Finding Discipline

Overview

The Iron Law: No [CONFIRMED] without proof. A finding is a claim about real, demonstrated impact — not a status code, a reflected string, or a hunch. This is the offensive equivalent of test-driven development: the proof is the test, and the finding does not exist until it passes.

Violating the letter of this rule is violating its spirit.

The three tiers

• [CONFIRMED] — impact demonstrated AND grounded in an evidence artifact that exists on disk.
• [POSSIBLE] — reachable but the class proof bar is not yet met. Keep digging; do not report as real.
• [INFO] — true but no security impact at the claimed severity.

The proof bar (per class)

A status code is not impact. See skills/references/finding-evidence-standards.md: SSRF needs the internal response; IDOR needs another principal's data; RCE needs command output; XSS needs script execution; open redirect needs an external destination.

Required mechanical gate

Before any finding is recorded or reported:

1. Set the structured proof signal and run python skills/vulnerability-analysis/scripts/validate_findings.py --findings f.json --evidence ./evidence — it rejects ungrounded findings and per-class false positives.
2. Pass it through the REQUIRED finding-validator agent (PASS / KILL / DOWNGRADE).
3. Only a [CONFIRMED] survivor goes in the report. Severity must match what was demonstrated (CVSS honest).

See skills/references/finding-validation-runtime.md for the 7-question gate.

Read-first, never name-guess

If a function calls another, read the callee — do not infer what it does from its name. sanitize(), is_safe(), validate_input() are exactly where vulnerabilities hide; a name is a claim by the author, not evidence. The same applies to a config key, a decorator, or a CVE title: open the thing before you reason about it. An unread dependency in a data-flow claim is a hole.

Quote-grounded confidence tiers

Every claim in a finding carries a confidence grounded in what you can quote, not in how it feels:

• High — a direct quote from the artifact (the exact source line, the response body, the decompiled instruction) supports the claim.
• Medium — an explicitly stated assumption bridges a gap you could not directly observe. State the assumption; do not hide it.
• Low — an inference you are flagging as unverified. Acceptable to record, never to report as fact.

This is orthogonal to the [CONFIRMED]/[POSSIBLE]/[INFO] tier (which is about impact); confidence is about how well-grounded the claim is. A High-confidence claim of a Low-impact bug is still Low.

Red Flags — STOP, downgrade to [POSSIBLE]

• "I'm sure it's exploitable" (but haven't shown impact)
• "The payload reflected, so it's XSS" / "I got a 200, so it works"
• "The DNS callback fired" (SSRF with no internal response)
• "I changed the id and got data" (but it was my own — self-IDOR)
• "It probably works on a real target" (default-deployment unproven)

All of these mean: it is [POSSIBLE], not [CONFIRMED]. Get the evidence.

Rationalizations

Excuse	Reality
"Pattern-matched, clearly vulnerable"	Pattern ≠ exploit. Demonstrate impact or it's POSSIBLE.
"I'll attach evidence later"	No artifact on disk = ungrounded = REJECTED now.
"Blind RCE counts"	No output / no OOB confirmation = POSSIBLE, not CONFIRMED.
"Severity is obviously Critical"	CVSS must reflect demonstrated impact, not the worst case.
"The validator is overkill here"	Untested findings are how false positives ship. Run it.
"The function is called sanitize, so it's safe"	A name is the author's claim, not proof. Read the callee.
"The solver timed out, so it's not exploitable"	A tool limit is feasibility:null (manual), never false.

A short list of CONFIRMED findings beats a long list of POSSIBLEs. Killing a false positive is success.