rhi

Appearance

Domain generator corpus ​

Project(s) touched: all (claude-config layer); meta-investigation lives in github-io

Status: Open

Surfaced in: docs/automated-introspection/investigations/2026-05-20-whats-wrong/synthesis.md, "Hypotheses no one tested but the data implies" (H-DOMAIN-GENERATOR-COMPRESSION); the May 13 "Counterweight: don't fake confidence" rule (7521987c) is held up as the working case

The question ​

Would a generator-corpus (worked examples + counterexamples) outperform a rule-corpus (CLAUDE.md prose) for the constraint classes that keep failing — no-specialcase, no-bandaid, specialcasing-vs-generalization?

Why this is a registry thread ​

Every repo's CLAUDE.md is downstream of the answer. If examples beat rules for these classes, the propagation tooling and the shape of the ecosystem-common region both change.

What's still open ​

  • No comparison of rule-corpus vs generator-corpus on the failing classes.
  • The working case (7521987c, "don't fake confidence") is a single data point; bandaid-style rules are the failing cases.

Working answer ​

None — hypothesis only.