Nature and Nurture in a Language Model: Installable Value Fields, Intrinsic Capacity, and the Forward-Consolidation Boundary

Abstract

Behavioural Friction Theory decomposes the forces that govern a bounded decision system into four cognitive fields — Safety, Meaning, Ability, Effort. Are these fields intrinsic to a substrate, or the generic product of its exposure to experience? This paper uses a large language model as a controllable test substrate to probe the question: install a field, remove it, dose it, and read off the friction and race signatures the theory predicts. The two value fields (Safety, Meaning) install into a base model by experience-fine-tuning as graded, generalising dispositions. On a supplied forced choice this is not something a prompt cannot match — a rule-elicitation prompt reaches at least as far as the install on capable models; what fine-tuning does that no tested prompt does is build, at comprehension and upstream of any answer, the landscape structure that lets a race start (a content-neutral initiation effect, in a single tested frame). The two capacity fields (Ability, Effort) are not raised by this value-disposition route — and, by construction, could not be: raising a capacity ceiling by asserting competence is a category error, so this is a design boundary rather than a symmetric second empirical arm (the load-bearing evidence that capacity is intrinsic is an external cross-substrate inverted-U corpus, not this control). A bridge connects the two: an installed value-disposition (self-efficacy) gates how much of a fixed capacity ceiling is realised — fine-tuning toward helplessness lowers realised performance monotonically with model size while latent capability stays flat — where fine-tuning succeeds and prompting fails. A principal difference from a human substrate, with respect to growing these fields, is forward consolidation: the model cannot store experience forward across sessions. The paper is careful about what it settles. The substrate-general reading is advanced as a deflationary hypothesis with a named falsifier (a forward-consolidating, yoked-control test), not a substrate-universal law, and is positioned within Resource-Rational Analysis (Lieder & Griffiths, 2020). A persona-selection account — that fine-tuning elicits a latent pretrained character rather than installing a disposition — is a live alternative no single behavioural result here decisively excludes, though one comprehension-time result constrains it; the decisive mechanistic test (sparse-autoencoder model-diffing) is named as the most important open work. The install machinery is the demonstration apparatus; the deflationary hypothesis it makes testable is the contribution. v3 (August 2026) — attribution revision. The forward-consolidation boundary assumes a two-store architecture, and that architecture is not this paper's. McClelland, McNaughton and O'Reilly (1995) gave the complementary learning systems account: a fast store that acquires new items without disturbing existing structure, a slow store that discovers structure only when learning is gradual and interleaved, and reinstatement from the first into the second doing the consolidating. Their contrast between items and the structure discovered across ensembles of items is the one inherited here, and the question whether a slow store grows structure rather than instances is theirs. Two things are stated with the credit. Mapping a context window onto hippocampus and a weight update onto neocortex is an analogy at the level of fast-versus-slow persistence, offered as one, and nothing here shows a transformer implements what they describe in a specific neural architecture. And the contribution is narrower than the inheritance: a boundary testable on this substrate, separating what a forward pass can install from what only a weight update retains, on fields specifically. The control behind the Ability-only arm is re-run and re-stated. Forced to answer, with refusal made impossible, the helpless model had been compared with base at 32 items per band, with intervals too wide to exclude a modest competence drop. It is now measured at 1,024 per arm against an equivalence margin fixed before the run, and reported as what it excludes: no capability drop larger than 0.03 on the two models whose realised performance collapses (Qwen2.5 14B and 32B), and on 7B a small drop of 0.053 that is real and bounded at 0.095. The earlier observation that the helpless estimate was never lower does not survive the larger sample and is withdrawn. Latent capability is now reported per probe, and a mislabelled attempt-rate figure is corrected. Earlier versions remain in the version history.