Machinery persists by warrant, not position, in a reflective loop

Type: kb/types/note.md · Tags: learning-theory, deploy-time-learning, self-improving-systems

Sutton's Bitter Lesson has two modes, and its folk application remembers only one. The apply-mode is a selection rule over existing methods: prefer the techniques "that have been shown to scale." The build-mode is in Sutton's own conclusion — "time is better invested in finding simple scalable solutions" — and it is what compatibility means where no scaled method exists yet: building one. The two-axis reading locates the localized forms in exactly that situation.

The build-mode's own history shows that fixed human design and computationally produced competence can coexist. Every victory in Sutton's list uses human-designed machinery whose task content is found with computation: Deep Blue's alpha-beta search, the hidden Markov models that beat hand-crafted speech pipelines, and the convolutional architecture that beat hand-coded features. Sutton's closing makes the intended split explicit: “we should build in only the meta-methods that can find and capture this arbitrary complexity… We want AI agents that can discover like we can, not which contain what we have discovered.” The load-bearing question is therefore whether a fixed component implements a general production method over the declared reach or instead supplies the task- or family-specific competence the method claims to acquire.

Sutton's geometry fixes an outside; reflection makes selected machinery addressable

That prescription has a clean geometry: the meta-method sits outside the content its update rule learns. The researcher designs the architecture; the architecture learns the content; and the fixed meta-method also supplies gradient descent's stability because the machinery doing the selecting is not selected by that same update.

A reflective system does not erase its declared boundary. It is reflective only relative to selected aspects. When types, gates, validators, skills, or the loop's own instructions are represented and connected to a revision path, those pieces of machinery become addressable through the same retained forms as other content. That possibility is already operative in bounded cases: the traced tag-readme episode produced validator machinery from operational strain, oracle accumulation is machinery-production as a standing channel, and the symbolic layer can be a learning target with codification as its write path. Architectural position alone therefore does not justify exemption from a revision claim. It also does not require every component to enter the update surface: an unrepresented or protected kernel may remain.

What replaces the outside

Three substitutions, each already carried by a standing claim:

  • Provenance, update placement, and actor allocation remain separate. "Hand-crafted" says who produced the current version. It does not say whether the artifact is currently revisable or whether a human performs its operation. A bootstrap may move production of selected content and machinery into computation without requiring every inherited component to move.
  • Exemption by position becomes persistence by warrant where the claimed path reaches. A gate may survive because its criterion keeps discriminating, a type because its carve keeps earning use, or a kernel because exactness, safety, or bounded cost warrants protection. Applying an assessed standard to consequential free choices, machinery included, is this note's conjecture. The case-level diagnosis of a structured method's loss under scaling states when unsupported proxy scope may explain a loss and disclaims the converse guarantee that assessed structure persists.
  • The declared frame replaces the fiction that reflection has no outside. Objectives, commitments, authority grants, and fixed general learning machinery may remain supplied. The computational-closure account names those crossings rather than pretending they disappeared. Fixedness becomes a reach limit when the protected component embeds task- or family-specific production knowledge that a broader claim requires computation to acquire, as stated by the pressure for agentic factory development.

The cost is the fixed point

Gradient descent buys stability from a fixed meta-method. A reflective path that also places its evaluator or update machinery inside its revision surface gives up part of that fixed point and faces the trusting-trust problem with fuzzier tools than Thompson had. Governance must then carry the lost stability: the current candidate does not author its own admission decision, acceptance is allocated through an authority appropriate to the domain, changes stay localized and reversible, and accumulated oracles need an authority path that does not depend on the candidate currently under judgment.

That is a cost of recursive revision, not a requirement that every reflective system pay it everywhere. A fixed higher-level gate or trusted kernel may be the warranted design. If a separate authorized path later revises it, that path needs its own evidence, boundary, and protection; if no claimed path needs to revise it, fixed placement is not deferred human work.

Scope

  • The claim is about what licenses a consequential fixed choice, not a demand for universal revisability. One traced instance of loop-produced machinery plus one standing channel does not make the loop the main producer of its own machinery today.
  • Fixed provenance, location outside one update space, and current human decision supply are different properties. A fixed computational rule can be computationally closed for the scope it settles.
  • Bitter-Lesson pressure is strongest where human-authored task- or family-specific production knowledge claims broad reach. Fixed general machinery may remain when its generality and constraints are part of the stated claim and survive relevant tests.

Open Questions

  • Which general, safety, or exactness obligations warrant a fixed kernel, and how should its claimed reach be tested?
  • What warrant would license moving a piece of machinery's production into the loop, as distinct from its revision — the migration-earned criterion applied to authorship rather than judgment?

Relevant Notes: