Skip to content

DARK-FACTORY.md

Latest commit

 

History

History
94 lines (63 loc) · 3.82 KB

DARK-FACTORY.md

File metadata and controls

94 lines (63 loc) · 3.82 KB

🏭 Dark Factory Build Log

Agent X-Ray v1.0 was built entirely by the Dark Factory — an autonomous AI build system that orchestrates 6 specialist agents through a checkpoint-gated pipeline with sealed-envelope testing.

No human wrote the initial code. A human described the goal. The factory built v1.

Post-Factory Changes

The following features were added after the original Dark Factory build:

Version Feature Description
v1.1 Keyword Density Ratio Anti-gaming: penalizes keywords in thin sentences (<10 words)
v1.1 --profile flag Built-in scoring profiles: balanced, security, creative, ci-gate, assistant
v1.1 --strict flag Optional LLM evaluation of keyword quality (OpenAI-compatible)

These changes were human-directed enhancements built on the factory's foundation.

Run Details

Field Value
Run ID run-20260224-195250
Mode Full (6 phases)
Model Claude Sonnet 4.6
Sealed Hash sha256:6864da25...
Final Gap Score 0% (171/171 tests passing)

Pipeline

Phase 0 — Factory Setup

Initialized run directory, sealed envelope directory, SQL tracking tables.

Phase 1 — Product Specification

The Product Manager agent translated the goal into a 163-line PRD with 42 functional requirements and 17 acceptance criteria. No code, no architecture — just what to build.

Phase 2 — Architecture + Sealed Tests (parallel)

Two agents worked simultaneously:

  • Architect produced ARCH.md (409 lines) — component design, data flow, file structure
  • QA Sealed wrote 79 acceptance tests in a sealed envelope — hashed and locked before the engineer ever saw the spec

The sealed hash was recorded. No one could tamper with the tests after this point.

Phase 3 — Implementation

The Lead Engineer built the full CLI (578 lines) plus 92 open tests — all passing. The engineer never saw the sealed tests.

Phase 4 — Sealed Validation

The sealed envelope was opened. All 171 tests (92 open + 79 sealed) were run against the implementation.

Result: 12 failures (7% gap). The sealed tests caught issues the engineer's own tests missed:

  • --fleet --json parsing bug
  • Abbreviated column headers vs. expected full names
  • 3-digit hex badge color vs. 6-digit standard

Phase 5 — Hardening (2 cycles)

The Lead Engineer received only failure descriptions (test name + expected + actual) — never the sealed test source code. Two hardening cycles fixed all 12 issues.

Final gap score: 0% — all 171 tests passing.

Phase 6 — Delivery

Code delivered to working directory. Factory floor cleared.

What is Sealed-Envelope Testing?

The QA agent writes tests before the engineer builds anything. Those tests are hashed and locked — the engineer never sees them. After implementation, the seal is broken and the tests run against the code.

This catches the blind spots that happen when the same person (or AI) writes both the code and the tests. In this build, sealed tests caught 12 issues that open tests missed.

Timeline

Phase Duration Artifacts
Phase 1 — Spec 45s PRD.md
Phase 2a — Sealed QA 6m 34s sealed-tests.mjs, fixtures/
Phase 2b — Architecture 1m 44s ARCH.md
Phase 3 — Build 3m 00s agent-xray.js, tests.mjs
Phase 4 — Validation 2m 00s GAP-REPORT.md
Phase 5 — Hardening 4m 00s Patched source
Phase 6 — Delivery 1m 00s ✅ Shipped
Total ~19 minutes

Built on the factory floor. Tested behind a sealed envelope. Shipped with zero gaps. 🏭