Governance Irony: Why Information-Heavy Compliance Directives Mechanically Fail in Large Language Models

Abstract

Pour the complete rule-set into a system “to be covered” and adherence can collapse: with the relevant rule buried among competing directives, one language model’s compliance fell from 100% to 0% (another from 100% to 60–88%), while a short positive recipe held near 100% — length-, structure- and header-matched controls isolating the competing rules themselves as the cause. The paper argues that compliance is behaviour, not information, and that the dominant method — the monolithic rule-dump — is self-defeating. It uses a large language model as a measurable model organism: on the demands a rule-dump makes (attend to every clause, never tire, hold the whole manual at once) the machine is at least as favourably placed as a human reader, yet it still loses the buried rule, with human confounds (motivation, fatigue) absent — so the failure is structural, and capacity-graded (every model breaks once competing directives exceed its capacity). The fix is not less completeness but a different method: keep the complete record for audit, but install behaviour with a separate, short, retrieved, trained artefact (retrieval-augmented generation restores adherence to ~100% on a checkable task). The governance principle that follows: the completeness of a rule-record and the efficacy of the behaviour it installs are separable and routinely in tension; the target is not 100% adherence but optimal compliance, a calibrated point. Because the substrate makes a rule-set’s cognitive burden measurable, a language model can also serve as a regulatory-complexity test bed, pre-flighting a policy before it reaches people. The human-organisational extension is offered as a falsifiable hypothesis anchored in established human-factors literature (work-as-imagined / work-as-done; alert fatigue; the procedure paradox; information overload), not as a proof. Grounded in Behavioural Friction Theory (Paper 0) and its learning application, the Physics of Learning (Paper 16). What is new in v3. Deepens the §5 design rules with the paper’s own measurements. Installing durable compliance is not a single lever: fine-tuning is the right tool for the closed, deterministic residue of a rule-set (training on varied, paraphrased subsets buys conditional rule-following without over-blocking), while the open, conditional tail is better served by retrieving the relevant rule at the decision point — a whole-rule-set fine-tune over-applies on unseen cases. A new result prices the retrieval step itself, the triage: where a rule’s applicability trigger is checkable, a deterministic triage recovers the relevant rule at oracle level; and the in-context over-block is a failure of applying all the rules at once rather than of knowing which rule applies, so performing the selection as its own low-competition step restores adherence — the reduce-competition mechanism, one level up. For triggers that require judgement the decomposition remains net-positive at the cost of a small selection error. Adds design rules 5–7 (train the closed set and retrieve the open set; dose a must-all-hold rule-set in a couple of passes; co-install abstention in one bundle) and an agent-layer illustration. All new measurements are pilot-scale directional corroboration on two instruct-tuned model families, with magnitudes in the supplementary materials; the load-bearing argument (§2.5/§4.2/§4.5) is unchanged.