FP v0.5.0 — Evidence-Based Optimization

What changed

SKILL.md: 162 → 77 lines — templates 35 → 7, on-demand profiles, zero ceremony creep.

fp-minimal — 3-rule variant for non-reasoning models (gpt-5.3, etc.)

Evidence (1,416 real LLM API calls)

Model	Winner	Score
gpt-5.3-codex-spark	v-minimal	3.08
gpt-5.6-sol	v-final (77 lines)	3.57
deepseek-v4-pro	v-final (77 lines)	3.14

E2E multi-turn: 45% fewer tokens, 89% fewer FP template reads, correct on-demand profile triggering.

Templates: 28 dead templates archived. Agents no longer waste tokens reading them by default.

Key insights

• Less is more only on weak models. Reasoning models need structure.
• Simulation predicted v7 wins at 4.73. Real eval: v0 baseline won.
• Prompt interference is real: adding CRITICAL modifiers degrades performance.

Full report: benchmarks/results/ARTICLE.md