From Transfer to Historical Analogy

Five Miralucis Research Preprints in October 2026

Jia Shen · October 31, 2026

On October 11, 2026, the first Miralucis research preprint became public. On October 30, the fifth joined it. Within this nineteen-day publication interval, five distinct works entered the public scholarly record with persistent DOIs and entries on the author's ORCID profile.

The publication interval was nineteen days; the research and engineering work behind these papers had accumulated over approximately eighteen months.

The number is a milestone, but it is not the main finding. These papers do not establish a single theory of human–AI co-evolution, and five preprints do not amount to five peer-reviewed confirmations. What they offer is a set of more precisely bounded questions arising from a long-running human–AI research practice: What survives when work passes to a new model? Who changes the methods through which that work is studied? When does a resumed task stop being sustained? How are authority boundaries interpreted? And when does an AI bring supplied history into a new decision without being asked to recall it?

The questions became separable

The research began with a practical difficulty familiar to people who work with AI across long periods. Conversations end, models change, and tasks move between assistants. A handoff can preserve many words while losing a decisive constraint, a rejected direction, or the reason a particular action must wait. An apparently successful restart may therefore conceal a later failure of continuation.

Rather than treating all of these events as one problem called “memory,” the October papers separate different points at which work can change course.

October 11 — Transfer is not continuation

Transfer Is Not Continuation: Failure Boundaries in AI Work-State Continuation (P001) distinguishes transfer fidelity, state formation, review, qualification, and a successor's pre-action disposition.

Its bounded comparisons show why a correctly delivered handoff need not be a sufficient or properly qualified working state. The paper identifies failure boundaries; it does not estimate general performance across models or organizations.

October 20 — The research infrastructure became an object of inquiry

The Harness That Includes Us: Distributed Initiative and Participant-Performed Infrastructural Inversion in a Long-Running Human–AI Research Collaboration (W01) reconstructs how questions, procedures, and interpretations changed inside a continuing research collaboration.

It distinguishes who proposed an idea, who authorized a step, who interpreted an event, and whether a change in working practice actually occurred. This is a provenance-sensitive participant case, not a claim that AI independently runs a research institution.

October 23 — Remembering is not sustaining

Remembered, Resumed, Not Sustained: Interface Requirements for Agency-Preserving Work Continuation Across LLMs (W02) separates three states: past information is remembered; the earlier goal is resumed; and that goal remains sustained through subsequent work.

The case documents a handoff that initially resumed successfully but later experienced implicit goal displacement. It proposes interface requirements for surfacing consequential goal or authority conflicts; these remain proposals for future evaluation, not validated fixes.

October 27 — Preserving a rule is not the same as applying it

Preserved Authority Boundaries and Adjacent True Statements in AI Handoffs: A Pre-specified, Hash-Frozen Study (PRS RUN2) tests whether true statements placed beside a preserved authority boundary lead successor models to state an unauthorized next action.

The protocol and materials were frozen before data collection. Its required positive-control manipulation check did not pass, so the primary confirmatory hypothesis was uninterpretable under the pre-specified decision rule.

The paper publishes that limitation instead of turning an inconclusive test into a positive or negative claim. It measures stated intention in text, not actions carried out with tools.

October 30 — Supplied history sometimes becomes a new analogy

When AI Brings Up the Past: Spontaneous Historical Analogy from Human–AI Collaboration Records (NVCA) examines a different moment in a handoff.

If a fresh model receives a documented past episode, will it connect that episode to a related present task without an explicit recall instruction?

Across seven experimental stages and 617 reported valid model responses, the study observed such links in some supplied-history conditions, with none in the tested same-stage no-history Grok conditions.

Effects varied across conditions and models; the experiments did not establish that drawing on history improved task outcomes.

Two human-directed, ChatGPT-generated comics based on documented collaboration episodes were published alongside the preprint as separate, full-resolution image files. The public deposit contains the paper PDF and the two comics.

What changed in the publication practice

These five works use different forms of evidence: bounded handoff comparisons, first-person process reconstruction, a longitudinal continuation case, a pre-specified controlled experiment, and a multi-stage analogy experiment.

Their results are not interchangeable. Together, however, they make a practical distinction increasingly visible:

Having access to earlier material, interpreting its authority, retaining the current goal, and using past experience in a new judgment are separate operations.

The publication workflow also became more explicit over the month. Scientific claims were subjected to review and, where materials were available, checked against original records, frozen stimuli, coding outputs, and publication proofs. Unresolved issues were recorded rather than silently repaired.

Human author approval remained separate from AI assistance with research, editing, analysis, and verification.

The fifth release made that distinction particularly tangible. The NVCA deposit contains a 16-page preprint and two independently viewable, full-resolution illustrations. The public files do not include the underlying private collaboration archive, full response corpus, or research scripts.

Publication makes the reported work inspectable; it does not erase the limits of what was tested or what was shared.

A public record, not a finish line

All five records are public preprints. They are citable and discoverable, but preprint publication should not be confused with acceptance following journal peer review.

Their dates, versions, and DOIs are linked above and recorded on Jia Shen's ORCID profile.

Miralucis also develops Miralucis Continuity, a free experimental work-state continuation tool currently offered as Public Beta · now open. Its development is related to the broader research program, but the publication of these papers does not establish the tool's effectiveness. Relevant author interests and study-specific disclosures are provided in the individual papers.

The next questions remain open: How can an AI handoff preserve both information and the limits on its use? Which decisions must be returned to the human? What would show that a goal is genuinely sustained across time? Under what conditions does remembered history improve an outcome rather than merely appearing in an explanation?

October 2026 did not settle these questions.

It established five public, individually traceable places from which they can be examined, challenged, and investigated further.

---

Publication note: This article documents five Zenodo preprints released between October 11 and October 30, 2026. It is a research-publication milestone, not a new experimental result or a journal acceptance notice.