Skip to main content
Xerg treats a pattern as monetary waste only when the source supplies enough content-free evidence to correlate the economic event, establish the relevant counterfactual, and assign each cost component once. Everything else remains a neutral signal or is marked unavailable.

​
Three monetary findings

​
Retry waste

retry-waste requires a stable logical request ID, ordered attempt IDs, an explicit charged failure or abort, and a higher explicit successful attempt. Affected spend covers the complete attempt chain. Avoidable spend covers only the charged failed or aborted attempts. An error followed later by another call, a same-model timing window, or an attempt number without stable request correlation is not enough.

​
Tool-loop waste

tool-loop requires the same tool name, identical content-free input fingerprint, unchanged result or repeated error fingerprint, unchanged state-progress fingerprint, exact tool/model correlation, and exact cost correlation. The first execution is necessary; only later no-progress repetitions are avoidable. Same tool names with different inputs, changed results, or demonstrated state progress do not produce a finding.

​
Cache churn

cache-thrash requires a stable cache-entry fingerprint, an observed write/read lifecycle, cache token counts, applicable cache pricing, and an uncached-input price. It emits only when the actual cache lifecycle costs more than the uncached counterfactual. The positive cache premium is the avoidable amount. Aggregate cache ratios, an unknown cache identity, or beneficial discounted cache reads do not produce a finding.

​
Seven neutral signals

The current signal registry is:
  • deep-loop-activity
  • context-outlier
  • context-growth
  • fixed-cadence-activity
  • premium-model-routine-label
  • cache-read-concentration
  • max-mode-concentration
Signals include only bounded observed metrics and optional associated spend with an observed, catalog-estimated, or mixed basis. Human-readable output labels associated spend as “not classified as waste.” Signals have no waste/opportunity classification, avoidable-spend value, recommendation, optimization record, or CI impact. Deep work, scheduled work, premium models, large context, cache reuse, and Max Mode can all be justified; the observed pattern alone cannot decide that question.

​
Cost ownership

Xerg’s component-level claim ledger is keyed by economic event and cost component:
  • retry claims take precedence over tool-loop claims for the same model charge
  • tool-loop claims own only repeated model charges after the first execution
  • cache findings own only the positive cache-premium component
  • aggregates and timelines sum owned claims once
This means affectedSpendUsd can describe a broader pattern while avoidableSpendUsd remains the only amount used for savings, recommendations, waste totals, and gates.

​
Coverage is part of the result

Each active finding and each signal has separate coverage. Missing stable IDs, fingerprints, lifecycle data, progress evidence, or pricing produces an explicit unsupported reason. Xerg never fills those gaps with heuristic inference. QM and Cursor currently report all active monetary findings as unavailable because their present source contracts cannot prove them. They may still emit the neutral signals their data directly supports. OpenClaw, Hermes, Claude Code, and generic ingest gain coverage only when the underlying record actually includes every required observation.

​
Legacy audits

Wire v5 and earlier findings remain readable with a Legacy detection method badge and retain their historical values. They cannot create or recur current optimization records. Wire v6 recurrence includes detectorVersion, so legacy status cannot silently transfer to an evidence-strict finding with the same textual kind. For the exact payload and recommendation fields, see findings, signals, and recommendations.