Self-reference classics, read into the harness
Parent: recursive-harness · Reference: self-reference-classics
Each classic self-reference tradition maps onto a harness piece:
- Rational metareasoning → our
P×Cbudget over decisions and deduction search: satisfice on your own computation. - Good Regulator / IMP → the observability ceiling (convergence-needs-an-observable-target): control needs a model; an unmodelable hidden variable breaks regulation.
- Gödel machine ("modify only under proof") → our credentialing / verify-before-commit; it avoids the unsound drift of current self-evolution (self-evolution-is-commodity); the computability core is the Kleene recursion theorem (recursion-is-a-phase-transition).
- Bellman contraction → literally our "convergence = contraction" (recursion-convergence-contraction).
Two payoffs
- Self-evolution drift = a second-order-cybernetics problem: the evaluator is inside the system (self-judge) → can't measure itself. The classic fix — separate the meta-level / the observer — is exactly
/goalusing a separate Haiku judge to break the self-reference (drift-is-world-wandering, weak-auditor-by-design). We re-derived an old result. - The Gödel machine's "modify only under proof" is what convergent self-evolution should look like: sound self-modification (prove, then change) vs today's feel-good unsound self-evolution that drifts. Same line as verify-before-commit / learn-only-against-a-pinned-world.
Two most useful anchors for "doing work": rational metareasoning (decide how much to think) + Gödel-machine-style sound self-modification (prove before you change) — the harness has both (P×C / verify-before-commit).