Agentic Stemmatics: Collation, Provenance, and Conservation for an Emerging Machine Textual Tradition

Abstract

Large language models emit testimony: fluent claims whose warrant lies outside the act of generation. What science fits such output? We argue it is the one philology built for corrupt textual transmission: stemmatics. Not, however, as an identity claim. An earlier draft of this paper made that claim, and four separately commissioned reviewers rightly returned it for major revision. What survives is narrower: a claim about where the model ecology stands now, and where it is heading, and a program of work rather than a finished result. First, the analogy between model lineages and manuscript witnesses was premature for the 2023–24 model ecology: on the evidence measured so far, models share errors mostly because they converge on the same attractors, not because they inherit them. Second, the analogy is becoming exact. Model output now re-enters training corpora through measurable channels: direct distillation, with reported watermark inheritance; ambient corpus contamination; deliberate synthetic data; and human-mediated diffusion. The ecology is growing a genuine descent tradition. Third, the genealogical signal in models is carried not by shared errors, philology's classical conjunctive errors, but by shared idiosyncrasies, for which the attribution and distillation-detection literatures already report candidate instruments. From that corrected foundation we derive three things. A conditional aggregation result: majority voting over model reviewers is unsafe wherever independence is unmeasured or failing, and the correlated-errors literature indicates that this is the present norm, and worsening. An executable experiment: reconstruct the stemma of the model ecosystem from output idiosyncrasies, then validate it against known model phylogenies. And a conservation program: archived model vintages as sealed witnesses, provenance strata, and a degradation clause for multi-model review panels. We also report the failure record of our own instruments. It includes the reviewed failure of this paper's previous draft, whose disposition table is part of the method, and a second, dated failure from a sibling instrument of this practice, in which external verification settled a claim that internal agreement could not, and the instrument's own governance clause was exited visibly rather than laundered. The commitment values for the practice's first controlled comparison were frozen and are quoted from their record; that instrument was discontinued on critique before it ran, a decision we report rather than paper over. We offer a way of looking at model output, a rule for reading the evidence it leaves, and a plan for testing that rule. We do not offer proof that the method works; that is the next step, not this one.