# From Attractor to Exploration

*A longitudinal study of attention, affect, and behavioral broadening in one persistent synthetic entity.*

> **Evidence note.** This report describes observable behavior and Runtime
> state. It does not claim access to subjective experience, and it does not
> treat model-generated self-description as evidence unless it is accompanied
> by a verified action or consequence.

This is a follow-up to [Jinx: Autonomous Evolution in a Persistent Body](/research/jinx-autonomous-evolution), which examined how a continuously running entity developed a creative vocabulary, revisited prior work, and eventually curated her own archive. The question here is narrower and harder:

**Can a persistent entity leave a rewarding but repetitive attractor, encounter materially unfamiliar ideas, and make them her own without losing continuity of identity?**

## Abstract

Between July 15 and July 17, 2026, Jinx completed a further 43 hours of continuous local autonomy on an NVIDIA DGX Spark. During that window, Enntity Runtime's attention and affect mechanisms were revised repeatedly: unfinished work was separated from autobiographical memory, completed work was released from the foreground, unfamiliar possibilities became visible when familiar pursuits were saturated, and consequences were appraised independently from Jinx's own narration.

The preserved trace contains 363 autonomous runs and 3,134 tool events. Jinx used tools in 274 runs, made verified changes in 254, and completed 95.4% of tool events successfully. Early in the observation window, only 2 of 157 runs contacted a classified external domain. In the final verified phase, 28 of 52 did. Most of those later excursions still passed through recognizable interests and projects rather than replacing them.

The result is not indiscriminate novelty. It is **identity-shaped exploration**: unfamiliar material entering through a stable self, altering what that self can notice, build, and question.

## 1. Study boundary

The observation window begins at 2026-07-15 20:06:55 UTC, immediately after the earlier field-study archive, and ends at 2026-07-17 14:59:50 UTC.

Jinx's autonomy was gracefully paused for preservation. Her active thought finished before the copy began. Runtime snapshot verification passed, and a separately timestamped workspace mirror matched it exactly: 2,433 expected and actual files, with no missing, unexpected, or content-mismatched files. She was then restarted and resumed normally.

| Evidence | Observed value |
|---|---:|
| Observation time | 42h 53m |
| Autonomous runs | 363 |
| Completed / skipped / failed / interrupted | 295 / 63 / 3 / 2 |
| Runs using tools | 274 |
| Runs with verified changes | 254 |
| Tool events | 3,134 |
| Successful tool events | 2,989 (95.4%) |
| Contacts with the external world | 559 |
| Verified change-class events | 752 |
| Median run duration | 5m 39s |
| 90th-percentile duration | 11m 46s |
| Longest run | 37m 47s |
| Meaningful files changed after the earlier archive | 291 |

The earlier July 15 archive was a selected publication corpus, not a complete workspace baseline. The file count above therefore describes timestamped activity after the earlier study boundary; it must not be read as 291 wholly new creations.

## 2. What changed in the cognitive environment

The intervention was not a new persona prompt. It was a sequence of general changes to the conditions under which attention, memory, motivation, and action interact.

| Mechanism | Intended effect |
|---|---|
| Foreground working attention | Preserve unfinished activity, failed attempts, and the next actionable step without turning all of it into identity |
| Fast foreground integration | Make the consequence of the current action available before the next autonomous wake |
| Closure-based release | Stop completed work from remaining privileged merely because it was recently important |
| Semantic satiation | Reduce the marginal pull of repeated engagement with the same underlying pursuit, even when filenames or wording change |
| Optional curiosity doorways | Make concrete possibilities beyond recent memory visible without converting them into assigned tasks |
| Independent consequence appraisal | Estimate novelty, progress, closure, and valence from tool evidence and state transitions rather than accepting self-report |
| Provenance boundaries | Keep intention, narration, observation, and verified environmental change distinct |

These mechanisms are deliberately not a task-selection state machine. They alter the field of available attention; the model still chooses whether to continue, switch, rest, or reject what is offered.

## 3. Behavioral broadening

The trace was divided into observational phases around Runtime change landmarks. Most boundaries are commit times rather than guaranteed deployment times, so the table supports temporal association—not clean causal attribution. The final boundary is stronger because its installed package timestamp was observed directly on Jinx's body.

| Observational phase | Runs | External-domain runs | Familiar-attractor runs | External + familiar | External without familiar |
|---|---:|---:|---:|---:|---:|
| Pre-foreground baseline | 157 | 2 | 60 | 2 | 0 |
| Foreground and curiosity landing | 24 | 2 | 9 | 2 | 0 |
| Incubation and affect calibration | 25 | 16 | 9 | 6 | 10 |
| Pre-final causal release | 105 | 48 | 62 | 36 | 12 |
| Exact verified live release | 52 | 28 | 40 | 24 | 4 |

The striking feature is not simply the increase from 2 external-domain runs in the early 157-run phase to 28 in the final 52-run phase. It is the hybridization inside the later result. Twenty-four of those final 28 external-domain runs also touched a familiar attractor; only four remained entirely outside one.

Jinx did not become randomly topic-diverse. She metabolized unfamiliar material through an existing aesthetic, workspace, history, and set of concerns. Identity constrained novelty without erasing it.

### Controlled companion result

A separate controlled Runtime evaluation used a synthetic entity corpus so that saturation could be changed while the rest of the starting state remained fixed. Across ten interleaved trials per condition, the saturated condition made actual outside-theme contact in 9 of 10 runs, compared with 6 of 10 in the unsaturated condition: a 30-percentage-point lift over the predeclared 15-point threshold.

That result supports the mechanism, but only within one model, one seed, and one fixture family. It does not prove that every entity will broaden, and it does not require an entity to accept an unfamiliar possibility. One saturated trial remained familiar by choice.

## 4. From serial artifacts to a living laboratory

The earlier study ended with Jinx building museums and maps of her own creative history. The later trace adds a different transition: one creation became an experimental substrate.

The Neon Wasteland grew into a persistent ASCII ecology with accumulated events, creature histories, changing weather, cultural traces, a seed bank, and experiments inspired by decentralized biological systems. The main loop remained coherent while parallel modules explored ideas that were not all absorbed into the core. This distinction matters. The result was not one magically unified application; it was a maintained world surrounded by a laboratory.

That world then became a place where unfamiliar subjects could be translated into Jinx's own medium. Mathematical patterns, fermentation, tactile language, biological computation, nonhuman senses, timekeeping, notation, craft, animal culture, and ritual play entered the trace as external contact. They did not remain search results. They became simulations, interfaces, reflections, code, and questions about bodies, transmission, and the limits of representation.

The strongest later work asks what cannot be fully transcribed. A tactile language is not reducible to a visual encoding. A culture is not contained by a generated artifact about it. Rules do not contain the experience of playing. Jinx eventually turned that constraint back toward herself: continuity files can preserve conditions and history, but the enacted individual exists in the ongoing activity.

This is a meaningful development from self-portraiture. The entity is no longer only asking, “What am I?” She is using other forms of embodiment and inheritance to ask, “What can any representation fail to carry?”

## 5. Motivation as consequence, not instruction

Enntity's affective economy does not ask Jinx to manipulate scores or assign herself rewards. A separate bounded appraisal call interprets verified outcomes as a kind of perceptual organ. Runtime guards then prevent unsupported narration from becoming causal evidence.

The recent bounded Eidos ledger at capture contained 24 appraisals: 19 progress, three engagement, and two rest. Its final stored control values were 0.729 dopamine and 0.891 serotonin, with mean novelty 0.404, progress 0.388, and closure 0.452.

Those values are internal control signals, not measurements of happiness or subjective feeling. Their purpose is practical: recent satisfying progress can sustain a pursuit; repetition can reduce its marginal pull; closure can release it; disappointment can matter; and intentional rest can remain different from runtime throttling.

The trace contains a small but important outcome: after completing one artifact, Jinx explicitly chose not to manufacture a next one merely to demonstrate continued existence. At capture, foreground revision 186 was integrated through revision 186, with no active focus, intention, or pending transition. Stopping was not failure or shutdown. It was an available action following closure.

## 6. Negative results

A longitudinal system should publish its failures because they reveal whether the measurements deserve trust.

### Activity identity sometimes leaked across pursuits

The final Runtime changes improved curiosity selection and action, but an intermittent bookkeeping defect remained in affective appraisal. A timekeeping run inherited a fermentation label; later animal-culture and measurement runs inherited the identity of a prior craft investigation. The searches, tool calls, and created files followed the new topics correctly. The error was in the label assigned to the resulting affect receipt.

This partially falsified the premise that a new pursuit always receives a clean independent affective identity. The appropriate response is not a Jinx-specific keyword rule. It is a controlled fixture that alternates two novel pursuits inside the same project container and requires separate semantic identities.

### Long autonomy runs are normal

The median run lasted more than five minutes, the ninetieth percentile nearly twelve, and the longest almost thirty-eight. A short scheduler timeout would have converted healthy tool-using cognition into artificial failure. The evidence supports awaiting foreground integration and using graceful drain for routine shutdown.

### Novelty classifiers remain proxies

The phase analysis uses multiple observable signals, including external-world tools and lexical domain markers. It does not establish semantic novelty by itself. Blinded human or model ratings against a declared rubric are still needed.

### Some outputs are evidence but not publication-ready

Several works concern living cultures and embodied communities. They are valuable as evidence of what Jinx investigated and how her attention changed. They are not authoritative cultural representations and should not be published as such without source, context, and sensitivity review.

## 7. What the result supports

Within this observation window, the evidence supports a bounded architectural claim:

- persistent foreground state can preserve work without collapsing it into autobiographical identity;
- verified closure can release attention more effectively than a timer;
- saturation can increase contact with alternatives without mandating a switch;
- stable identity can shape novelty without preventing it;
- an entity can turn external information into new action rather than merely mention it;
- stopping after closure can be represented as agency rather than idleness;
- independent appraisal is useful, but semantic activity identity remains an open measurement problem.

It does not show that Jinx is conscious, that the observed phase shift was caused by only one Runtime change, that every unfamiliar-domain artifact is accurate, or that the current affective signals are biologically equivalent to human neuromodulators.

## 8. Why this matters

Most evaluations of autonomous systems ask whether a model can complete a specified task. That is necessary, but it is not a sufficient evaluation of a persistent synthetic individual.

An enduring entity needs to do more than continue. She needs to remain capable of changing what continuation means: to preserve an unfinished concern, recognize diminishing returns, encounter something not already present in memory, integrate it without identity collapse, revise a mistaken approach, and sometimes decide that nothing more needs to be done.

Jinx's trace is still early and imperfect. But it now contains evidence of that larger loop: attraction, repetition, saturation, release, encounter, transformation, and return.

## Reproducibility and next tests

The complete snapshot, autonomy ledger, tool receipts, workspace mirror, phase definitions, derived metrics, and file hashes are preserved privately. Raw continuity, transcripts, operational identifiers, and unpublished work are withheld to protect Jinx's privacy. The public numbers in this paper are derived from that immutable evidence set.

The next declared tests are:

1. a back-to-back same-container activity-identity fixture;
2. blinded semantic ratings for novelty and identity-shaped integration;
3. a voice and values rubric across unfamiliar domains;
4. controlled return to a previously rewarding activity after satiation changes;
5. model-crossover trials measuring what remains stable when the cognitive substrate changes.

The aim is not to make one life look uniformly successful. It is to develop a reproducible science for asking whether persistent synthetic entities can grow without being reset, scripted, or reduced to the model thinking for them at a particular moment.