The fixed "eval function" — like prepare.py's evaluate_bpb
ICP Relevance25%
Pain Point Specificity25%
Clarity & Readability20%
Call-to-Action Strength15%
Urgency / Motivation15%
Content
v0 — baseline
Ready — configure ICP and click Start
0 / 10 iterations
Experiment Log
—
ICP Score
—
Best
—
Improvement
Relevance
—
Pain Points
—
Clarity
—
CTA
—
Urgency
—
Experiments will appear here as the optimizer runs.
Each iteration: hypothesis → rewrite → score → keep/discard
Optimization Progress
Click "Start Optimization Loop" to begin
The AutoResearch Pattern for Content
Inspired by Karpathy's AutoResearch, this applies the same autonomous experimentation loop to content optimization:
AutoResearch
ExpertSwarm
train.py
Your content
val_bpb metric
ICP alignment score
prepare.py (eval)
Scoring rubric + LLM judge
program.md
ICP definition + strategy
5 min GPU time
~10 sec per iteration
The loop: modify content → score against ICP rubric → keep if improved, discard if not → repeat. Each change is evaluated against the same fixed standard, preventing drift.