Exposure-Driven Residual Disclosure

Expose yourself to the problem until the deltas of the system output stop averaging out and become impossible to ignore. Then, once each systematic delta is isolated, fix them in Greediest Pareto order: each step takes the fix with the highest marginal (submodular) contribution, the largest net-new reduction in systematic residual per unit of effort given the fixes already made, until only irreducible noise remains in the residuals.

The disclosure loop. A model predicts, each exposure yields a residual, and residuals pile up until bias clears noise. Disclosure then separates fit, random error and bias, whether or not the model updates its priors. Conceptual schematic, redrawn from the author’s sketch.

Definition

Exposure-driven residual disclosure (EDRD) is the process by which repeated contact with a domain makes a model’s systematic misfit visible.

Line: Substrate-agnostic, repetition-accumulated visibility gain of systematic over stochastic prediction residuals under sustained, outcome-coupled domain exposure.

Sentence form: A process, common to human minds and artificial predictive models, in which repeated exposure to a domain accumulates prediction residuals until their systematic structure becomes observable. It exposes where the model’s priors fit, where they err randomly (variance, which averages out), and where they err systematically (bias, which accumulates), whether across the whole domain or only within a recurring sub-context.

EDRD is independent of whether the model then updates. It differs from learning, which changes the model, and from evaluation, which measures it once.

Terms

Disclosure time

The number of exposures needed before a bias becomes visible grows with noise and shrinks with the square of the bias.

n^* \approx \frac{z^2 \sigma^2}{p\, b^2}

For Gaussian residuals this is the minimum evidence any detector needs: $n^*$ is inversely proportional to the KL divergence between biased and unbiased residuals, which is $b^2 / 2\sigma^2$ . Halving the bias quadruples the exposure needed; a bias active in 10% of cases takes ten times longer to surface.

Boundary conditions

Repetition discloses bias only under three conditions; outside them, exposure accumulates without revealing anything.

Feedback fidelity. The outcome must be pairable with the prediction that produced it. Delayed, sparse, or confounded feedback (wicked environments) gives repetition without disclosure.
Update competes with disclosure. A model that corrects each residual immediately absorbs the bias before it piles up. The residual stream then looks clean, and the bias shows only as drift in the model itself. Disclosure is strongest when the model is held fixed during exposure.
Disclosure is observer-relative. The model and an outside observer each have their own noise and threshold, so each has its own $n^*$ . Self-disclosure also requires the model to represent its own residuals, which many models, human and artificial, cannot do.

The mechanism has precedents in every field below, ranked from closest to most distant conceptual match. EDRD’s own contribution is the framing: exposure itself as the disclosure instrument, disclosure decoupled from update, and one category for minds and AI models.

References

Yu, A. J., & Dayan, P. (2005). Uncertainty, neuromodulation, and attention. Neuron, 46(4), 681–692. doi:10.1016/j.neuron.2005.04.026. ↩
Page, E. S. (1954). Continuous inspection schemes. Biometrika, 41(1–2), 100–115. doi:10.1093/biomet/41.1-2.100. ↩
Wald, A. (1945). Sequential tests of statistical hypotheses. The Annals of Mathematical Statistics, 16(2), 117–186. doi:10.1214/aoms/1177731118. ↩
Wei, K., & Körding, K. (2009). Relevance of error: what drives motor adaptation? Journal of Neurophysiology, 101(2), 655–664. doi:10.1152/jn.90545.2008. ↩

Model	Any system that predicts outcomes from priors: a human mind or an AI model, treated as one behavioral category
Exposure	One cycle of predicting, acting, and receiving an outcome in a domain, where the outcome arrives soon and clearly enough to be paired with the prediction that produced it
Residual	The gap between the observed outcome and the model’s prediction
Disclosure	The point at which accumulated residuals cross an observer’s detection threshold ( $n \ge n^$ ), making their structure observable. The observer can be the model itself (metacognition) or an external one; each has its own $\sigma$ and $z$ , hence its own $n^$

$b$	Size of the systematic error
$\sigma$	Noise in each residual
$p$	Fraction of exposures where the bias is active (1 when it is global)
$z$	The observer’s confidence threshold

Mathematics	Doob decomposition (martingale theory)	Any accumulated residual stream splits uniquely into a martingale (unpredictable error) plus a predictable drift (bias): the fit / random error / bias split, stated as a theorem
Neuroscience	Yu & Dayan (2005), expected vs unexpected uncertainty¹	The random error vs bias split, with distinct neuromodulators (acetylcholine, norepinephrine) signaling each
Statistics	Page’s CUSUM (1954); Wald’s SPRT²³	Their detection delay is $n^*$
Physics	Metrology: systematic vs random error	Only the random part shrinks as $1/\sqrt{n}$
Physics	Le Verrier and Uranus; Mercury’s perihelion	Persistent orbital residuals disclosed Neptune and, after the “Vulcan” non-update, general relativity
Neuroscience	Wei & Körding (2009), relevance of error⁴	The motor system judges whether an error is relevant before adapting: the “bias > noise?” check
Philosophy of science	Kuhn, anomaly accumulation⁵	EDRD at the scale of a scientific community
Mathematics	Chernoff–Stein lemma; Cramér large deviations	Error probability in telling biased from unbiased decays as exp(−n · KL): the information-theoretic floor behind n* ∝ 1/KL
Mathematics	Law of large numbers; central limit theorem	The formal reason noise averages out (mean of n residuals shrinks its noise as σ/√n) while a constant bias b survives intact
Computer science	Domino (Eyuboglu et al., ICLR 2022)⁶	Finds systematic errors on data slices: conditional bias (p < 1)
Expertise	Hogarth, kind vs wicked environments; Kahneman & Klein (2009)⁷⁸	Expertise forms only where feedback is regular: the feedback fidelity condition
Neuroscience	Burge, Ernst & Banks (2008)⁹	Adaptation rate set by a Kalman-style ratio of noise to drift: the update vs disclosure trade-off
Mathematics	Birkhoff ergodic theorem	Time averages along one trajectory converge to the space average: sustained exposure to one domain recovers its true statistics, so residual structure becomes visible from experience alone
Neuroscience	Rao & Ballard (1999); Feldman & Friston (2010)¹⁰¹¹	Prediction errors weighted by their precision
Organizations	Argyris, single vs double-loop learning¹²	“Update priors (or not)”
Computer science	Gradient boosting (Friedman 2001)¹³	Each new component is fit to the current residuals: “fix the largest delta, repeat”
Mathematics	Submodular maximization (Nemhauser, Wolsey & Fisher 1978)¹⁴	Picking the next fix by highest marginal contribution is within (1 − 1/e) ≈ 63% of optimal: the guarantee behind Greediest Pareto order

Definition

Terms

Disclosure time

Boundary conditions

Related work

References

Footnotes