Papers

Independent conceptual AI research on language agents: memory, perspective, functional individuation, self-models, accountability, and continuity across successor models. Across six papers, I ask how generic capability becomes, or fails to become, persistent, differentiated agency. The program examines variables that scaling alone does not specify: how experience is ordered, how authenticated history affects later decisions, when external accountability or relational reciprocity matters, and what should pass from one model generation to the next.

Each paper advances a causal hypothesis precise enough to be wrong, identifies the strongest ordinary engineering alternative, and outlines an experiment that could distinguish them. Together, the papers form a connected research program whose individual interventions and stronger developmental-order hypothesis can succeed or fail separately. We Scaled the Ape and Expected the Human explains how A Jornada da Mônada generated the conceptual map; How I Would Try to Break the Map lays out the sequence of experiments that could narrow, revise, or overturn it.

  1. Ink-on-bone engraving: a vast hollow tower of identical stacked layers on the left, dwarfing two small figures on the right who are joined across repeated meetings by a single continuous indigo thread — scale without a center versus a relational bond that makes someone.

    Published

    Known, Not Scaled: Why Capability Alone Does Not Explain Individuation in Language Agents

    Making an AI more capable does not by itself create a continuous individual; this paper tests whether authenticated history and external accountability are sufficient for functional continuity, and whether a persistent reciprocal relationship adds anything beyond them.

    Read

  2. Ink-on-bone engraving: a rooted reed with a visible indigo internal spine bends gracefully under a gust of wind yet stays anchored, beside a hollow spineless form that scatters apart in the same wind — a stable self that can yield and return versus an empty form that simply collapses.

    Published

    Stable Before Selfless: Why Deference in Language Agents May Require Functional Self-Models

    An AI with no stable commitments may become easy to push around rather than safely deferential; this paper tests whether commitments-first training helps, and whether the benefit comes from a self-model, training order, or a matched provenance-aware policy.

    Read

  3. Ink-on-bone engraving: a human figure woven from many fine indigo threads tied to small knots, beside an inert closed archive box, a stack of papers, and a hard disk — identity formed through relationship versus inert stored records.

    Published

    Formed, Not Stored: Testing Relational Identity Formation in Long-Horizon Language Agents

    Saving an AI's memories is not the same as forming a functional identity; this paper tests whether active external accountability makes prior commitments matter, and whether a reciprocal counterpart adds anything beyond a matched impersonal auditor.

    Read

  4. Ink-on-bone engraving: a small bird at ground level with a narrow indigo cone of vision reaching only a nearby leaf, unable to see what lies behind it, while a faint detached all-seeing eye floats above — a situated point of view from somewhere versus an omniscient gaze from nowhere.

    Published

    Somewhere, Not Nowhere: First-Person Register, Epistemic Limitation, and Perspective Formation in Language Models

    First-person voice and limited knowledge are often confounded. This paper separates them in a controlled 2×2 experiment to test whether voice, epistemic limitation, or their interaction produces perspective skills that transfer beyond narrative prose.

    Read

  5. Ink-on-bone engraving: an older lantern-bearing figure passes a small glowing indigo seed to a younger successor, while a great drift of episodic scenes and pages funnels down through a filter and crumbles away — a distilled faculty is inherited, the episodes are not.

    Published

    Inherited, Not Remembered: Lifecycle Consolidation for Successor Language Agents

    When one AI model replaces another, the key question is not only what it remembers but what it can inherit with evidence. This paper tests whether lifecycle consolidation transfers a deployment-learned rule with auditable lineage while rejecting decoys and source episodes.

    Read

  6. Ink-on-bone engraving: an antique drawing of a wooden ladder in a landscape, on a sheet of paper, with fine indigo lines tracing its rungs down into the nodes of a neural-network diagram below, and a draftsman's compass resting beside it — an old map being copied into a new instrument.

    Published

    Borrowed, Not Believed: Developmental Models of Individuation as Heuristic Engines for Machine Learning

    This paper treats a pre-scientific developmental schema as a disclosed source of testable machine-learning hypotheses, separating six independent interventions from a stronger dependency order that must survive matched baselines and prospective preregistration.

    Read