Checked procedures, manual reuse, and a selection gap.

CE-EXEC-01: generation and manual reuse

Four local CPU model conversations generated procedures for two policies, comparing ordinary sources with manually prepared CE context. Both conditions had the same read, search and bounded execution tools, six-turn caps and 3,200-generation-token caps. All four procedures passed frozen hidden-input checks and separate source/code meaning review. Each condition completed two of two activities.

Two accepted ordinary procedures were retained under a predeclared selection rule. A human-directed root selected and dispatched them on four new-input batches; a nonauthor consumer used them on two more. All six were correct, with zero model calls for dispatch. Two isolated source-section changes caused refusal before execution. Selection was manual; automatic successor regeneration and general dependency maintenance were not demonstrated.

Prepared execution used 21,103 reported prompt-plus-generation tokens versus ordinary 11,151. Engineering, preparation, review, warmup and energy costs were not fully measured. There is no demonstrated end-to-end savings. The retained review records 189 passing tests plus recovery; that supports implementation correctness within its checked scope, not general agent efficacy.

CE-REUSE-02: negative study with separate review

The frozen four-attempt comparison finished and received separate review within the research effort on September 30, 2026. Ordinary source inventory was compared with the same inventory plus both accepted skill cards and code, using common tools and new inputs. Each condition accepted one of two activities: deduplication succeeded; quality classification failed.

No agent called run_skill, read evidence or searched it. All chose to generate code. The quality attempts invented thresholds and also encountered execution restrictions. Citations named available sources but do not prove evidence consumption. Skill availability used 11,818 reported model tokens versus ordinary 11,317, 4.43% more. Autonomous selection and efficiency were not demonstrated. This is a negative study accepted for its recorded evidence, not acceptance of all agent answers.

Source, correctness criteria and scope

The local source was Stanford CS336 assignment 4, printed edition 26.0.1, Spring 2026, preserved with SHA256 2b0660d0d1e96366c4688f9606c707c668a1fa76396c563b21a6fc75863fb59e, plus explicit local exercise policies and input schemas. Quality candidates were checked on 22 hidden rows each; dedup candidates on seven hidden corpora each, alongside 14 oracle controls. Review assessed policy meaning separately from output equality.

These are two known public policies and a finite local cohort, using qwen2.5-coder:14b on CPU. They do not establish new-domain transfer, general curriculum learning, source truth, reviewer authentication, a security sandbox, scalable intake economics or integrated Library-to-agent operation. Changes in tools and budgets prevent single-cause comparison with earlier context-only studies. The older compact-inventory flywheel comparison is a separate research line; these studies do not establish its success or a company-wide autonomous loop.

What a visitor can reproduce

The aggregates above were checked by inspecting retained local output/check records, recorded nonauthor reviews and process validations. The private receipts, source data and runtime are not published. Visitors cannot reproduce these private outcomes from the public ledger or fixture.

The pinned September 27 ledger checks public process records and integrity. The separate released synthetic fixture is runnable and passed nine hosted OS/Python jobs. Those jobs demonstrate synthetic replay, not private product efficacy or external human acceptance.

Explore the product, run the public fixture, or inspect the research record.