← Research map
Finite-state pilotBecome

Reflective Self-Models

When does tracking a changing part of a system's own machinery help, and does an explicit self-model add anything?

Read Online Or start with the explanation below.

Watch: when does a self-model help?

Follow a command through a changing actuator, see a Bayesian update, and compare the five policies in R4.

10:34 · English captions · Brian HQ narration

Watch larger, with chapters and transcript

A changing machine and a changing estimate

Imagine a robot sends a command to turn left, but its wheel turns right. Was that a random error, or have its controls reversed? To act effectively, its controller needs a useful estimate of what its commands will do.

R4 studies a small version of this problem: three successive actions, three noisy preliminary checks, and one possible actuator reversal. It is a finite mathematical model, with inherited benchmark results, rather than an experiment with a physical robot.

The controller’s estimate concerns its own actuator: are the controls normal or reversed? That is the paper’s minimal functional self-model. The question is whether explicitly organizing and naming such an estimate improves what a system can do.

Two comparisons answer different questions

The explicit self-model, compiled version, and external observer run the same probability update and decision code. Given the same sequence of events, their actions and probability reports match on every admitted path. A separate compiled lookup table was not built. Changing the label or attributed location adds no measured benefit in this comparison.

Two further rules use the available evidence differently. One keeps its starting estimate. The other allows for the reversal chance but waits for two consecutive contradictory observations before swapping its weights, instead of applying a Bayesian likelihood update after each observation.

In the supplied benchmark, the Bayesian trio averages about 23 failed actions per hundred, compared with 27 for the two-contradiction rule and 29 for the keep-starting-estimate rule. These are expected model rates. The gain supports this use of evidence relative to these two restrictions; it does not establish that an explicit self-model is generally superior to other architectures or change-detection methods.

All five share information access and the same three initial checks. Equal memory, calculation cost, energy, latency, and engineering effort were not established across all five. A lower probability-prediction score also does not, by itself, establish better calibration or a human feeling of confidence.

Being, Becoming, and the awareness question

The actuator has a condition now: normal or reversed. That condition can change. Observations let a controller revise its description and use it to act. This gives a small operational example of keeping a useful account of Being through Becoming.

Actual change and a changing estimate are different events. A reversal may happen before the controller detects it; new evidence may also correct an estimate while the actuator stays unchanged.

The study examines tracking a property of one’s own machinery. It does not establish a sense of identity, introspection, or subjective experience. The broader connection to The MIND remains a research question.

Read the study and consider the next test

Read R4 online with rendered mathematics, download the PDF. The accompanying video above introduces the mechanism and comparison; its larger watch page includes chapters and a transcript.

A useful next comparison would fix the memory budget and ask whether an explicitly organized actuator model helps across new actuator changes. The alternatives would need the same information and comparable resources, declared before the outcomes are known. That comparison has since been carried out in one finite model: the working note Fixed-Memory Comparison finds no advantage for the explicit organization, and no increase in optimal task loss, with two to four memory states. Transfer across actuator families and tasks, which the manuscript also sketches, was not tested.

Technical addendum: when checking takes time

When Checking Takes Time gives an exact two-probe example in which delayed diagnostic evidence makes a first check harmful in isolation yet useful within a contingent plan. Every executed probe exposes the mechanism to another transition. The result depends on that changed timing contract and does not imply that merely offering optional information worsens optimal performance. It is a modest technical addendum with internal AI review, not a claim of architectural advantage or verified historical priority.

The technical addendum Finite Noisy Audits with Learned Baselines extends this matched-controller caution to learned baselines, exact noisy-audit risks and output-dependent memory. Its supplied structural and noise assumptions remain explicit.

The Sensor Decisions and Paid Calibration addendum separates learning a shared measurement law from decisions about a fresh target. Its finite latent-sensor model gives an exact interval frontier and fixed-sample certification bounds, without claiming architectural superiority or empirical validation.

The short diagnosis-and-repair family guide distinguishes the fixed-target foundations report, the changing-state repair manuscript and R4’s matched-tracker pilot. The reports use different observation, transition and recovery contracts; the guide does not merge their guarantees.

Related manuscript

When Does an Explicit Self-Model Help? Calibration and Control under Internal Change in a Finite-State Pilot

Analytical manuscript based on an inherited finite benchmark, finalized October 14, 2026. Internal AI review is not external peer review; no new simulation or human data is claimed.

Reading scope: Full manuscript. Results are inherited source evidence; the reading copy does not independently rerun the benchmark or establish subjective awareness.

Site reading copy 1 from Final R4 manuscript, October 14, 2026; site reading formats prepared October 16, 2026 with typesetting and navigation additions.

Watch video

Technical material

Exact source package

See this manuscript's index entry
Phase: BecomeBecomeManuscript: When Does an Explicit Self-Model Help? Calibration and Control under Internal Change in a Finite-State PilotWhen Does an Explicit Sel…Video: When does a self-model help?When does a self-model he…Project: Causal Self-DiagnosisCausal Self-DiagnosisProject: Diagnosing While RepairingDiagnosing While RepairingProject: Finite Noisy AuditsFinite Noisy AuditsProject: Fixed-Memory ComparisonFixed-Memory ComparisonProject: Paid audits and repeated revisionPaid audits and repeated…Project: Recoverable Memory Under ReplacementRecoverable Memory Under…Project: Risk-constrained solver selectionRisk-constrained solver s…Project: Sensor Decisions and Paid CalibrationSensor Decisions and Paid…Project: Shared Clocks and Historical TargetsShared Clocks and Histori…Project: When Checking Takes TimeWhen Checking Takes TimeConcept: Being and BecomingBeing and BecomingConcept: Finite-state systems with known dynamicsFinite-state systems with…Concept: Noisy readouts and error–cost trade-offsNoisy readouts and error–…Concept: Self-model calibration under internal changeSelf-model calibration un…Concept: The MIND — proposed cognitive readingsThe MIND — proposed cogni…Reflective Self-Models
  • Phases
  • Manuscripts
  • Videos
  • Projects
  • Concepts

Conceptual and navigational links, not proof dependencies.