This paper develops a constructive no-meta, observable-only framework for autonomous discovery, evaluator redesign, target expansion, constitutional drift, and self-modification under public evidence alone. The core object is a coded evaluation package with public transcript laws, feasible public actions, semantic target realizations, verifier-induced replay semantics, and transcript-indexed completion operators. Within this setting, the paper defines robust public risk and derives selector-based envelope identities, actionwise worst-case decomposition, local minimax duality, a common-action witness margin, and an oracle disambiguation gap for ambiguity-reduction limits. The paper then studies exact and approximate comparison between packages under constitutional extension. It proves no-improvement and non-worsening results under structural simulation, isolates deterministic public coarsening as a correct no-improvement case, shows how latent-only internal computation collapses to public randomization under action-pasting closure, and introduces legacy anchors, bounded coherent cocycles, constitutional atlases, and safe carrier quotient compression to preserve certified semantic control across long-lived system change. At the constructive layer, proposal triples, comparator domains, and benchmark covers are generated by predictable constitutional operators from public generator states rather than externally declared meta-evaluators. Replay semantics are verifier-induced; generator upgrades are judged by compiler-aware public-coding criteria; admissible target morphisms are constrained by witnessed predictive geometry; public carrier growth is controlled by time-filtered distributional or macro-action fallback; and constitutional exhaustion is defined through certified frontier non-improvement together with proof-of-search and residual-randomness certificates. The result is a first-principles framework for auditable observable invention and self-extension without hidden states or privileged referees. Archived on Arweave (Data Tx ID) PDF: ₛmjYCQQgmb89paQbaFZgwvNPHbnt2t4sjJvmJBLkGsTeX source: tiRygGDkDwnU5R8U9cPeOXWxx5jd-JbFyxlVBIFliBI
K Takahashi (Wed,) studied this question.