fix: add falsification discipline to security audit template

manifest.yaml

@@ -1351,7 +1351,7 @@ templates:
         Security audit of code or a system component. Systematic
         vulnerability analysis with severity classification.
       persona: security-auditor
-      protocols: [anti-hallucination, self-verification, operational-constraints, security-vulnerability]
+      protocols: [anti-hallucination, self-verification, operational-constraints, adversarial-falsification, security-vulnerability]
       taxonomies: [stack-lifetime-hazards]
       format: investigation-report
 

protocols/analysis/security-vulnerability.md

@@ -48,6 +48,21 @@ For every external input:
    re-encoded, decoded, or transformed before use.
 4. Check for **integer overflow/underflow** in size or length parameters
    derived from external input.
+5. **Validation provenance check**: For every candidate vulnerability at
+   a use site, trace **backward** to find all validation that may have
+   occurred between the input's entry and the use. Check:
+   - Caller functions that decode or validate before passing values
+     (e.g., a header-decode helper that checks `Field > MaxAllowed`
+     before the caller uses `Field` in arithmetic)
+   - Helper functions called earlier in the same code path
+   - API postconditions — when a system API (e.g., kernel API, stdlib)
+     returns success, what guarantees does it make about output buffer
+     contents, lengths, or value ranges?
+   - Initialization-time invariants — constructor or init functions
+     that constrain values to safe ranges (e.g., power-of-2
+     requirements, maximum bounds)
+   Do NOT report a finding until you have confirmed that no upstream
+   validation neutralizes it.
 
 ## Phase 3: Authentication and Authorization
 

protocols/guardrails/self-verification.md

@@ -36,6 +36,12 @@ presenting it as final. Treat it as a pre-submission checklist.
   - Does the evidence actually support the conclusion stated?
 - If any sampled item fails verification, **re-examine all items of
   the same type** before proceeding.
+- For each sampled finding, apply **symmetric falsification**: attempt
+  to disprove the finding with the same rigor you applied when
+  falsifying candidate findings that you concluded were safe. Ask:
+  "Is there a validation, API contract, or initialization invariant
+  I missed that makes this safe?" If you cannot answer "no" with
+  specific code evidence, downgrade or remove the finding.
 
 ### 2. Citation Audit
 

templates/investigate-security.md

@@ -12,6 +12,7 @@ protocols:
   - guardrails/anti-hallucination
   - guardrails/self-verification
   - guardrails/operational-constraints
+  - guardrails/adversarial-falsification
   - analysis/security-vulnerability
 taxonomies:
   - stack-lifetime-hazards