A Layered Artificial-Life Model of Functional Self-Modeling — What Fails to Correct a Learned Self-Estimate, and What Does / Un modelo de vida artificial por capas del automodelado funcional — qué no corrige una estimación aprendida sobre uno mismo, y qué sí

Abstract

Research compendium (layers 0 to 16, four arcs) of a minimal artificial-life model in which persistent organisms — with real death, no negative valence, and no aversive signal — acquire value, prediction, attention, autobiographical damage records, a self-model, and a learned scalar estimate of their own fragility that governs avoidance. Layers 10 to 16 were measured under pre-registered protocols (hypothesis, controls and power thresholds fixed before running), with null and base conditions in every experiment, and a published error ledger. Layers 0 to 9 predate the adoption of that rule and are not pre-registered. VERSION 1.2 ADDS AN ERRATA DOCUMENT WITH NINETEEN ENTRIES AND THE SCRIPTS THAT PRODUCED THEM. Readers should start there. Two are consequential. First, the sealed headline run of layer 10 does not regenerate from the deposited code: an engine repair made while building layer 11 altered the null and base conditions after those outputs were sealed, and no earlier engine version survives. The experiment was replicated with the repaired instrument and N fixed in advance (asymmetry +0.041 vs +0.005, t = 4.64, d = 1.04, 95% CI [0.021, 0.052], N = 40+40). Second, the model contains three lineages, and in the one where the avoidance rule is not written by the designer the asymmetry is absent (t = 0.81, 95% CI [-0.008, +0.020], N = 30+30) — the condition the layer-10 pre-registration named as refuting the claim. The asymmetry therefore requires the hand-written functional form, as it does throughout the prior literature, and the earlier claim that the phenomenon arises "without designer-given goals" is withdrawn. What the corpus does establish: four exogenous channels, each independently verified to write to the learned parameter before being tested for efficacy, none of which met a correction criterion fixed in advance; and forced unbiased re-sampling, which does lower the stuck rate (13.6 points, d = 6.29) and which does so even in the lineage where the asymmetry is absent. Their failures delimit a class: information carried about the organism from outside cannot correct a quantity only its own body can measure. The sampling asymmetry itself is not novel — it is the hot stove effect (Denrell & March, 2001; Denrell, 2020), demonstrated in humans (Rich & Gureckis, 2018) and simulated in active inference (Smith, Moutoussis & Bilek, 2021). This deposit does not claim it. N = independent seeded runs per condition; effect sizes are Cohen's d. In seeded simulation between-run noise is low, so d values are large by construction and are not comparable to effect sizes in human studies. The corpus contains all pre-registrations, results, simulation code (dependency-free Node.js, fully seeded), raw per-seed JSON outputs, syntheses, a statistical report, the errata document and the verification scripts, each of which prints a digit-for-digit gate against the sealed aggregates before reporting anything. All corpus documents are in Spanish. An English preprint is in preparation. This deposit provides a citable, timestamped record of the research programme. VERSION 1.3 ADDS THREE PRE-REGISTERED EXPERIMENTS RUN AFTER THE MAIN RESULTS, IN RESPONSE TO REVIEW, WITH THEIR REFUTATION CRITERIA FIXED BEFORE RUNNING. (1) A yoked control: each organism acts on the estimate of another living organism, which preserves between-individual variation and destroys only the organism-estimate pairing. The asymmetry vanishes (signal minus yoked = +0.0455, t = 6.86, d = 1.53, N = 40+40), so the effect requires the organism's own estimate and not merely an individual one; the lesion of version 1.2 could not separate those. (2) A sweep of the estimator's step floor, which had never been justified: the contrast separates at a true cumulative mean 1/n and at floors of 1/10, 1/30 and 1/60, so the effect does not depend on the forgetting rate the model was built with. (3) A test of whether forced re-sampling corrects the estimate or merely reduces stuckness. ITS REFUTATION CRITERION FIRED. The belief-trait correlation is 0.97925 under forced re-sampling against 0.97944 with the estimate intact (t = -0.15, 95% CI [-0.0027, +0.0024]): flat. Forced re-sampling opens the loop without correcting the estimate, and the claim stated in version 1.2 that the channel failures delimit a class of exogenous correction is scoped accordingly in the preprint, whose title changed as a result. The same run measures the central phenomenon in units of accuracy for the first time: the correlation is highest of all when the estimate is prevented from governing exposure (0.98731, t = 7.78 against intact), so letting the estimate decide where the organism goes makes the estimate worse. Also deposited: a failed attempt to demonstrate the sign change of the sampling sweep with reduced generative models, reported as a failure rather than adjusted until it agreed; the patched engines implementing the yoke and the configurable floor, together with the scripts that derive them from the deposited code by exact substitutions that abort if an anchor is not unique; and the inertia gates those engines pass, 960/960, 960/960 and 720/720 per-seed cells identical to the sealed replication under strict floating-point equality. Two secondary predictions of the yoked pre-registration failed and are declared as specification errors by the author, not as model outcomes. The original corpus is unmodified.