What can be learned without the answer?

Question

Can a retained predictive state support useful choices without being given the hidden regime or a teacher's action values?

Experimental setup

In an isolated partially observed task, a recurrent model was trained to predict observations. Hidden-regime labels and teacher action-value targets were excluded. The comparison asked whether retaining state improved later decisions.

The comparison

Compare the retained-history condition with a state-reset control, and retain a structured hidden Markov model as a simpler competing explanation.

Result

Reported mean utility · higher is better · task-specific scale
Retained recurrent state		0.805667
State reset		0.577487
Structured HMM		0.817264

History retention improved the reported utility relative to resetting state. The structured hidden Markov model remained stronger than the recurrent model, so neural complexity was not the winning explanation.

What remains open

This is a frozen-weight, isolated experimental comparison. It does not establish continuous learning in live Luna, experiential memory, or subjective awareness.

Next test

Test whether a retained estimate changes a real later decision in a continuing project, against a simple estimator with the same information and budget.

Evidence & access / NR-2026-001

Observation-only learning study, reviewed 16 September 2026. The publication reviewed internal source and research reports. It did not independently execute these full comparisons. The numerical results above are author-reported; they are not independently replicated findings.

Source family: Observation-only learning study, reviewed 16 September 2026. Internal, non-public source material; the public identifier above gives a stable reference for enquiries.
Public methods: This narrative and its machine-readable study summary (JSON).
Public executable artifact: Not released here. This page is not sufficient to reproduce the full experiment.
External research replication: Not established.

To request the protocol, implementation snapshot, run configuration or validation log, identify NR-2026-001 when contacting Neurasoft. Release availability requires review; no delivery is implied. This is not a live operational report about Luna.

All research records How we classify evidence