← Research map
Working noteBecome

Fixed-Memory Comparison

At a fixed number of persistent memory states, does an explicitly organized actuator model help across new actuator changes?

Read Online Or start with the explanation below.

A proposed test, carried out in one model

Reflective Self-Models found that an explicitly organized self-model performs exactly as well as a compiled controller given the same evidence. It proposed a further test: fix the memory budget and ask whether an explicit actuator model helps across new actuator changes.

This working note carries out that test in one finite model. The model, the memory budget, the controller classes, the loss and the rules for the verdict were written down and approved before any computation.

What was compared

An actuator has a polarity, normal or reversed, that may flip at every step. The controller sees only a noisy residual of each action and must guess the polarity for the next one. It may keep at most K states of memory from one step to the next, with K = 2, 3 or 4, over 8 steps.

The comparison was made under two conventions. In the first the controller has a free step counter, so its tables may differ from step to step. In the second it has none, and one table is used at every step. “Explicit” means something different in each:

  • With a step counter, an explicit controller is a calibrated belief register. Each memory state carries the actual probability of the polarity given that state. The belief is updated by Bayes’ rule, allowing for a possible flip, and the updated beliefs are rounded monotonically to at most K states, so that each state receives a run of neighbouring beliefs. The rounding points may move with the step.
  • Without a step counter, an explicit controller is a fixed-grid belief register. It has K fixed levels, makes the same update and moves to the nearest level. The levels are design parameters; they need not be the actual probabilities of the states.

In each convention the compiled class is every K-state controller that the convention allows.

The best loss in each class was found exactly. The searches are exhaustive, with one exception: for three and four states with a step counter, the compiled optimum comes from the proof described below together with an exhaustive search of the explicit class.

What was found

The headline under the frozen rules is no advantage for the explicit organization. There is also no increase in optimal task loss: the best explicit and the best compiled controllers have the same loss in every case compared.

  • With a free step counter, a short proof shows that the best K-state compiled controller can always be written as a calibrated belief register with monotone rounding. This holds for every K, horizon and parameter value of the model.
  • Without a step counter, exhaustive computation at two parameter points finds that the best K-state controller is a saturating counter: one residual moves it one state up, the other one state down, and it stops at the ends. A fixed-grid belief register produces the same table. This is a finding for K up to 4 at the stated parameter points and horizons, not a theorem.
  • The designs that are best without a step counter at the primary parameter point were then frozen and run under four change processes not used in their design. For each K the best explicit designs and the best compiled designs are the same two behaviours, the counter and its mirror image, so the two classes have the same task loss in all four. Transfer losses for the best step-counter designs are reported without a verdict.

Two elementary facts limit what any comparison of this kind can show. An explicit class is a subset of the compiled class at the same budget, so its best loss on the regime it was designed for cannot be lower. And the contract counts a transfer advantage only if it holds whichever best designs are compared; that requires the explicit class to have a strictly higher best loss on the design regime.

The effects that do appear in the tables belong to the number of states and to the free step counter, not to how the memory is organized.

What this does not show

The model has one binary actuator parameter. Its flip and noise laws are given, and the residual isolates the actuator from the target by construction, so attribution between body and world is not tested. The comparisons use at most four memory states and eight steps. Of the comparison results, only the step-counter equality is proved for every number of states and every horizon.

Only the task loss is scored; there is no confidence report or recovery measure. Controllers are deterministic. Code length, table size, operation count, energy and latency are not constrained, so nothing here says which organization is cheaper to write, learn, verify or maintain.

Without a step counter, the equality says that the best table is one a fixed-grid belief register produces. It does not say that the table was found by modelling.

The transfer test moves a frozen controller to a different change process for the same parameter. It is not transfer across actuator families and tasks. Because the best designs coincided, that test could not separate the two classes.

For three and four states with a step counter, the claim that no better rounding exists rests on one exhaustive search. Separate code re-evaluated the values and class membership of those optima but did not repeat the search.

“Self-model” here labels a machine property of a programmer-defined variable. The note makes no cognitive, biological or consciousness claim. It claims no novelty or priority: finite-memory tracking of a binary Markov source is an established subject, and the note may duplicate known results. It is AI-assisted work with internal checks, not external peer review.

Connection to B-Theory

The budget counts the states kept at every step, as in Finite-memory adaptive tracking and Adaptive bit acquisition and memory. The designer’s recursion comes from the first of these, and its published values were used to check the code. Finite Noisy Audits shows that a required report can change the memory needed; this note scores task loss only.

The note and its supporting files

Read the note online. The proofs are in Part I and the exact tables in Part II.

The frozen contract, its amendments, the standard-library Python code and the exact outputs are kept in the research working folder. They are not published with this page, so the computations cannot be reproduced from the site alone.

Related manuscript

Explicit and compiled actuator trackers under a fixed memory budget: An exact finite comparison under a contract frozen in advance

Working note v0.1, 16 November 2026. AI-assisted research draft with internal checks only; not external peer review. No novelty or priority is claimed.

Reading scope: One binary actuator parameter; task loss only; deterministic controllers. The finite comparisons use two to four persistent memory states and horizon 8. The equality proved for a free step counter (Theorem 7) holds for every number of states and every horizon of this model. Code length, table size, operation count and the cost of learning a controller are not constrained. No cognitive or consciousness claim and no novelty or priority claim.

See this manuscript's index entry
Phase: BecomeBecomeManuscript: Explicit and compiled actuator trackers under a fixed memory budget: An exact finite comparison under a contract frozen in advanceExplicit and compiled act…Project: Adaptive bit acquisition and memoryAdaptive bit acquisition…Project: Finite Noisy AuditsFinite Noisy AuditsProject: Finite-memory adaptive trackingFinite-memory adaptive tr…Project: Reflective Self-ModelsReflective Self-ModelsConcept: Self-model calibration under internal changeSelf-model calibration un…Fixed-Memory Comparison
  • Phases
  • Manuscripts
  • Projects
  • Concepts

Conceptual and navigational links, not proof dependencies.