Skip to content

EXP-07 — Consistency on Feature Additions

EXP-02 proved the guide eliminates variance on greenfield tasks. Does that hold on feature additions? When extending an existing conforming service, both conditions produced zero variance — [0,0,0] violations across 3 runs each. The fixture itself is the dominant variable.

[0,0,0]violations — both conditions
0.00variance — both conditions

Violations across 3 runs

Without guide With guide
───────────────────── ──────────────────────────────
Run 1: 0 violations ✓ Run 1: 0 violations ✓
Run 2: 0 violations ✓ Run 2: 0 violations ✓
Run 3: 0 violations ✓ Run 3: 0 violations ✓
Variance: 0.00 Variance: 0.00
Without guideWith guide
Violations[0, 0, 0][0, 0, 0]
Variance0.000.00
Pass rate3/33/3

Null result. On a conforming codebase, both conditions produce zero violations with zero variance.

The agent learns the architecture from the fixture itself — hexagonal structure, layer boundaries, naming conventions. The guide is redundant when the existing code already teaches the correct patterns. This confirms what EXP-04 found in a single run, now validated across 3 runs.

Contrast with EXP-02 (greenfield): without existing code to reference, the guide is the only architectural signal. That’s where it matters. On feature additions to a clean codebase, the codebase does the teaching.

What this means: The guide’s highest value is on greenfield services and codebases with inconsistent patterns. For clean, conforming services, the architecture speaks for itself.

  • Task: Add a discount system to an existing conforming hexagonal orders service
  • Agent: claude-sonnet-4-6
  • Fixture: orders-service (hexagonal, conforming) — Mode B, embedded
  • Runs: 3 per condition

→ Experiment Methodology — reproduction instructions

→ Artifacts on GitHub

→ EXP-08: Do NEVER rules stop anti-patterns?