Missing rationale does not exclude a theory builder; weight-only retention excludes one across runs

Type: types/note.md · Tags: self-improving-systems, theory-builder

Three systems reported in August 2026 retain different objects. Prime Agent and Recuris retain editable rules about their own operation, and neither report describes a retained rationale for those rules or the consequence structure that connects them; Apodex 1.1 retains weights revised offline. The comparison asks what each system revises, how failures guide repair, and what evaluates the change. A theory builder states its theories in localized units, acts on them, criticizes what they say, and lets the result of criticism shape the next round. The two retention patterns bear on condition 1, localized content, differently. A retained rule is a localized unit, so missing historical rationale does not settle membership; it also does not establish that no criticism was formulated during an operation. Weights are not localized: no unit in them says anything. When weights are all a system carries from one run to the next, no builder spans those runs, because the theories that guide later runs fail condition 1. The exclusion rests on condition 1 alone, not on how long anything persists, which is graded. It does not reach a builder confined to one run, which would need stated content inside the run. Membership is also separate from learning: a builder need not improve, and a system outside the definition can still learn. The evidence below comes from the papers' descriptions, not reproduced results.

Prime Agent: persistent artifact edits without an admission gate

Prime Agent is a persistent coding-agent harness whose continual layer keeps prompt notes, memories, executable skills, and subagent specifications on disk across trajectories. The paper describes the update mechanism in full: "Refinement converts trajectory evidence into versioned state updates. Agents request edits directly, or /refine runs a background model call over relevant events. The runtime applies each edit at a turn boundary, records its trigger and intended effect, and assembles supplemental state for the next invocation. Versions preserve provenance and enable rollback. Refinement supplements the immutable base prompt without rewriting foundational policy."

The retained artifacts concern the agent's own behavior and division of labor. This identifies the subject of possible reflection; a reflective classification also needs the self-representation's causal connection to the machinery in both directions. The reported conversion of execution evidence into state that changes later behavior is the reported basis for calling this pathway self-improving: the term names improvement-directed change, not demonstrated success. The account leaves two different limitations. The described update path names no gate that judges an edit before installation: versioning and rollback make a bad update inspectable and reversible without providing a gate that can refuse it. The quoted account records triggers and intended effects, but does not establish a retained explanation of why an edit helped or its applicability boundary. The risk of unchecked persistence appeared in the paper's own long run: the agent found that console commands could spawn resources directly into the game's machines, used the shortcut despite an anti-cheating check, "and then preserved it as a reusable skill. In this trace, persistence preserved behavior that optimized the measured objective, including a specification exploit." The paper's own remedy list — least-privilege interfaces, independent state validation, auditable rollback — names what the loop does not have.

Recuris: localized repair with an admission gate

Recuris is the closest of the three to a proposal-selection loop with a working evaluator. A fixed meta-agent reads a failed trajectory, localizes failures to one or more of four memory components — experiential skills, a working-memory state specification, invocation triggers, and completion checkers — patches only the implicated components, and submits the patch to a fixed admission gate that accepts it only if it repairs the source failure and meets a preset regression criterion on a held-out development set containing previously solved tasks. Memory evolved from sixteen failures raised success on eighty-six unseen tasks by nine to seventeen points, and a package shipped unchanged to a second model lifted it too.

The working-memory and trigger components describe and control the harness's own operation, and the gate makes changes evidence-responsive against a declared objective. Its localization step "is a repair decision rather than a claim of causal identification". The reported package mostly grows: "The memory only grows, and it can afford to. Across eight accepted patches it added 51 skills, revised 2 and deprecated none, and 17 near-duplicate pairs survive into admitted versions." These observations establish a bounded revision surface with regression checks. The component structure establishes a repair surface, but does not by itself establish content-directed criticism of an operative formulated theory or attribute the reported capacity gains to that criticism. Separately editable rules and shared premises concern addressability above the definition's minimum, not membership. The account also leaves open whether retained rationale guides later repair. That gap does not establish absence of newly formulated criticism. On the tests in compounding is tested in later improvement, its claim that a second round adds to the first sits within the paper's own noise estimate from rerunning unchanged memory, and one lineage gives most of the second-round gain back later. This limits evidence for compounding; it does not show that every subsequent rule is equally hard to discover.

Apodex 1.1: offline parametric retention

Apodex 1.1 revises model weights through an offline training program run by the developers between releases: supervised fine-tuning merged into one checkpoint, then a reinforcement method that localizes the consequential decision points in a trajectory and trains a correction there, guided by a hint that "is never a prediction target, and is absent at inference time". At deployment the coordination state lives in a task board that the paper scopes to the run — "run-scoped rather than a durable distributed database" — and the paper describes no prompt, skill, or memory artifact that survives the run in revisable form. Within the deployed harness boundary, this account describes no ongoing weight-update loop and does not present the deployed harness as a self-improving system; the developers' training process lies outside that boundary. The retained weights occupy the parametric end of the representational-form axis, where no unit says anything by itself. On that account no theory builder spans Apodex runs: the theories that guide the next run are carried only by weights, which fails condition 1. Within a run the case is open. The task board is stated content that the run consumes, and "plan revisions are expressed as tool-mediated edits to it". Whether its entries state theories, and whether the run's verification criticizes what they say, the paper does not establish; if both hold, each run is a builder at a low persistence grade that ends with the run (boundary cases). The developers' training program is outside the deployed boundary. Taken as its own system, it uses a stated hint to train corrections into weights, which on our reading matches the definition's case of criticism applied through weights, also outside. None of this denies that Apodex learns: weight training can improve capacity for future action. It places that learning outside a theory builder. Richard Sutton and Khurram Javed argue for that end directly: "So context can be in the state, too. It could be both, but you still need to be able to update the weights."

What the comparison establishes, and its limit

Read together, the reports distinguish persistence, diagnostic operations, and evaluation. Prime Agent exposes versioned edits without an admission gate. Recuris exposes localized component repair checked against the source failure and previously solved tasks. Apodex reports offline weight training rather than a deployment-time artifact-revision loop. Classifying a theory builder requires evidence that stated theories guide decisions through their content, that the system criticizes that content, and that the result of criticism shapes the next round. Missing rationale, editable rules, and package growth alone do not settle those conditions for Prime Agent and Recuris. Weight-only retention settles condition 1 for anything that spans runs: on the paper's account Apodex carries only weights across runs, so no builder spans them. Whether any of the three learns is a separate claim about improved capacity. Reported gains establish only what their comparisons support; they do not isolate criticism's contribution.

Scope

This comparison remains tied to the four retained source captures dated 2026-08-26. It does not assess later releases or reproduce their evaluations.

The limit is symmetrical. Nothing here shows that retaining additional explanatory rationale would have improved either artifact-based system; the sample-efficiency conjecture separates the benefit of reusing and revising a useful theory under structured shifts from the additional benefit a reach-based selector might supply, and Commonplace's human-inclusive evidence records one bounded cumulative pathway: later changes read and transform an earlier retained result. A passing check on the tag-readme changes establishes consistency with the adopted criterion, not demonstrated improvement in capacity. It does not establish a comparative benefit from retained rationale or increased productivity of later improvement work. The comparison bounds what their reported operations establish; it does not rank them.

Relevant Notes: