Can sequencing edits across three surfaces avoid hidden costs?

MetaRSI-v1 treats a deployed model as a triple of three writable surfaces — data, harness (the execution scaffold: system prompt, memory, built-in tools, skills, MCP resources), and model weights — and argues that editing only one of them always leaves a cost the others can't see. The paper states it plainly: "each single-surface operator closes a valid loop on its own surface but incurs a cost it cannot see": harness edits are "re-paid as context at every inference," data amplification "cannot reach capability the rollouts never exhibit," and training "may regress already-held behaviour." Its fix is a four-step composed sequence — "extend the scaffold, amplify what it reveals, internalize into weights, then retire the extension" — which the paper says "retains the capability while returning the scaffold budget to baseline," paying neither cost.

The three operators — Data-RSI, Harness-RSI, Model-RSI — share "one loop kernel and artifact vocabulary" and are linked by "Transition Agent-v1 adapters" that pass one compiled learning signal between them, so a change on one surface becomes legible input to the next. Above the operators, an "RSI2 Agent-v1" decides two things: the order operators run in ("some orders are ill-posed") and each operator's own proposal policy. The paper is explicit that composition is only a reallocation, not a source of new capability: "the system composed of these operators alone is a closed amplifier, whose ceiling is the best arrangement of what was already present," and "every genuine gain requires information originating outside the loop" — a verifier's ruling, a retrieved document, an instrument reading, or "a commissioned expert note," all entering "on equal terms" through the learning signal.

This sharpens How does the substrate change which behaviors an optimizer can reach? by naming exactly which persistence and cost trade each surface makes — harness: no training cost, re-paid every inference; data: cheap, bounded by what rollouts exhibit; weights: persistent, risks regression — and by proposing typed adapters as the mechanism for moving changes between surfaces instead of leaving them siloed. It also confirms Can models reliably improve themselves without external feedback? with a more mechanistic argument: even a system that composes all three surfaces, not just a single self-refinement loop, is still bounded by what the model already holds and still needs an external signal to produce genuine gain. And it extends What separates self-improvement from policy improvement? by giving a concrete reason the external-standard dial matters operationally: the paper's "three separations" are "what let a loop be trusted when the verifier is weak," because "they keep the loop from improving its own definition of success."

The excerpt validates MetaRSI-v1 only "on code and closed-form science, with no external teacher," so the claim that this composition "must next operate across real, diverse scientific, engineering, and meta-scientific domains" is the paper's stated target, not a result shown here. The harder claim — that the framework works where verification is expensive or partial — is argued for but not demonstrated in what's excerpted. The "closed amplifier" ceiling is likewise asserted rather than measured: no bound, benchmark gap, or failure case is given to show how tightly composition is constrained by "what was already present." Until a genuinely open-ended domain is run through it, the paper's contribution should be read as a typing and scheduling framework for existing single-surface techniques, not evidence that composing them unlocks capability single-surface RSI cannot reach.