Skip to content

Precision-Weighted Error Gate

Residual-prioritisation method — instantiates Predictive Residual Processing

Scores each residual by magnitude, uncertainty, source reliability, consequence, and capacity cost, and admits only the ones worth the scarce bandwidth.

A Precision-Weighted Error Gate does not compute residuals; it decides which of them deserve to be seen. Every candidate deviation is scored by combining its raw magnitude with the precision of its source — how much to trust it, given sensor reliability and current uncertainty — and then by its consequence and the capacity cost of forwarding it. Only high-scoring residuals pass; the rest are logged or dropped. Its defining idea, borrowed from precision-weighting in predictive coding, is that a large error from an untrustworthy source can matter less than a small error from a trusted one: it is the uncertainty-weighted, cost-aware rationer that keeps a finite channel pointed at the surprises that are both real and important.[1]

Example

A patient in intensive care is wired to monitors that, unfiltered, throw hundreds of threshold-crossings a day — most of them artifacts, expected variations, or the momentary noise of a probe being jostled. A precision-weighted error gate sits between the monitors and the nurse's pager. Each deviation is scored: magnitude (how far from predicted), precision (is the SpO₂ probe firmly attached and reading cleanly, or loose and noisy right now), consequence (a rhythm deviation outranks a mildly high skin temperature), and capacity cost (a nurse covering six beds can meaningfully attend to only so many alarms an hour). A deep SpO₂ drop from a firmly-seated probe scores high and pages immediately; the same numeric drop from a probe the system knows is slipping scores low and is quietly logged. The point is not fewer alarms for their own sake — it is spending a nurse's finite attention on the deviations most likely to be both real and dangerous.

How it works

  • Score, don't just threshold. Each residual gets a composite weight from magnitude × precision × consequence, discounted by the cost of forwarding it — not a single fixed cutoff.
  • Precision as gain. The inverse of a source's current uncertainty multiplies its errors, so a reliable channel's small deviations can outrank a flaky channel's large ones.
  • Ration against a finite budget. Residuals compete for a limited number of downstream slots; the gate admits the top-scoring few and holds total suppressed error under an explicit budget.
  • Log the losers. Suppressed residuals are not discarded silently but retained, so the cost of the gate's own choices can later be audited.

Tuning parameters

  • Precision sensitivity — how sharply low source-reliability discounts a residual. Aggressive discounting silences noisy channels but can mute a genuine event from a chronically-noisy sensor.
  • Consequence weighting — how much a high-stakes source is boosted. Set from harm, not from volume; weighting by how often a source fires is how minority-but-critical signals get buried.
  • Bandwidth budget — how many residuals per unit time may pass. The core rationing dial: tight budgets protect attention but raise the bar a real surprise must clear.
  • Suppressed-error budget — the total unsent deviation the design tolerates before the gate is forced to widen. Keeps "saving bandwidth" from quietly becoming "hiding a lot."

When it helps, and when it misleads

Its strength is fighting saturation: when a channel or a human is drowning in deviations, the gate concentrates finite attention on the ones that are both trustworthy and consequential, and it makes the cost of a surprise explicit instead of pretending attention is free. It is what stops a residual pipeline from degenerating into an unreadable stream of noise.

It misleads when its weights encode the wrong priorities. A mis-set precision or consequence term will systematically suppress an entire class of real signals — a minority sensor, a rare failure mode — and because the suppression is silent, no one notices the gate has been quietly editing reality. The textbook misuse is running it backwards: tuning the weights until the alarm rate is comfortable, which optimises for a quiet room rather than a safe patient and is precisely the dynamic behind alarm fatigue.[1] The discipline is to set consequence weights from harm analysis, to keep and periodically audit the suppressed stream (that is what Shadow Raw-Channel Sampling is for), and never to tune the gate against alarm counts.

How it implements the components

This method owns the archetype's prioritisation-and-rationing slice:

  • precision_weighting_rule — the composite scoring of each residual by magnitude, precision, consequence, and cost is this mechanism's core rule.
  • attention_and_bandwidth_budget — the finite downstream capacity it rations against; the gate exists only because attention is scarce.
  • residual_error_budget — it holds total suppressed error under an explicit ceiling, so prioritisation never quietly drops more than the design allows.

It does not compute the residuals it ranks — that is the Innovation Residual Filter and the comparators upstream — nor does it audit what it suppressed; checking the dropped stream against ground truth is Shadow Raw-Channel Sampling, and reverting entirely to raw is the Raw-Signal Fallback Switch.

Notes

The gate is the dynamic, multi-factor cousin of the Confidence Threshold Table: where the table is a static lookup ("flag anything past this line"), the gate weighs several factors and the cost of acting, so the same-sized deviation can pass or fail depending on source trust and current load. A system with abundant bandwidth may need only the table; one running near its attention limit needs the gate.

References

[1] In predictive coding, prediction errors are precision-weighted — multiplied by an estimate of their reliability (inverse variance) — before they drive inference, so unreliable channels are automatically down-weighted. The applied-side failure this mechanism guards against is alarm fatigue: when monitor alarms are too frequent or too often false, responders desensitise and miss the real ones, which is why tuning a gate to minimise alarm counts rather than to preserve true positives is a known clinical hazard, not a virtue. ↩