The target is institutional learning, not an autonomous self-modification mandate. Research should be able to produce validated improvements to its own routing, formalization, evidence handling, execution, review, and repair machinery.
MOONSHOTS · Research that improves its own machinery
Self-Improving Research Infrastructure
A governed feedback loop where successful research can improve the representations, methods, tools, and infrastructure used by later research.
OBJECTIVE06 / 08
PROGRAMMoonshots
STATUSLong-horizon research goal
IMPORTANT LIMITThis is a direction, not a completed result.
This page describes the goal, work that already exists, and the main unresolved problems. It does not claim scientific validation, deployed capability, external adoption, market value, or inevitability.
Moonshots indexWHAT THIS GOAL MEANS TODAY
The current interpretation, footholds, and open problems.
Any such change remains bounded by provenance, validation, rollback, authority, and human consequence gates. The central risk is self-confirming drift: machinery that optimizes for its own internal signals while degrading contact with external evidence.
Improvement must remain reversible and inspectable
Required guardrails
The machinery should be able to learn from use without making its own success criteria unchallengeable.
- 01
Separate genuine process improvement from local metric optimization.
- 02
Promote new machinery through explicit review rather than silent replacement.
- 03
Regression-test new machinery against prior successful and failed cases.
- 04
Preserve provenance across versions so changed behavior remains reconstructable.
- 05
Retain rollback and human authority for consequential changes.
PROGRAM CONTEXT