Trajectory Identity: A Mathematical Framework for Enactive AI Self-Hood

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Version 0.17 (2026-10-02). The abstract, conclusion, §1.3, §7.2 and §8.3 are brought in line with the body: the within-agent evidence is stated with §6.4's figures, and the multi-agent discrimination criterion is stated as open (the §6.5 pilot is confounded by role and harness). No framework or empirical change. Version 0.16 correction (2026-08-14). Appendix A is corrected in this version. Through v0.15 the appendix stated that the linked open-source implementation (anima-mcp) used the five informationally-independent weights (0.18, 0.18, 0.30, 0.22, 0.12) for (Π, β, α, ρ, Δ) with η held out of the weighted similarity sum, per §3.6 and §4.1. It did not, and never had: the similarity function implemented the pre-v0.11 six-component sum with η inside it at weight 0.15, over weights (0.15, 0.15, 0.25, 0.20, 0.10). The claim entered in v0.13, when a corrected code listing embedded in the appendix was replaced by a repository pointer and the sentence describing that listing was carried across to code which had never been changed. §3.6 and §4.1 are unaffected and remain the paper's specification, and no empirical result depends on the linked code — §6.4 and §6.5 are computed by standalone analysis code that does not import anima_mcp. The implementation has since been corrected (anima-mcp commit 3ac59e6, 2026-08-14), and that change is inert for the deployment reported in §6.4: Lumen's genesis signature stores η as null, so the term could never fire against genesis and its lineage similarity is unchanged to twelve decimal places. Full note in Appendix A. Version 0.15 correction (2026-07-28). The multi-agent discrimination pilot reported in §6.5 is marked confounded in this version. Evidence that arrived after the v0.14 deposit — a client migration under which one agent's production trajectory score against its own genesis signature collapsed to 0.123 and stayed flat, and a cross-agent audit finding between-agent similarity (0.63) exceeding within-agent-across-era similarity (0.12) — indicates this family of signature is dominated by era and regime rather than by individual. The §6.5 figures stand as computed and are not retracted: that analysis uses a different, reduced instrument which omits two of the components later found defective. What transfers is a confound. The four agents compared in §6.5 differ systematically in harness and duty cycle, and the pilot's most informative features are cadence descriptors, so separation by role is a live alternative explanation for the reported accuracies, and the within-era temporal split cannot exclude it. The discrimination criterion of Definition 2.3 therefore returns to open, pending a within-agent/across-harness test and a between-agent/same-harness test. No security implication: the governance write gate is credential-based and never depended on this score. Full note at the end of §6.5.Current approaches to AI agent identity rely on static identifiers (UUIDs, session tokens) or accumulated memory stores. We propose an alternative grounded in enactive cognition and dynamical systems theory: identity as trajectory. Rather than asking "what ID does this agent have?", we ask "what patterns persist in this agent's behavior over time?" We present a mathematical framework for computing trajectory signatures from time-series data including homeostatic state, learned preferences, self-beliefs, and recovery dynamics. The trajectory signature Sigma captures the quasi-invariant characteristics that define an agent's identity — not where it is at any moment, but how it tends to behave, where it tends to rest, and how it returns from perturbation. The framework is aimed at (1) identity continuity across sessions without unbounded memory growth, (2) principled semantics for agent forking and merging, (3) anomaly detection as trajectory deviation, through an asymmetric two-tier threshold scheme intended to distinguish drift from hijacking, and (4) inter-agent recognition based on behavioral signatures rather than credentials. We ground the framework in the UNITARES governance architecture and the Anima embodied AI system. The evidence is preliminary and mixed. On one embodied agent (Lumen: 65 calendar days, 47 of them with at least 100 samples, ~226,029 state observations), per-dimension state means stay within a small band (between-window variance of window means below 0.015), and a recovery time constant was estimable in all 12 perturbation episodes, from a small and dispersed sample (Section 6.4). This is consistent with within-agent quasi-invariance within one era and harness. The framework's defining claim, that distinct agents carry distinguishable signatures, is not established. A four-agent pilot separated agents (Section 6.5), but a later audit of the production similarity measure found between-agent similarity (0.63) above within-agent similarity across eras (0.12): that measure is dominated by era and harness rather than by the individual. The audit used a different instrument from the pilot, so it does not refute the pilot directly, but the confound carries over: the pilot is confounded by role and harness, and the discrimination criterion remains open.