Hostile reading of the builder-loop comparison
Question attacked
Does the comparison survive the strongest reading of the experimental records, or does it obtain Commonplace's advantage by calling every editable prompt, agent, or harness part "inside a fixed organization"?
The strongest counterexample is not the Darwin Gödel Machine alone. Continual Harness creates, edits, and deletes sub-agent definitions, and the paper's GPP record includes a structural rewrite in which per-decision logic moved into a master agent that dispatches to named sub-checks. Autogenesis versions and reuses agent prompts, tools, and code, while its bus architecture is designed to let participants be replaced. Self-Harness can propose subagent and middleware structure, although its reported subagent and skill branches were rejected and the retained structural change was narrower middleware. These are organizational changes, not merely parameter changes inside otherwise untouched roles.
What fails
A binary classification of an entire system as having either an internal or external builder loop fails. Redesign reach is relative to an aspect and horizon. Continual Harness demonstrates a repeatable path for reorganizing sub-agent roles while leaving the four-part harness partition, Refiner, evaluator, and reward design supplied. Autogenesis demonstrates versioned revision of agent implementations while leaving the five-resource ontology, named specialist arrangement, bus protocol, evaluator, and acceptance rule supplied. The Darwin Gödel Machine demonstrates inherited agent-code reorganization while leaving the population controller, admission rule, objective, and evaluator outside descendant edits.
The earlier phrase "the Darwin Gödel Machine is the important partial exception" is therefore false by omission. Several systems internalize narrower builder loops. What none of the five records demonstrates is a retained path for revising its reported objective or evaluator; their outer update protocols also remain supplied.
What survives
The operative-path criterion survives if it is applied to a named redesign class rather than to a whole system. It also needs a sharp distinction between an internal builder-level event and an internal builder loop:
- Representation, reachable evidence, identifiable authority, installation, and later dependence establish that one redesign became operative inside the declared system.
- Continued availability of the resulting organization, plus whatever evidence or rationale the path needs for another challenge, establishes a retained loop rather than a one-off intervention.
This formulation classifies the counterexamples rather than excluding them. Continual Harness can satisfy the loop criterion for sub-agent definitions. Autogenesis can satisfy it for enabled agent resources. DGM can satisfy it for descendant agent code. Their fixed outer machinery then becomes the next named redesign class to test.
The ordinary research-organization objection also survives. An issue, design record, review decision, CI check, merge, and later deployment can satisfy the same obligations. If that workflow's own rules are represented and revisable through the retained path, it is another internal builder loop. Commonplace is evidence for one implementation, not a technologically unique category. The published-paper comparison can say only what the reported experimental pathways establish; it cannot infer the absence of unreported laboratory machinery.
The newly captured Agno practitioner report is a useful hostile control. Its coding agent can alter instructions, tools, parameters, and code, but the target specification, probe derivation, judge, coding agent, platform architecture, and stopping rule remain fixed. The author explicitly distinguishes this convergent fitting loop from improvements to improvement ability. Broad code access therefore does not by itself defeat the criterion.
Further correction
"Something must stand outside each change" is easy to misread as requiring a permanently external component. The causal requirement is temporal: incumbent conditions govern each transition. A current acceptance rule, authorization, update law, or scope boundary determines whether a candidate becomes the successor; a later transition may replace that condition under the conditions then incumbent. This clarification belongs inside the builder-loop note for now. It needs its own theory note only if another artifact needs to cite it independently.
Integration consequences
- Replace whole-system internal/external language with aspect-bounded redesign reach.
- Credit Continual Harness, Autogenesis, Self-Harness, and DGM for the organizational changes they actually report.
- Reserve the comparison for the remaining fixed machinery: evaluators, objectives, update protocols, resource ontologies, and population controllers.
- Keep the Commonplace cases differently calibrated and retain the evidence-asymmetry caveat.
- Add the builder-loop note to the self-improving-systems curated head.