Precision Weighting¶

Version: v1 · 2026-08-24 · History
Prime #: 1077
Origin domain: Cognitive Science
Subdomain: bayesian brain and predictive processing → Cognitive Science

Core Idea¶

Precision weighting is the structural pattern by which evidence about the same latent quantity receives influence in proportion to estimated reliability — specifically precision, the inverse of variance — so lower-noise evidence contributes more to the resulting estimate or update. Its defining commitments are a shared target, per-source precision estimates, and an influence or integration rule whose weights scale with those precisions. In the Gaussian case the posterior mean is a precision-weighted average of prior and likelihood means; in Kalman filtering the gain is a precision ratio; in parallel cue integration the cues combine by inverse variance. Precision estimates may be fixed for one calculation or may vary with context. Attention, instrumentation, preprocessing, or learned routing can actively change them, but active modulation is an extension of the core rule rather than a condition every instance must satisfy.

This skeleton recurs across substrates with unusual exactness. In Bayesian inference, Gaussian posterior beliefs combine prior and likelihood by precision.^[1] In Kalman filters and sensor fusion, the gain deciding how much weight a new measurement receives is precision-driven. In neural predictive processing the brain is hypothesized to weight prediction errors by context-estimated precision, with attention as one modulation mechanism.^[2] In weighted wisdom-of-crowds, contributions can scale with demonstrated accuracy; in mixture-of-experts and ensemble averaging, components are weighted by estimated reliability.^[3] Strip the substrate vocabulary and what remains is: estimate the reliability of evidence about a shared target and make influence proportional to that reliability. Context-sensitive revision is available where the substrate can support it, but a fixed inverse-variance meta-analysis remains a complete instance of the same rule.

How would you explain it like I'm…

Trust the Sure Friend

Imagine asking two friends which way to the park. One knows the way really well, and one is just guessing. You listen more to the friend who's sure. That's how you mix what people tell you — trust the steady one more than the shaky one.

Believe the Reliable One

Precision Weighting is a way to combine several reports about the same thing by trusting the more reliable ones more. Say three thermometers tell you the temperature: one is brand new and accurate, two are old and jumpy. You'd lean on the accurate one and only half-believe the jumpy ones. The amount you trust each source is its 'precision,' and your final guess is a blend where reliable sources pull harder. And trust can change with the situation — a thermometer that's great indoors might be useless in the rain.

Weighting by Reliability

Precision Weighting is the pattern of combining multiple signals about the same hidden quantity by weighting each one by how reliable it is — specifically by precision, which is one divided by the variance (the noisiness). Low-noise signals get proportionally more influence over the final estimate. So if your eyes and your ears disagree about where a sound came from, you lean on whichever sense is sharper in that moment. Crucially the reliabilities aren't fixed: the same channel can be precise in one context and noisy in another, so the weights shift with context. And a system can actively boost a channel's precision — by paying attention to it, cleaning up its input, or trusting a source more — reweighting the blend without changing the raw signals themselves.

Precision Weighting is the structural pattern by which a system integrates multiple signals about the same underlying quantity by weighting each by an estimate of its own reliability — precision, the inverse of variance — so that lower-noise signals gain proportionally more influence over the estimate or update. Five commitments define it. The system processes multiple sources of evidence (sensory channels, observations, votes, sensors, witnesses) about a shared state. Each source has an estimated precision measuring its trustworthiness in the present context. The integration rule is precision-weighted averaging or updating: each signal's contribution scales with its precision — in the Gaussian case the posterior mean is the precision-weighted average of likelihood and prior means, and in Kalman filtering the gain is itself a precision ratio. The precision estimates are dynamic and context-dependent rather than fixed. And the system can modulate precision actively — through attention, instrumentation, preprocessing, or social trust — reweighting the integration without changing the underlying signals. The skeleton recurs with unusual exactness: Bayesian posteriors, sensor fusion, predictive-processing accounts where attention increases a channel's precision, weighted wisdom-of-crowds, juries, and mixture-of-experts routers that learn to allocate precision.

Structural Signature¶

the evidence about one shared quantity — the per-source precision estimate — influence proportional to precision — optional contextual revision — the vulnerability to misestimated precision

The pattern is present when each of the following holds:

Multiple signals about a shared state. Several sources — channels, observations, votes, sensors, witnesses — bear on the same underlying quantity.
A per-signal precision estimate. Each source carries an estimate of its own reliability — precision, the inverse of variance — measuring how trustworthy it is in the present context.
A precision-weighted integration rule. Contributions scale with precision: lower-noise signals gain proportionally more influence. In the Gaussian case the posterior mean is the precision-weighted average of likelihood and prior; in Kalman filtering the gain is a precision ratio. This is the load-bearing relation, and it is provably optimal among linear combinations.
Optional contextual revision. A source's precision can be fixed for one calculation yet change across conditions; systems capable of attention, instrumentation, preprocessing, or learned routing can revise the weights without changing the underlying evidence.
A precision-of-precision vulnerability. Misestimated precision — especially overweighting a noisy source — corrupts the result in proportion to the misplaced confidence, whether or not the system actively audits that error.

These compose into a reliability-weighted influence rule: estimate each source's reliability, weight in proportion, revise the estimate when context warrants, and guard against the error that follows from getting a precision wrong.

What It Is Not¶

Not superposition. superposition is the linear combination of signals as-is; precision weighting combines them weighted by reliability, scaling each by its inverse variance rather than summing them on equal footing.
Not Bayesian updating in general. bayesian_updating is the full belief-revision machinery; precision weighting is the specific rule — weight each source by its precision — that Bayesian updating reduces to in the Gaussian case.
Not calibration. calibration is the process of making a source's confidence match its accuracy; precision weighting uses calibration estimates to set weights but is the integration rule, not the calibration procedure.
Not sampling representativeness. sampling_representativeness concerns whether sources fairly cover the population; precision weighting concerns how to combine sources of differing reliability, independent of how representative they are.
Not attention as a resource. attention is the allocation of limited processing; precision weighting is the integration rule that attention modulates — attending raises a channel's precision, but the weighting is the structure, not the spotlight.
Common misclassification. Treating precision estimates as ground truth. Overweighting an actually-noisy source — an overconfident expert, an overfit model — corrupts the result in proportion to the misplaced confidence; a mature integrator tracks the precision of its precisions.

Broad Use¶

Computational cognitive science — predictive-coding models weight prediction errors by precision; attention is precision modulation of selected channels; multisensory integration follows precision-weighted Bayesian rules, as in visual-haptic integration.^[4]
Probability and statistics — Bayesian updating as precision-weighted prior plus likelihood; inverse-variance weighting combining studies in meta-analysis; weighted least squares.
Engineering and control — Kalman filtering and sensor fusion, where the Kalman gain is the ratio of prior precision to total precision and redundant measurements are weighted by inverse noise.^[5]
Machine learning — mixture-of-experts, ensemble averaging weighted by validation performance, attention mechanisms interpreted as soft precision allocation across tokens, uncertainty-aware losses.
Organizational decision-making — expert panels weighting members by track record, review boards weighting reviewers by calibration, intelligence assessments weighting sources by reliability ratings.
Forecasting and markets — aggregation by historical accuracy or Brier score; prediction markets weighting by capital-at-risk as implicit confidence.^[6]
Clinical reasoning — physicians weighting diagnostic signals by sensitivity and specificity, with positive predictive value a precision-weighted combination at a given prevalence.^[7]

Clarity¶

Naming precision weighting separates what the signal says from how much it should count. A naive integrator treats all signals equally; a precision-weighted integrator scales contribution by reliability, and the difference is operationally decisive: averaging high- and low-precision sources equally drives the estimate toward the noisy one, whereas precision-weighting drives it toward the precise one, with a closed-form optimum in the Gaussian case. A second clarifying move is the signal-versus-weight separation: an estimate can be improved either by improving the signals themselves (more measurements, better instruments) or by improving the precision estimates that determine weighting (better assessment of which channels to trust), and these are distinct intervention families, so a failing estimate may be repaired not by acquiring more data but by recognizing that one source was over-trusted or under-trusted. A third is precision-as-attention: interpreting attention as precision modulation explains why it matters operationally — it changes neither the world nor the signals but the weighting in the integration, the same structural move as a journalist deciding which source to lead with. A fourth is the failure mode of misestimated precision: when a system overweights an actually noisy source — an overconfident expert, an overfit model — the integrated belief is corrupted in proportion to how confidently the bad source is trusted, the structural diagnosis of confirmation-bias spirals and expert-deference failures.

Manages Complexity¶

Precision weighting compresses a large family of integration problems onto a small schema: a set of signals about a shared state, a set of precision estimates (one per signal), an integration rule weighting contributions by precision, and a meta-process updating the precision estimates themselves. Once named, problems that share nothing in surface vocabulary — Bayesian belief update, Kalman sensor fusion, cortical attention, ensemble averaging, expert-panel weighting, source reliability in journalism, weighted meta-analysis — collapse onto the same axes, and substrate-specific terms (gain, weight, attention, trust, credibility) become local instantiations of one role. The compression is not merely descriptive but action-guiding, because the same intervention vocabulary applies across all of them: improve the signals, improve the precision estimates, modulate precision actively, or audit the calibration of the precision-estimation process. An analyst who has internalized the schema carries four reusable levers rather than a scattered list of domain tricks, and recognizes that tuning a Kalman gain, recalibrating an expert panel, and redirecting attention are the same kind of action applied to different substrates.

Abstract Reasoning¶

Treating precision weighting as a unit licenses several substrate-independent moves. The signal-versus-weight intervention split: integration quality can be improved by changing what is measured or by changing how much each measurement is trusted, two independent families with different costs. The optimality argument: in the Gaussian case precision-weighting is optimal, no other linear combination yielding lower variance, so "weight by reliability" is the structural target of optimal-aggregation theorems across statistics, control, and decision theory. The active-modulation argument: attention, instrumentation choice, and social trust are all acts of precision allocation, so choosing to attend, verify, or trust is structurally the same move as picking a Kalman gain. The precision-overestimation failure mode: integrators that misestimate precision systematically degrade, a unifying account of overconfidence bias, expert-deference failure, ML overfitting on noisy labels, and groupthink. The context-dependence argument: precision is not a stable property of a source but conditional on context — a forecast precise at 24 hours and noisy at 7 days, an expert calibrated in their specialty and miscalibrated outside it — so precision estimates must be conditional. And the second-order argument: a mature integrator also tracks how good its precision estimates are, the recursive meta-uncertainty layer that distinguishes it from a naive integrator that takes its precisions as ground truth.

Knowledge Transfer¶

The role mappings are stable: the integrated state maps onto the quantity being estimated from multiple signals; the signal set onto the sources, channels, or observations bearing on it; the precision estimates onto per-source reliability (inverse variance, calibration, historical accuracy); the integration rule onto the precision-weighted combination, closed-form optimal in the Gaussian case; the precision-estimation process onto the meta-procedure maintaining and updating per-source precisions; and the active modulation onto the deliberate reweighting via attention, instrumentation, or trust. Because these are pure relational notions, both the mathematics and the intervention logic transfer, and the historical record makes the ports concrete. The closed-form Kalman precision-weighted update ports directly to expert-panel aggregation once historical calibration scores are read as precisions, so weighting reviewers by Brier score is a Kalman-style update on a shared latent question. The brain's "attend to precise channels" rule ports to editorial decisions about which source to lead with, both being precision allocations rather than signal changes. Inverse-variance weighting from meta-analysis ports to combining noisy forecasts in finance or policy with an identical structural rule. Mixture-of-experts routing ports to multi-source diagnostics that weight tests by conditional sensitivity and specificity, the routing decision being a precision allocation. The systematic-degradation result for misestimated precision supplies a substrate-independent account of confirmation-bias spirals (over-weighting an agreeing source) and dismissive failure (under-weighting a disagreeing one). And active precision modulation reframes trust intervention in organizations and societies — deciding whom to trust rather than what to do — as the same structural move as attention in cortex. A physician integrating history, ECG, and a time-dependent troponin level by their context-dependent precisions, a meta-analyst combining studies by inverse variance, and a control engineer setting a Kalman gain are all doing the same structural work: estimate each source's reliability in context, weight in proportion, and guard against the bias that follows from getting a precision wrong.

Examples¶

Formal/abstract¶

Consider the fusion of two Gaussian estimates of one quantity — the prime's mathematical core, where the precision-weighting rule is a closed-form optimum. Suppose two independent sensors report measurements \(x_1\) and \(x_2\) of the same true value, with variances \(\sigma_1^2\) and \(\sigma_2^2\); the multiple signals about one shared quantity are \(x_1, x_2\), and the per-signal precision estimate is the inverse variance \(\tau_i = 1/\sigma_i^2\). The precision-weighted integration rule gives the optimal combined estimate as \(\hat{x} = \frac{\tau_1 x_1 + \tau_2 x_2}{\tau_1 + \tau_2}\) — each signal contributing in exact proportion to its precision — with combined precision \(\tau_1 + \tau_2\) (precisions add).^[1] This is provably the minimum-variance unbiased linear combination, so "weight by reliability" is not a heuristic but the structural target of an optimality theorem.^[8] The same rule is the Kalman update: in a Kalman filter the gain \(K = \frac{\tau_{\text{meas}}}{\tau_{\text{prior}} + \tau_{\text{meas}}}\) is exactly the fraction of total precision contributed by the new measurement, deciding how much a fresh observation shifts the prior.^[5] The context-dependence of precision enters when a sensor's noise depends on conditions (a rangefinder precise up close, noisy at distance), so \(\tau_i\) must be re-estimated per context rather than fixed. The precision-of-precision layer is the failure guard: if \(\sigma_2^2\) is underestimated — the sensor is actually noisier than believed — the fusion overweights \(x_2\) and the combined estimate is corrupted in proportion to the misplaced confidence. The diagnostic the prime forces: a bad fused estimate may be fixed not by adding sensors (better signals) but by recognizing that one source's precision was mis-set (better weights) — two distinct intervention families.

Mapped back: The two sensor readings are the multiple signals, inverse variance the precision, the inverse-variance-weighted mean the integration rule, the Kalman gain the precision ratio, and underestimated variance the precision-misestimation failure — the closed-form optimal aggregator.

Applied/industry¶

Consider a meta-analysis combining clinical trials, alongside the structurally identical case of a physician integrating diagnostic signals — two genuine domains running the same precision-weighted aggregation. In the meta-analysis case the multiple signals about one shared quantity are the effect-size estimates from several trials of the same treatment; the per-signal precision estimate is each trial's inverse variance, which scales with sample size — large, tight trials are high-precision, small noisy trials low-precision. The precision-weighted integration rule is inverse-variance weighting: the pooled effect is \(\hat{\theta} = \frac{\sum w_i \theta_i}{\sum w_i}\) with \(w_i = 1/\sigma_i^2\), structurally identical to the Gaussian fusion above — a large trial dominates a small one in exact proportion to precision, which is why naively averaging trial results (equal weights) is wrong and drives the estimate toward the noisy small studies.^[9] The context-dependence appears in random-effects models, where between-study heterogeneity inflates variances and down-weights over-precise outliers.^[10] The precision-of-precision layer is the meta-analyst's guard against a single large but biased trial whose reported precision overstates its trustworthiness — the same overweighting failure the prime names. The clinical parallel maps role-for-role: a physician integrating patient history, an ECG, and a time-dependent troponin level weights each by its context-dependent sensitivity and specificity (positive predictive value is itself a precision-weighted combination at a given prevalence), attends more to the test that is precise for the suspected condition (active precision modulation), and guards against anchoring on an overconfident but unreliable signal. A meta-analyst pooling trials by inverse variance and a clinician weighting tests by conditional reliability do the same structural work: estimate each source's reliability in context, weight in proportion, and guard against the bias from getting a precision wrong.

Mapped back: Trial effect sizes (or diagnostic test results) are the multiple signals, inverse variance (or test sensitivity/specificity) the precision, inverse-variance pooling (or PPV combination) the integration rule, random-effects inflation (or prevalence conditioning) the context-dependence, and an overweighted biased trial (or overconfident test) the precision-misestimation failure.

Structural Tensions¶

T1 — Improving Signals versus Improving Weights (sign/direction). A failing integrated estimate can be repaired two structurally independent ways: improve the signals (more measurements, better instruments) or improve the precision estimates that determine weighting (better assessment of which sources to trust). The failure mode is reaching reflexively for one — acquiring more data when the real fault is that one source was over-trusted, or recalibrating weights when the signals themselves are simply too noisy. Diagnostic: ask whether the error comes from what the sources say or how much each was counted — adding sensors does nothing if the existing fusion mis-set a precision, and reweighting does nothing if every signal is genuinely uninformative; misdiagnosing which family the fix belongs to wastes the intervention.

T2 — Trusting Precision Estimates versus Precision-of-Precision (measurement). The integration rule weights by precision, but the precision estimates themselves can be wrong, and a naive integrator takes them as ground truth. The failure mode is the prime's central pathology: overweighting an actually-noisy source — an overconfident expert, an overfit model, an agreeing-but-unreliable witness — corrupts the result in exact proportion to the misplaced confidence, producing confirmation-bias spirals and expert-deference failures. Diagnostic: ask whether the system tracks how good its precision estimates are, not just the estimates — an integrator that never audits the calibration of its own reliability assessments will degrade silently and most severely precisely when it is most confident in a bad source.

T3 — Context-Conditional versus Fixed Precision (temporal/scopal). Precision is not a stable property of a source but conditional on context — a forecast precise at 24 hours and noisy at 7 days, an expert calibrated in their specialty and miscalibrated outside it, a rangefinder precise up close and noisy at distance. The failure mode is treating precision as a fixed per-source attribute: carrying a reliability rating across contexts where it no longer holds, so a source trusted for its in-domain accuracy is over-weighted out of domain. Diagnostic: ask whether the precision estimate is conditioned on the current context or inherited from a different one — a precision measured in one setting and applied in another silently violates the conditionality the prime requires, and the integration inherits the error.

T4 — Optimal Weighting versus Correlated Signals (coupling). Precision-weighting is provably the minimum-variance linear combination — but the optimality theorem assumes the signals are independent. When sources are correlated (witnesses who conferred, models trained on shared data, sensors with a common disturbance), inverse-variance weighting double-counts the shared component and overstates the combined precision. The failure mode is summing precisions (\(\tau_1 + \tau_2\)) as if independent, producing false confidence in the pooled estimate. Diagnostic: ask whether the sources share an error component — if they do, the naive precision-weighted rule mis-weights them and the additive-precision claim fails, so correlation must be modeled (a covariance, a random-effects term) before the optimality of weight-by-reliability can be claimed.

T5 — Active Modulation versus Distorting the Integration (sign/direction). The system can reweight without changing signals — through attention, instrumentation, or trust — which is a powerful lever and an attack surface. The same modulation that sharpens a precise channel can, mis-applied, amplify a noisy one or suppress a disconfirming one. The failure mode is active precision allocation driven by motivation rather than reliability: attending to (up-weighting) sources that agree and down-weighting those that disagree, which is confirmation bias dressed as precision modulation. Diagnostic: ask whether a reweighting tracks an actual change in contextual reliability or merely a change in what the integrator wants to hear — active modulation is legitimate only when the precision shift reflects the world, and indistinguishable from bias when it reflects the integrator's prior.

T6 — Weighted Aggregation versus Where the Prior Dominates (scalar). Precision-weighting balances likelihood (signals) against prior, each by its precision — but when the prior's precision vastly exceeds the signals', the integration is dominated by the prior and new evidence barely moves it. The failure mode is mistaking a high-precision prior for confirmation by data: a system that "integrated the evidence" but whose posterior is essentially its prior, because the signals were down-weighted into irrelevance. Diagnostic: ask what fraction of total precision the signals contribute versus the prior — if the Kalman gain is near zero, fresh observations are being ignored regardless of what they say, and the estimate reflects the prior wearing the appearance of an evidence-weighted conclusion. This is where precision-weighting hands off to questions of prior calibration, distinct from the weighting of the signals themselves.

Structural–Framed Character¶

Precision weighting sits at the structural end of the structural–framed spectrum: it is a bare relational pattern — integrate multiple signals about a shared quantity by weighting each in proportion to its estimated precision, the inverse of variance — and signal, precision, and weighted average are pure relational notions carrying no normative or institutional load. Every diagnostic points one way.

The pattern carries no home vocabulary that must travel with it: the same precision-weighted combination is the Gaussian posterior mean, the Kalman gain, the predictive-processing account of attention, weighted wisdom-of- crowds, and a mixture-of-experts router — each field naming it in its own terms while the underlying rule arrives unmodified. It carries no inherent approval or disapproval: weighting a reliable channel more heavily is neither good nor bad until you specify what is being estimated; down-weighting a noisy witness and up-weighting a precise sensor are value-neutral operations. Its origin is formal — a precision ratio applied to signals about a common state — with no appeal to human institutions, and it runs indifferently in Bayesian inference, sensor fusion, and any estimator that must combine noisy measurements, substrates with no human practice in them. And to invoke it is to recognize a weighted-integration structure already present — the Kalman gain is a precision ratio whether or not anyone calls it that — not to import an interpretive frame. On vocabulary, evaluative weight, institutional origin, human-practice dependence, and import-versus-recognize alike, it reads structural, which is exactly the all-zeros profile the aggregate of 0.0 records.

Substrate Independence¶

Precision weighting is a maximally substrate-independent prime — composite 5 / 5 on the substrate-independence scale. On domain breadth, the reliability-weighted-signal-integration pattern recurs with identical force across computational cognitive science (predictive-coding error weighting, attention as precision modulation, multisensory integration), probability and statistics (Bayesian updating, inverse-variance meta-analysis, weighted least squares), engineering and control (Kalman filtering and sensor fusion), machine learning (mixture-of-experts, confidence-weighted ensembles, attention as soft precision allocation), organizational decision-making (expert panels weighted by track record), forecasting and markets (aggregation by historical accuracy, prediction-market capital-at-risk), and clinical reasoning (weighting diagnostic signals by sensitivity and specificity) — physical, computational, biological, and institutional substrates alike, a clear 5. On structural abstraction, signal, precision, and weighted average are pure relational notions with no normative or institutional load; the precision-weighted combination is the Gaussian posterior mean, the Kalman gain, and the MoE router, each field naming it in its own terms while the rule arrives unmodified, and it runs in sensor fusion with no human practice — a 5. On transfer evidence, the prime scores a 5: the closed-form Kalman precision-weighted update ports directly to expert-panel aggregation once calibration scores are read as precisions, inverse-variance weighting from meta-analysis ports to combining noisy forecasts, mixture-of-experts routing ports to multi-source diagnostics, and the misestimated-precision degradation result supplies a substrate-independent account of confirmation-bias spirals — documented, bidirectional transfer. Every component reads maximal, anchoring the composite at 5.

Composite substrate independence — 5 / 5
Domain breadth — 5 / 5
Structural abstraction — 5 / 5
Transfer evidence — 5 / 5

Relationships to Other Abstractions¶

Current abstraction Precision Weighting Prime

Parents (1) — more general patterns this builds on

Precision Weighting is a kind of Aggregation Prime

Precision weighting is aggregation specialized to signals about one target whose influence scales with estimated inverse variance or an equivalent reliability measure.

Children (1) — more specific cases that build on this

Bayesian Cue Integration Prime is a kind of Precision Weighting

Bayesian cue integration is precision weighting specialized to several partly independent simultaneous cues about one latent quantity, with inverse-variance fusion and a lower-variance combined estimate.

Hierarchy path (1) — routes to 1 parentless root

Precision Weighting → Aggregation → Micro Macro Linkage

Neighborhood in Abstraction Space¶

Precision Weighting sits in a moderately populated region (50^th percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.

Family — Measurement, Attestation & Signal Weighting (14 primes)

Nearest neighbors

Bayesian Cue Integration — 0.78
Correlated-Source Attribution Failure — 0.73
Measurement Uncertainty and Observational Noise — 0.70
Imputation — 0.70
Monte Carlo Simulation — 0.70

Computed from structural-signature embeddings · 2026-09-10

Not to Be Confused With¶

The deepest confusion is with bayesian_updating, because precision weighting is what Bayesian updating becomes in the Gaussian case, and the two are routinely treated as synonyms. The distinction is one of scope. Bayesian updating is the general machinery of revising a probability distribution in light of evidence — applicable to any likelihood and prior, producing a full posterior distribution. Precision weighting is the specific rule that emerges when the distributions are Gaussian: the posterior mean is the inverse-variance-weighted average of the signals, and precisions simply add. Holding them apart matters because the precision-weighting rule is a clean, closed-form, computationally cheap aggregator that applies wherever sources can be summarized by a reliability scalar — even in settings (expert panels, sensor fusion, meta-analysis) where no one is running full Bayesian inference. Conversely, when the distributions are not Gaussian, the precision-weighting shortcut can mislead, and the fuller Bayesian machinery is required. Treating every reliability- weighted average as "Bayesian" overstates the assumptions; treating Bayesian updating as always reducible to precision weighting ignores the non-Gaussian cases where it does not.

It must also be distinguished from calibration, with which it is tightly coupled but distinct. Calibration is the property (and the process of achieving it) that a source's stated confidence matches its actual accuracy — that a forecaster who says "70%" is right 70% of the time. Precision weighting consumes calibration as an input: it needs each source's precision to weight correctly, and a well-calibrated source supplies a trustworthy precision. But precision weighting is the integration rule that combines sources; calibration is the upstream property that makes the precision estimates reliable. The prime's central failure mode — overweighting an overconfident source — is precisely a calibration failure feeding into the weighting: the source reports high precision it does not deserve, and the weighted average inherits the error. Confusing the two leads to fixing the wrong thing: recalibrating sources when the integration rule is sound, or retuning weights when the underlying confidences are miscalibrated.

A third confusion is with attention, because the prime interprets attention as precision modulation and the two are genuinely linked. Attention is the allocation of a limited processing resource — a spotlight that selects some channels over others. Precision weighting is the integration rule that attention acts on: attending to a channel raises its precision and so its influence over the integrated estimate, without changing the signal itself. But attention is the act of allocation; precision weighting is the structural combination that the allocation tunes. The value of the distinction is that it explains why attention matters operationally — it changes neither the world nor the signals but the weighting in the integration — while keeping clear that the weighting rule exists independently of whether any attentional spotlight is moving.

For a practitioner these distinctions sort the interventions. A Bayesian-updating frame brings the full distributional machinery (needed only off the Gaussian case); a calibration frame fixes the reliability of the inputs; an attention frame reallocates the spotlight. Precision weighting itself is the integration rule — weight by contextual reliability, modulate actively, and audit the precision of the precisions — and knowing which neighbor a failure belongs to (miscalibrated source, wrong distributional assumption, misallocated attention) is what directs the fix to the right layer.

Solution Archetypes¶

Solution archetypes in the catalog that build on this prime — directly (this prime is a source ingredient) or as a related prime.

Built directly on this prime (4)

Also a related prime in 7 archetypes

Conditioned Probability Frame Specification: State what is being taken as given before interpreting, comparing, or acting on a probability.
Event-Rate Magnitude Encoding: Encode intensity as event frequency and decode it by counting or integrating over a calibrated window rather than by inspecting any single event.
Mode-Setting Gain Modulation: Use a separate noncontent channel to retune how many content channels are processed, so the system changes sensitivity or mode without rewriting the content itself.
Model-Guided Signal Separation: Recover a target component from mixed observations by stating what the target is, modeling how target and nuisance combine, applying a calibrated separator, and proving what the output preserves, suppresses, and still leaves uncertain.
Prediction-Error Learning Calibration: Teach from the signed gap between expected and received value so surprise updates the model while expected outcomes do not keep pretending to teach.
Reference Tracking Bandwidth Alignment: Make the demanded trajectory trackable by matching reference update speed to the loop bandwidth that can actually observe, decide, act, and settle.
Shared-Source Variance Isolation: Prevent a single hidden source from making multiple supposedly independent dimensions look more correlated than they really are.

References¶

[1] Gelman, Andrew, John B. Carlin, Hal S. Stern, David B. Dunson, Aki Vehtari, and Donald B. Rubin. Bayesian Data Analysis. 3^rd ed. Boca Raton: CRC Press, 2013. Derives the Gaussian posterior mean as the precision-weighted average of prior and likelihood, with precisions adding. registry ↩a ↩b

[2] Feldman, Harriet, and Karl J. Friston. "Attention, Uncertainty, and Free-Energy." Frontiers in Human Neuroscience, vol. 4 (2010): article 215. Models attention as inferring and weighting by the precision of prediction errors in predictive processing. registry ↩

[3] Jacobs, Robert A., Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. "Adaptive Mixtures of Local Experts." Neural Computation, vol. 3, no. 1 (1991): 79–87. Introduces mixture-of-experts, with the gating network a learned reliability-weighted allocator across experts. registry ↩

[4] Ernst, Marc O., and Martin S. Banks. "Humans Integrate Visual and Haptic Information in a Statistically Optimal Fashion." Nature, vol. 415, no. 6870 (2002): 429–433. Demonstrates that multisensory integration follows precision-weighted (inverse-variance) maximum-likelihood combination. registry ↩

[5] Kalman, Rudolf E. "A New Approach to Linear Filtering and Prediction Problems." Journal of Basic Engineering, vol. 82, no. 1 (1960): 35–45. Introduces the Kalman filter, whose gain is the ratio of measurement precision to total precision. registry ↩a ↩b

[6] Wolfers, Justin, and Eric Zitzewitz. "Prediction Markets." Journal of Economic Perspectives, vol. 18, no. 2 (2004): 107–126. Describes aggregation of dispersed information in prediction markets, with capital-at-risk as implicit confidence weighting. registry ↩

[7] Sackett, David L., R. Brian Haynes, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: A Basic Science for Clinical Medicine. 2^nd ed. Boston: Little, Brown, 1991. Develops diagnostic-test reasoning by sensitivity and specificity, with positive predictive value as a prevalence-conditioned combination of test signals. registry ↩

[8] Kay, Steven M. Fundamentals of Statistical Signal Processing, Vol. I: Estimation Theory. Englewood Cliffs: Prentice Hall, 1993. Proves that inverse-variance weighting yields the minimum-variance unbiased linear combination of independent estimates. registry ↩

[9] Borenstein, Michael, Larry V. Hedges, Julian P. T. Higgins, and Hannah R. Rothstein. Introduction to Meta-Analysis. Chichester: Wiley, 2009. Standard reference deriving inverse-variance weighting for pooling effect sizes across studies. registry ↩

[10] DerSimonian, Rebecca, and Nan Laird. "Meta-Analysis in Clinical Trials." Controlled Clinical Trials, vol. 7, no. 3 (1986): 177–188. Introduces the random-effects model, inflating variances to down-weight over-precise outliers under between-study heterogeneity. registry ↩