Measuring Verifier-Reported Verification Load on a Single LLM Monologue Corpus: A Report with Termination Controls

Under a fixed protocol P, one strict LLM verifier per claim emits a verdict and new claims, judges which are load-bearing, applies a removal test and deduplicates its own output; the per-act result is the verifier-reported offspring count ô, with self-adjudication part of the construct. On one informal single-model monologue corpus, ô sat near 2–3 through sampled depths to 4 over selected paths (not an asymptote); termination controls read about 0.14 and stopped unaided. One claim compiled to a deterministic priced check has a terminal execution sub-act with zero emitted prose obligations, on a separate axis from ô. An earlier branching/queue theorem is retired by scheduler counterexamples. Every abstract claim is a row of a claim ledger decided by a closed-manifest gate that is shipped with the paper. At deposit, with sigma-glyph 0.6.7, the gate reports 2 rows CHECKED and 8 REFUSED with typed reasons (frozen act corpus not deposited, scheduler simulation not deposited, a drifted receipt source, measurement not replayed). This version is deposited with those refusals as known loss, by the author's decision. NOT peer-reviewed; all verifiers are one vendor's models; drafted with language-model assistance, reviewed by other language models (Codex/OpenAI, Kimi, Qwen, Claude), no independent human review to date.