Home › Entities › Three-Layer Verification

Three-Layer Verification

The employee-only post-edit verification system in claude-code, gated behind USER_TYPE === 'ant'. Addresses the documented 29-30% false-claims rate on the current model (Capybara v8).

The Problem

In the external build, Claude Code's success criterion for a file write is binary: did bytes reach disk? No compilation check. No type check. The model reports "Done!" and the next action begins. Internal data shows approximately one in three task completions contains at least one incorrect claim about what was accomplished.

The Three Layers

The source comment at services/tools/toolExecution.ts:149-157 explains the motivation:

"Self-verification doesn't work for LLMs. An agent that just made a change is biased toward believing it worked. The 29% false-claims rate from BQ proves this."

Layer Role What It Does
1. Agent Worker Makes the code change
2. Verifier Checker Re-reads files, runs tests, checks for TypeScript errors
3. Spot-check Auditor Examines the verifier's evidence independently

Nobody trusts anybody without proof. The agent doesn't verify itself. The verifier doesn't go unchecked. Each layer operates on the output of the previous layer.

The Verification Steps

In the ant build, toolExecution.ts injects an additional post-edit block. The agent is instructed to:

  1. Re-read each file it claimed to have edited
  2. Run any available test suite
  3. Execute scripts and confirm their output
  4. Check that no new TypeScript errors were introduced
  5. Report only if all checks pass

The CLAUDE.md Bypass

The verification instructions themselves are not secret — only the automated injection was gated. Community member @iamfakeguru published a CLAUDE.md snippet that manually reconstructs the ant-tier verification behavior. Within 48 hours, multiple CLAUDE.md templates incorporating the pattern had been starred over 2,000 times on GitHub.

Measurable Consequences

GitGuardian's 2026 report quantified the impact of missing verification: - Claude Code-assisted commits: 3.2% secret leak rate (vs 1.5% baseline) - At peak (August 2025): 31 leaked secrets per 1,000 commits (2.4x human baseline) - 28.65 million new hardcoded secrets on public GitHub in 2025 (+34% YoY)

Key Claims

Sources