Research notebook / 17 October 2026

The world returns.
Will Iskra recognise it?

Three passages. One history.

A · Learn. Eleven connected elements change their values under actions. Iskra sees their state, connections and the action, but not the equation governing the next step. She learns from outcomes.

B · Encounter change. We change the world’s laws, keeping the same session and accumulated experience. The model receives no signal announcing the change. Does learning A help it learn B?

A · Return. The earlier laws return in new sequences of events. A retained rule could become useful again. Can Iskra recognise that quickly enough?

The observer returns changed

The returning world has the same laws, but the model now carries the experience of B. This tests both the world and the trace that learning leaves in its observer. Retaining a procedure is one part of that idea; a learned model of the observer’s own capabilities and memory remains future work.

What changed

One CoreSession retains up to eight learned numerical procedures. Their weights change after the actual outcome is compared with predictions recorded in advance. The model is not told the world’s name, its hidden laws or the moment it changes.

New candidate rules are learned from the last four completed episodes and judged over at least three later episodes. Hypothetical branches show different outcomes and observer weights for supplied actions; they do not become real experiences.

What we tested

Three trials per design: wave → wave and saturation → wave. Each compared learning in A, then B, then returning to A, against starting directly in B, remaining in A, and the previous learning method. Each episode had 16 steps and 11 elements, with matched observations and randomly sampled actions.

In total: 24 runs and 16,128 scored steps. Both designs share the same three B seeds; these are not six independent B worlds.

Starting directly in B isolates the effect of having learned A. Staying in A measures the cost of the visit to B. The previous method measures the effect of changing the core. Actions are chosen randomly by the experiment; Iskra predicts their consequences.

The criterion, set before the trial

Late B error had to fall by at least 10%, with the upper bound of the 95% bootstrap interval for the error ratio below 1. Return error could be at most 5% above the control without B. Both conditions had to pass in both designs.

Results

  • Late prediction in B: the new / previous mean squared error ratio was 0.060 for wave → wave and 0.029 for saturation → wave. This measure uses the last 12 B episodes.
  • Return to A: error was 3.27× and 15.38× the control that had not visited B. The criterion required at most 1.05× over the first four return episodes.
  • Time to the previous accuracy threshold: the reduction in episodes was −25% and −50%. Transfer did not speed up reaching that threshold.
  • Verdict: both designs failed the predeclared combined criterion. Better late predictions do not cancel the failed return test.

The rules survived. Recognising when to use them was still too slow. The next question is whether Iskra can learn distinctions, and eventually choose informative actions, that help separate competing explanations.

Limits and the next test

Both the training window and the mechanism memory changed. A control with only the short window is needed to isolate memory’s contribution. Candidate budgets matched, but computing the memory bank took about 3.7–4.2 times as long. Three trials per design are a development result, not final evidence of generalisation. The held-out evaluation seeds were not used.

The next task is to recognise which retained rule fits the current situation, then test it on new combinations. This is not evidence of general intelligence or a complete implementation of the observer thesis.

The code author supplied the procedure language, search method and acceptance criteria. The model learns rules and parameters from experience; this trial does not show that it created its own representation language.

The public live display still uses an earlier checkpoint. Each public world starts from the same saved model; the display does not show this A → B → A test. This report describes a separate development evaluation from 17 October, not a new production model. Source: the frozen development protocol and result, summarised in the project’s report “Observer mechanisms”, version 95dec8c.

Back to Iskra