Epistemic Compression: The Case for Deliberate Ignorance in High-Stakes AI

Steffen Lukas

cs.LGMar 26, 2026

Epistemic Compression: The Case for Deliberate Ignorance in High-Stakes AI

Authors: Steffen Lukas

Organizations: Charité – Universitätsmedizin Berlin, Germany

Abstract

Foundation models excel in stable environments, yet often fail where reliability matters most: medicine, finance, and policy. This Fidelity Paradox is not just a data problem; it is structural. In domains where rules change over time, extra model capacity amplifies noise rather than capturing signal. We introduce Epistemic Compression: the principle that robustness emerges from matching model complexity to the shelf life of the data, not from scaling parameters. Unlike classical regularization, which penalizes weights post hoc, Epistemic Compression enforces parsimony through architecture: the model structure itself is designed to reduce overfitting by making it architecturally costly to represent variance that exceeds the evidence in the data. We operationalize this with a Regime Index that separates Shifting Regime (unstable, data-poor; simplicity wins) from Stable Regime (invariant, data-rich; complexity viable). In an exploratory synthesis of 15 high-stakes domains, this index was concordant with the empirically superior modeling strategy in 86.7% of cases (13/15). High-stakes AI demands a shift from scaling for its own sake to principled parsimony.

Figures & tables

Figure 1: The geometry of generalization by regime. a, In the Shifting Regime (low signal stability), OOD error (red curve) rises monotonically after an optimal compression point ( $D^{*}$ ), creating a “Fragility Zone” (red shading) where excess capacity fits spurious correlations. The optimal strategy is “staying in the left valley” (green shading). b, In the Stable Regime (high signal stability), the “double descent” phenomenon occurs 15 , where massive overparameterization (blue shading) reduces error by capturing fine-grained invariant structure. Note that double descent applies to in-distribution performance on a fixed data distribution; its benefits do not extend to the distribution-shifted setting that characterizes the Shifting Regime.

Figure 2: Empirical simulation of the robustness–capacity tradeoff. a, Robust accuracy (minimum of in-distribution and OOD accuracy) versus model capacity (number of accessible features). In Shifting Regime environments (red line, $\rho=0.25$ ), increasing capacity—defined here as the inclusion of spurious covariates—degrades robustness ( $\downarrow$ ) as the model exploits non-stationary signals. In the Stable Regime (green line, $\rho=0.8$ ), the invariant signal dominates, rendering capacity increases benign. b, Regime migration: at fixed high capacity, robustness improves sigmoidally with signal stability ( $\rho$ ), illustrating the phase transition between regimes. The inflection point suggests a critical threshold where the dominant signal transitions from noise to invariant structure. Error bars denote s.d. over $n=20$ repetitions.

Figure 3: The regime phase diagram. A diagnostic map for high-stakes model selection. The boundary curve (black line) is a conceptual frontier delineating the minimum data richness ( $\log_{10}N/D_{\text{eff}}$ ) required for robust generalization at a given level of signal stability (derived in Supplementary Note 5). The shaded region (top-left) indicates the “Horizon-Data Constraint,” where non-stationarity prevents the accumulation of sufficient data to train high-capacity models—data half-life creates a physical ceiling on achievable sample size. Domain positions represent illustrative schematic placements drawn from and beyond the 15-domain formal synthesis; all positions are approximate and not authoritative coordinates. Domains are grouped by regime: red circles (Shifting), blue squares (Stable), orange triangles (Borderline/mixed). For the complete formal scoring of the 15 synthesis domains, see Supplementary Table 1. High-stakes societal applications (red circles) are shown clustered in the Shifting Regime based on this benchmark evidence. Foundation models trained on massive data (blue squares) occupy the Stable Regime but risk the “Fidelity Trap” (red arrow) when applied to non-stationary Shifting Regime tasks.

Regime	Examples	Mean RI	Winning Strategy	Concordance
Shifting Regime	Medicine, Finance, Policy	3.0	Compression (Simple Models)	80% (8/10)
Stable Regime	Physics, Games, Vision	0.2	Complexity (Deep Learning)	100% (5/5)

Table 1: Summary of the 15-Domain Exploratory Synthesis. Concordance denotes the proportion of domains where the Regime Index was concordant with the empirically superior model class. Statistical Analysis : Overall concordance is 86.7% (13/15; 95% CI [62.1%–96.3%], Wilson score interval; two-sided Binomial Test,

H_{0}:p=0.5

p<0.01

). This synthesis is illustrative and hypothesis-generating, not a formal meta-analysis; the 15 domains were selected to span the regime spectrum rather than to constitute a representative sample of all AI applications. Full domain scoring and references are provided in Supplementary Information.

Figure 4: Geometric conceptualization of epistemic compression. a, The Shifting Regime data characterized by low-dimensional manifolds entangled with high-variance noise. b, Black-box behavior (schematic): likelihood maximization unconstrains the feature space, fitting specific noise instances to separate classes, resulting in high-energy, brittle representations. c, White-box behavior (compression-based architecture): optimization of rate reduction ( $\Delta R$ ) acts as a geometric sieve, collapsing noise onto the underlying low-dimensional manifold and orthogonalizing class subspaces, thereby recovering the invariant causal structure.

Indicator	Shifting Regime Criteria (Score +1)	Stable Regime Criteria (Score 0)
1. Temporal Stability	Unstable : Relationships degrade ( $\Delta$ AUROC > 0.05) over 2–5 years. (e.g., Clinical practice, Finance)	Stable : Mechanisms are invariant over decades. (e.g., Physics, Protein folding)
2. Context Invariance	Low : Models fail to transfer across institutions without retraining.	High : Data generation is standardized globally. (e.g., Mass spec, Go rules)
3. Data-to-Complexity	Poverty : $N/D_{\text{eff}}<100$ . (Transition Zone: 100–1000).	Abundance : $N/D_{\text{eff}}>1000$ . Data saturates signal complexity.
4. Ground Truth	Subjective/Proxy : Labels are noisy proxies (e.g., “readmission”) or low-agreement ( $\kappa<0.8$ ).	Objective : Labels are unambiguous (e.g., “Win/Loss”, “Crystal structure”).
5. Causal Priors	Weak : Causal graph is unknown, contested, or unencodable.	Strong : Invariant physical laws can be encoded as constraints.

Table 6

Domain	RI	Simple Model (OOD)	Complex Model (OOD)	$\Delta$	Winner	Criterion	Conc.
1. ICU Mortality (MIMIC-IV v3.1)	4	LR: 0.716	MLP: 0.740	+0.024	Simple	Robustness ( $\Delta$ stability)	+
2. 30-Day Readmission	3	LACE: 0.68	DNN: 0.65	+0.03	Simple	Robustness ( $\Delta$ stability)	+
3. Sepsis Prediction	3	qSOFA: 0.74	Epic: 0.63	+0.11	Simple	Robustness ( $\Delta$ stability)	+
4. Credit Default (LendingClub)	3	LR: 0.674†	FT-Transf.: 0.671†	+0.059‡	Simple	Robustness ( $\Delta$ stability)	+
5. Recidivism (COMPAS)	3	2-var: 0.65	COMPAS: 0.65	0.00	Simple	Robustness ( $\Delta$ stability)	+
6. Income Prediction (Adult)	3	LR: 0.80	MLP: 0.88	-0.08	Complex	OOD performance	-

Table 7

Figure S1: The attention-to-noise mechanism.

Domain	$\tau_{1/2}$ (years)	Primary Drivers
Fraud Detection	<1	Adversarial adaptation
Credit Scoring	~2	Economic cycles, regulatory changes
ICU Mortality	~3	ICD transitions, protocol updates, EMR changes
Protein Structure	>100	Physics is invariant
Board Games	$\infty$	Rules never change

Table 9

Domain	Proxy for $D_{\text{eff}}$	Rationale
ICU Severity	APACHE-II variables (12)	Clinical consensus on mortality predictors
Credit Risk	Basel III factors (5-8)	Regulatory-validated risk drivers
Sepsis	qSOFA components (3)	Physiological deterioration markers
Drug Dosing	PK compartments (2-4)	Mass-action kinetics constraints

Table 11

Method	Estimate	Notes
Domain Prior	$D_{\text{eff}}\leq 12$	APACHE-II score components
PCA (MIMIC-IV)	$D_{\text{eff}}\approx 8$	95% variance threshold
Learning Curve	$D_{\text{eff}}\approx 10$	Plateau at $N\approx 1000$
Consensus	$D_{\text{eff}}\approx 10$	Conservative estimate

Table 12

Aspect	Classical Regularization	Epistemic Compression
Mechanism	Post-hoc penalty on weights	Architectural constraint on representable functions
Flexibility	Model can represent complex functions; regularization discourages it	Model cannot represent certain functions; architecture forbids it
Analogy	A leash (can be pulled)	A fence (cannot be crossed)
Failure Mode	Can be overridden by strong gradients	Architecturally resistant by design; empirical validation in new domains required
Interpretability	Weights remain dense, hard to interpret	Sparse/structured, often interpretable
Domain Knowledge	Not encoded	Explicitly encoded in architecture

Table 13

Distribution Shift Robustness

Distribution shift robustness studies how to keep machine-learning predictions reliable when deployment data differ from training data.

See all 280 papers

Mar 8, 2026

From Garbage to Gold: A Data-Architectural Theory of Predictive Robustness

Tabular machine learning presents a paradox: modern models achieve state-of-the-art performance using high-dimensional (high-D), collinear, error-prone data, defying the "Garbage In, Garbage Out" mantra. To help resolve this, we synthesize principles from Information Theory, Latent Factor Models, and Psychometrics, clarifying that predictive robustness arises not solely from data cleanliness, but from the synergy between data architecture and model capacity. Partitioning predictor-space "noise" into "Predictor Error" and "Structural Uncertainty" (informational deficits from stochastic generative mappings), we prove that leveraging high-D sets of error-prone predictors asymptotically overcomes both types of noise, whereas cleaning a low-D set is fundamentally bounded by Structural Uncertainty. We demonstrate why "Informative Collinearity" (dependencies from shared latent causes) enhances reliability and convergence efficiency, and explain why increased dimensionality reduces the latent inference burden, enabling feasibility with finite samples. To address practical constraints, we propose "Proactive Data-Centric AI" to identify predictors that enable robustness efficiently. We also derive boundaries for Systematic Error Regimes and show why models that absorb "rogue" dependencies can mitigate assumption violations. Linking latent architecture to Benign Overfitting, we offer a first step towards a unified view of robustness to Outcome Error and predictor-space noise, while also delineating when traditional DCAI's focus on label cleaning remains powerful. By redefining data quality from item-level perfection to portfolio-level architecture, we provide a theoretical rationale for "Local Factories" -- learning from live, uncurated enterprise "data swamps" -- supporting a deployment paradigm shift from "Model Transfer" to "Methodology Transfer'' to overcome static generalizability limitations.

Terrence J. Lee-St. John, Jordan L. Lawson, Bartlomiej Piechowski-Jozwiak

Independent Researcher & Institute for Healthier Living Abu Dhabi, Abu Dhabi, UAE · Independent Researcher & Brown University, Center for Computation and Visualization, Providence, RI, USA

Statistical Learning Theory

Ask PDF

Apr 16, 2026

The Missing Knowledge Layer in AI: A Framework for Stable Human-AI Reasoning

Large language models are increasingly integrated into decision-making in areas such as healthcare, law, finance, engineering, and government. Yet they share a critical limitation: they produce fluent outputs even when their internal reasoning has drifted. A confident answer can conceal uncertainty, speculation, or inconsistency, and small changes in phrasing can lead to different conclusions. This makes LLMs useful assistants but unreliable partners in high-stakes contexts. Humans exhibit a similar weakness, often mistaking fluency for reliability. When a model responds smoothly, users tend to trust it, even when both model and user are drifting together. This paper is the first in a five-paper research series on stabilising human-AI reasoning. The series proposes a two-layer approach: Parts II-IV introduce human-side mechanisms such as uncertainty cues, conflict surfacing, and auditable reasoning traces, while Part V develops a model-side Epistemic Control Loop (ECL) that detects instability and modulates generation accordingly. Together, these layers form a missing operational substrate for governance by increasing signal-to-noise at the point of use. Stabilising interaction makes uncertainty and drift visible before enforcement is applied, enabling more precise capability governance. This aligns with emerging compliance expectations, including the EU AI Act and ISO/IEC 42001, by making reasoning processes traceable under real conditions of use. The central claim is that fluency is not reliability. Without structures that stabilise both human and model reasoning, AI cannot be trusted or governed where it matters most.

Rikard Rosenbacke, Carl Rosenbacke, Victor Rosenbacke +1

Faculty of Medicine, Lund University, Sweden · Department of Economics, Lund University School of Economics and Management, Sweden · Department of Health Services Research and Policy, London School of Hygiene & Tropical Medicine, UK

LLM Uncertainty Estimation AI Governance Human-AI Decision Making Human Oversight of AI

Ask PDF

Apr 9, 2026

Less Data Approximates More: Earning Faithful Confidence in High-Stakes Domains

Large language models are increasingly deployed in high-stakes domains, where confident yet incorrect inferences may cause severe real-world harm, bringing the long-overlooked issue of confidence faithfulness to the forefront. A promising solution jointly optimizes unsupervised Reinforcement Learning from Internal Feedback (RLIF) with reasoning-trace-guided Reasoning Distillation (RD), yet it faces three persistent challenges, namely the scarcity of high-quality training corpora, factually unwarranted overconfidence, and erroneous updates amplified by indiscriminate fusion. Inspired by how human confidence accumulates from uncertainty to certainty, we propose Progressive Reasoning Gain (PRG) to measure whether reasoning steps progressively strengthen confidence in the final answer. Building on PRG, we introduce HyTuning, a hybrid post-training framework that adaptively reweights RD and RLIF, using scarce supervised reasoning traces as a stable anchor while exploiting abundant unlabeled queries for scalability. Experiments on several domain-specific and general benchmarks demonstrate that HyTuning improves accuracy while achieving confidence faithfulness under limited supervision, supporting a practical ``Less Data Approximates More'' effect. Our code will be released upon acceptance.

Haokai Ma, Lee Yan Zhen, Gang Yang +4

National University of Singapore, Singapore · Singapore Management University, Singapore

RL for Language Models Reinforcement Learning Reasoning Distillation Confidence Estimation in Language Models LLM Fine-Tuning Language Model Calibration

Ask PDF

Abstract

Distribution Shift Robustness

Explore similar work

From Garbage to Gold: A Data-Architectural Theory of Predictive Robustness

The Missing Knowledge Layer in AI: A Framework for Stable Human-AI Reasoning

Less Data Approximates More: Earning Faithful Confidence in High-Stakes Domains