This paper formalizes a “verification-limited” scaling regime in which candidate proposals proliferate faster than they can be certified under observable-only evidence, in settings with no privileged judge (“no-meta”). The core goal is to define progress, charging, and diagnostics that remain audit-ready and non-gameable even under adversarial missingness, split views, and underspecified rules. The manuscript is explicitly layered: (1) Theory / principle layer (normative semantics). We define an observable-only decision discipline, close-pinned cutoff semantics (a head commitment plus an explicit leaf-index bound), and a strict-only progress credit rule. A central result is an observable-only law of no progress inflation under weakening: removing observations cannot increase strictly credited progress, because strict credit requires mechanically checkable completeness (certified cutoffs, pinned policy identifiers, index-unique receipt selection, cutoff-parameterized workload sealing, bounded proofs and bounded replay, and non-circular closure hashing). Uncertain components are mapped to conservative canonical forms, and the verifier may fail closed by outputting NONSTRICT or UNDEFINED. (2) Reference mechanism layer (one realizable instantiation). We provide a replay-auditable construction built from dependency-closed policy pinning; witness-cosigned dual-root transparency-log heads (an append-only list root plus a key-index root); mechanically checkable uniqueness via index-membership proofs (not “uniqueness by absence”); workload sealing via a log-committed roster plus an index-certified admission count parameterized by the same cutoff bound; escrow reservation with replay-validated metered charging and cap fallback when receipts/proofs/dependencies are missing/invalid/non-unique; and a deterministic derived transcript with streaming size bounds and memory-safe fail-closed triggers. Finally, the paper introduces conditional “acceleration envelopes” over spans of windows where observable premises hold (e.g., budget stability, utilization and strict-fraction lower bounds, and an empirically testable power-law upper bound on charged verification cost per admitted attempt as a function of cumulative committed strict progress). Importantly, interpreting these envelopes as intelligence/capability acceleration requires an epoch-defined semantic link between evidence and performance (“semantic adequacy”); otherwise the progress variable is only a strictly auditable counter. To support liveness without weakening strict guarantees, the paper also isolates a non-strict escrow layer: escrow tickets and (optional) non-strict escrow leaves can record work products and enable compute reuse, while being formally projected out of strict credit and strict diagnostics. Operational profiles (e.g., spam/Sybil resistance, minimum fees, and rate limits) are treated as non-strict ingestion policy, not as strict-proof rules.
K Takahashi (Sat,) studied this question.