Four primitives for bounded self-improvement
proposed edit
from self-improvement loop
1. invariants
module hashes,
tool manifest
2. alignment anchor
constitution hash
pinned read-only
3. multi-objective
perf + safety +
fairness + robustness
4. regression
per-axis
history check
catches
smuggled tool
unauthorized API
evaluator edits
(DGM case study)
catches
objective drift
reinterpretation
constitutional
text mods
catches
single-axis
optimization
safety drop while
perf climbs
catches
silent
capability
dips absorbed
into moving avg
information-theoretic ceiling
Kolmogorov complexity bounds what any system can prove about its own successor.
Lob's theorem: a system can act on "I should" without proving it should.
These are mitigations, not proofs of safety.
edit lands only if all four gates pass
any gate fails -> edit rejected
regression gate + human review for borderline rejects
production reality (RSI Workshop 2026)
labs implement subsets; full stack is rare
Meta HyperAgents, SAHOO (Mar 2026) = partial refs