

base_tier = 2  (Silver default)

# Positive signals
if network_calls == 0: tier += 1  (Silver→Bronze)
if response_capture > 0: tier -= 1  (can upgrade)
if test_framework > 0: tier -= 0.5

# Negative signals
if incomplete_markers > 0: tier += 1
if mock_data > 2: tier += 1
if commented_out > 0: tier += 1

tier = clamp(tier, 1, 4)

Step 5: Output

[POC-TIER] File: <path>
[POC-TIER] Positive signals: network_calls=N, response_capture=N, test_framework=N
[POC-TIER] Negative signals: incomplete=N, mock_data=N, commented_out=N
[POC-TIER] Tier: N (<name>)
[POC-TIER] Result: PASS (Tier 1-2) / BLOCK (Tier 3-4, submission forbidden)

BLOCK Actions

• Tier 3: "Script exists but no execution evidence. Run it and capture output.txt"
• Tier 4: "No PoC. Write exploit code first. Theoretical descriptions are 100% Informative"

DeFi PoC Extra Checks

• vm.deal() used + no honest disclosure → downgrade to Tier 3
• fork-url + real block number → Tier 1 signal
• assert proving profit → Tier 1 signal

REMINDER: Tier 3-4 = NEVER submit. Only Tier 1-2 pass to triager-sim.