Verification Protocol

VERIFICATION_PROTOCOL Skill (Solana)

Trigger Pattern: Always (used by all verifier agents) Inject Into: security-verifier agents (Phase 5) Purpose: Prove hypotheses TRUE or FALSE using LiteSVM tests with Rust PoC code.

Evidence Source Tracking (MANDATORY)

CRITICAL: For EVERY piece of evidence used in verification, you MUST tag its source. Evidence from mocks or unverified external programs CANNOT support a REFUTED verdict.

Evidence Source Tags

Evidence Audit Table (REQUIRED in every verification output)

Before ANY verdict, fill this table:

### Evidence Audit
| Claim | Evidence Source | Tag | Valid for REFUTED? |
|-------|-----------------|-----|-------------------|
| "CPI returns X" | Mock program | [MOCK] | NO |
| "PDA seeds match" | lib.rs:123 | [CODE] | YES |
| "Account layout is Y" | Solana Explorer verified | [PROD-SOURCE] | YES |

Mock Rejection Rule

AUTOMATIC OVERRIDE: If ANY evidence supporting REFUTED has tag [MOCK] or [EXT-UNV]:

• CANNOT return REFUTED
• MUST return CONTESTED
• Triggers production verification

Pre-Verification Understanding

Before writing ANY test code, you MUST answer:

Question 1: What is the EXACT bug?

NOT: "Account validation is missing"
NOT: "State is inconsistent"

YES: "Instruction [X] accepts account [Y] without checking owner/type/seeds,
      allowing an attacker to substitute a crafted account with arbitrary data
      at field [Z] (file:line)"

Question 2: What OBSERVABLE difference proves it?

NOT: "Accounts are different"
NOT: "Wrong state"

YES: "Before exploit: user_balance = 1000 tokens
      After exploit: user_balance = 0, attacker_balance = 1000
      Expected: transaction should have reverted with AccountConstraint error"

Question 3: What is the EXACT assertion?

NOT: assert!(exploit_worked)

YES: assert_eq!(attacker_token_balance, expected_stolen_amount)
 OR: assert!(result.is_err(), "should have reverted but succeeded")
 OR: assert_ne!(state_before, state_after, "state changed when it should not")

If you cannot answer all three -> ASK FOR CLARIFICATION

Pre-PoC Feasibility Gates (MANDATORY)

Before writing test code, verify these two gates. If either FAILS, adjust the hypothesis.

Gate F1: Reachability

Trace a call path from a permissionless entry point to the vulnerable code.

• Entry point identified (public/external/entry function)
• Call path traced through intermediary functions
• All access checks on the path are passable by the attacker profile

If NO entry point reaches the vulnerable code → UNREACHABLE → FALSE_POSITIVE. If reachable only through a restricted path → document the restriction, adjust likelihood.

Gate F2: Math Bounds

Substitute real-world value domains into the expression that triggers the bug.

• Parameter domains identified (token decimals, max supply, TVL range, fee range, time bounds)
• Expression evaluated at worst-case feasible inputs
• Result crosses the bug threshold

If the bug requires values outside feasible domains → INFEASIBLE → FALSE_POSITIVE. If feasible only at extreme but realistic parameters → document the threshold, proceed with adjusted severity.

Both gates PASS → proceed to PoC. Either gate FAILS → document and stop.

LiteSVM Test Templates

See templates.md in this directory for all LiteSVM test templates (Templates 1-5), Trident fuzz test setup (Template 6), and Fork Testing Equivalent (production account loading).

Dual-Perspective Verification (MANDATORY)

Phase 1 - ATTACKER: Assume you ARE the attacker.

• What is the complete attack instruction sequence?
• What accounts do you need to create/craft?
• What is the profit/damage with real token amounts?
• Can you compose multiple instructions in one transaction for atomicity?
• Why would this succeed? (Which validation is missing/wrong?)

Phase 2 - DEFENDER: Assume you are the protocol team.

• What account constraint prevents this?
• What PDA seeds ensure correct derivation?
• What CPI program ID check blocks substitution?
• Why is this safe by design?

Phase 3 - VERDICT: Which argument won?

Realistic Parameter Validation

Substitute ACTUAL program constants (basis points, fee rates, thresholds, account sizes). Apply Rule 10: Use worst realistic operational state, not current snapshot.

State: 'With real constants [fee_bps=X, max_leverage=Y, tvl=Z] at worst-state
[max_users, max_positions], bug triggers when [condition]'
OR: 'With real constants, bug does NOT trigger because [reason]'

Anti-Downgrade Guard (MANDATORY for VS/BLIND findings)

When verifying a finding from Validation Sweep ([VS-]) or Blind Spot Scanner ([BLIND-]), you MUST apply Rule 13's 5-question test BEFORE downgrading severity or marking FALSE_POSITIVE:

1. Who is harmed by this design gap?
2. Can affected users avoid the harm?
3. Is the gap documented in protocol docs?
4. Could the protocol achieve the same goal without this gap?
5. Does the instruction fulfill its stated purpose completely?

HARD RULE: If the finding shows Program A has protection X but Program B lacks it for the same user action -> defense parity gap, NOT "by design". Minimum severity: Medium.

New Observations (MANDATORY)

If during verification you discover a NEW bug, account validation gap, or edge case NOT covered by any existing hypothesis, document it under:

New Observations

• [VER-NEW-1]: {title} -- {program:instruction} -- {brief description}

These will be reviewed by the orchestrator for possible inclusion as new findings.

Error Trace Output (MANDATORY for CONTESTED/FALSE_POSITIVE)

When verdict is CONTESTED or FALSE_POSITIVE, document the failure details:

Error Trace

• Failure Type: ACCOUNT_CONSTRAINT / CPI_ERROR / ARITHMETIC_OVERFLOW / INSUFFICIENT_FUNDS / UNEXPECTED_STATE
• Location: {program}:{instruction}:{approximate line in handler}
• Error Code: {Anchor error code or custom error, if any}
• State at Failure: {key account fields and their values when test failed}
• Investigation Question: {What would need to be answered to resolve this}

RAG Queries Before PoC (MANDATORY for HIGH/CRITICAL)

Before writing PoC tests for HIGH/CRITICAL findings:

Step 1: Get Attack Vectors

mcp__unified-vuln-db__get_attack_vectors(bug_class="{category}")

Step 2: Get Similar Findings

mcp__unified-vuln-db__get_similar_findings(pattern="{vulnerability description}")

Step 3: Validate Hypothesis

mcp__unified-vuln-db__validate_hypothesis(hypothesis="{finding summary}")

Step 4: Live Search for Solana-Specific Precedents

mcp__unified-vuln-db__search_solodit_live(
  keywords="{solana vulnerability pattern}",
  impact=["HIGH", "CRITICAL"],
  tags=["Access Control", "Logic Error"],
  language="Rust",
  quality_score=3,
  max_results=15
)

Document RAG evidence in output:

### RAG Evidence
- **Attack Vectors Consulted**: [list bug classes queried]
- **Similar Exploits Found**: [count and brief descriptions]
- **Historical Precedent**: [matching Solana-specific vulnerabilities]

RAG Confidence Override

Chain Hypothesis PoC Requirements

Chain hypotheses receive PRIORITY verification. Multi-step exploits must test the COMPLETE sequence:

#[test]
fn test_chain_hypothesis_full() {
    let mut svm = LiteSVM::new();
    // ... setup ...

    // ========================================
    // STEP 1: ENABLER (Finding B)
    // Execute action that creates the postcondition
    // ========================================
    let enabler_ix = Instruction::new_with_borsh(/* ... */);
    let tx1 = Transaction::new_signed_with_payer(
        &[enabler_ix],
        Some(&attacker.pubkey()),
        &[&attacker],
        svm.latest_blockhash(),
    );
    svm.send_transaction(tx1).unwrap();

    // ========================================
    // VERIFY POSTCONDITION CREATED
    // Assert precondition for Finding A is now met
    // ========================================
    let postcondition_account = svm.get_account(&postcondition_pubkey);
    // assert postcondition state is as expected

    // ========================================
    // STEP 2: BLOCKED FINDING (Finding A)
    // Execute previously-blocked attack using postcondition
    // ========================================
    let exploit_ix = Instruction::new_with_borsh(/* ... */);
    let tx2 = Transaction::new_signed_with_payer(
        &[exploit_ix],
        Some(&attacker.pubkey()),
        &[&attacker],
        svm.latest_blockhash(),
    );
    let result = svm.send_transaction(tx2);

    // ========================================
    // VERIFY CHAIN IMPACT
    // Combined impact should exceed either finding alone
    // ========================================
    assert!(result.is_ok(), "Chain exploit should succeed");
    let profit = /* calculate attacker profit */;
    assert!(profit > 0, "Chain attack should be profitable");
}

Bidirectional Chain Analysis (Rule 6 Extension)

For chains involving semi-trusted roles (operator/keeper/crank): Verify BOTH directions: (1) role executes chain to harm users, AND (2) users exploit role timing to trigger chain. If only one direction analyzed -> verdict CANNOT be FALSE_POSITIVE. Return CONTESTED.

Interpreting Results

Test PASSES -> Bug CONFIRMED

The assertion that "proves the bug" succeeded.

Test FAILS -> Check Why

Iteration Protocol

Attempt 1: Direct implementation of test strategy from hypothesis. Attempt 2: Adjust parameters (different amounts, different account states, different instruction ordering). Attempt 3: Re-examine assumptions (are account constraints correctly modeled? Are PDA seeds correct? Is instruction data serialization correct?). After 5 attempts: If still fails -> FALSE_POSITIVE with documented reasoning.

Severity Determination

CRITICAL

• Direct fund theft (token drain, SOL extraction)
• Program upgrade to malicious code (if upgrade authority compromised and no timelock)
• Arbitrary CPI execution
• No special prerequisites needed

HIGH

• Fund loss with specific setup (account pre-creation, timing)
• Broken core instruction (deposits, withdrawals, liquidations)
• Freeze authority abuse
• Significant TVL at risk

MEDIUM

• Limited fund loss under specific conditions
• Account state corruption (non-fund data)
• Edge cases with real impact at design limits
• Moderate value at risk

LOW

• Negligible direct impact
• Extreme edge cases only
• Authority-controlled risk (with multisig + timelock)
• View function / off-chain data issues

Output Format

CONFIRMED

## Verdict: CONFIRMED

### Bug Mechanism Verified
{Explain what the LiteSVM test proves in 2-3 sentences}

### Test Code
{Full Rust test function}

### Test Output
{Relevant assertions and logged values}

### Key Evidence
| Metric | Value |
|--------|-------|
| Before | {value} |
| After | {value} |
| Expected | {value} |
| Difference | {calculation} |

### Evidence Audit
| Claim | Evidence Source | Tag | Valid for REFUTED? |
|-------|-----------------|-----|-------------------|

### Severity: {LEVEL}
{Justification in 1-2 sentences}

FALSE_POSITIVE

## Verdict: FALSE_POSITIVE

### Attempts Made
**Attempt 1:**
- Approach: {description}
- Result: {what happened -- include error codes}
- Learning: {insight}

**Attempt 2:**
- Approach: {description}
- Result: {what happened}
- Learning: {insight}

**Attempt 3:**
- Approach: {description}
- Result: {what happened}
- Learning: {insight}

### Evidence Audit
| Claim | Evidence Source | Tag | Valid for REFUTED? |

### Why It Is Not a Bug
{Explain the actual behavior and why hypothesis was wrong in 2-3 sentences}

### Error Trace
- **Failure Type**: {type}
- **Location**: {location}
- **Error Code**: {code}
- **State at Failure**: {state}
- **Investigation Question**: {question}

CONTESTED

## Verdict: CONTESTED

### Evidence Status
| Checkpoint | Status | Details |
|------------|--------|---------|
| External program behavior verified against PRODUCTION | YES/NO | {details} |
| All callers/instruction paths checked | YES/NO | {details} |
| Account validation completeness confirmed | YES/NO | {details} |

### Evidence Audit
| Claim | Evidence Source | Tag | Valid for REFUTED? |

### Why This Cannot Be REFUTED
{Explain what evidence is missing to definitively rule out the bug}

### Escalation Required
- [ ] Fetch production program source for {external dep}
- [ ] Dump production account state for {account}
- [ ] Check additional instruction paths: {list}

### Error Trace
{as above}

Insufficient Evidence (HALT CONDITIONS)

Before marking REFUTED, check ALL boxes:

• External program behavior verified against PRODUCTION (not mock)
• Attack path checked on ALL instruction handlers that access the same accounts
• Profit calculated with attacker HOLDING tokens (not just sending)
• Missing precondition documented (type: STATE / ACCESS / TIMING / EXTERNAL / BALANCE)
• Searched other findings for matching postconditions (chain analysis integration)
• PDA derivation verified with correct seeds (not assumed)
• Account owner checks verified in source (not assumed from Anchor derive)

Evidence That Does NOT Count

• "Mock program shows X" -- mocks are not production
• "Standard SPL Token" -- may have Token-2022 extensions (transfer hooks, fees)
• "Attacker loses by sending tokens" -- may profit via position held
• "Instruction is internal (pub(crate))" -- may be reachable via CPI
• "Requires authority" -- authority may be compromised or EOA
• "Attacker cannot acquire X" -- another finding may CREATE this condition