---
title: How Do Dark Factories Validate Without Reading Code?
slug: how-do-dark-factories-validate-without-reading-code
type: query
created: 2026-06-17
updated: 2026-06-17
status: active
tags:
  - query
  - validation
  - code-review
  - gates
sources:
  - dan-shapiro-you-dont-write-code-2026-02-13
  - jleechanorg-dark-factory-repo
  - jleechanorg-agent-orchestrator-repo
related:
  - code-review-vs-validation
  - digital-twin-universe
  - cxdb
  - healer
confidence: medium-high
contradictions:
  - Automated validation can reduce code reading, but security-sensitive changes still need human-owned policy and periodic audit.
bead_refs:
  - dark-factory-rr4
---

# How Do Dark Factories Validate Without Reading Code?

They replace routine diff reading with layered outcome evidence.

## Validation Stack

- Complete specs with no hidden requirements.
- Public tests that agents can use while building.
- Sealed holdouts that implementing agents cannot read.
- Independent adversarial review by a different backend or model family.
- CI, evidence gates, and reviewer verdicts.
- Digital twins for realistic external-system behavior.
- [[cxdb]] event logs and [[healer]] diagnosis for failure memory.
- Human gates for high-risk, ambiguous, or policy-sensitive decisions.

## Key Point

`inference:` The validation stack must be designed before code generation. If validation is added after a generated patch exists, humans end up reading code again because the factory did not define enough evidence up front.