Autoresearch

Thirteen interactive angles on the autonomous AI loops that iterate on d3-power-tools visualizations.

Autoresearch is a simple idea: let an AI propose a code change, run it, compute a score, have an auditor judge whether the result actually improved, and keep or discard the change. Repeat. d3-power-tools applies the pattern to D3 visualizations, using four quality dimensions and a composite score to decide what survives.

The first seven pieces explain how the loop works today. The last six explore ways to make it work better: vision auditors, multi-objective rewards, smarter proposer strategies, compressed context, ensemble auditing, and cross-block transfer learning.

How Autoresearch Works

Making It Better

The Vocabulary

Autoresearch has five recurring roles. Every glyph on this page colors its elements by these roles, and every article uses the same palette in its figures.

propose
The proposer writing code, the source of change.
edits · diffs · candidate blocks · strategies · forward flow from proposer to auditor
score
The composite number that decides fate, the quality measurement.
composite score lines · Pareto frontiers · score trajectories · threshold lines · feedback from audit to proposer
audit
The auditor judging, the eye looking at the result.
vision auditors · ensemble auditors · audit boxes · judgment arcs
keep
A change that survived, above threshold, not dominated.
kept paths · rising trajectories · non-dominated Pareto points · passing checkmarks
discard
A change that got rolled back, below threshold or dominated.
rejected proposals · declining trajectories · dominated Pareto points · failures