Correction Latency
The time between the detection of an execution error and the verified restoration of shared operational reality.
Concept origin: Correction Latency is dimension 5 of the Execution Integrity Score framework, developed by Duena Blomstrom and Dave Ballantyne as part of the Human Debt™ research programme.
What is Correction Latency?
Correction Latency is the elapsed time between the moment an execution error is detected — a verified divergence from shared operational reality — and the confirmed restoration of that reality. It is not the time to first response, nor the time to declare a fix, but the time to verified recovery. High latency means the organisation lives in a correcting state for extended periods; low latency means recovery is structural and fast.
Why does Correction Latency matter under AI acceleration?
AI acceleration increases both the frequency of divergence events and their downstream blast radius. An error that once affected one team now propagates across interconnected AI-assisted workflows before any single actor notices. The organisation that cannot correct fast cannot survive the compounding. Correction Latency is the measure of whether recovery capacity is keeping pace with drift velocity.
How is Correction Latency scored?
Scored 1–5 as dimension 5 (d5) of the Execution Integrity Score v1.0. A score of 1 indicates errors propagate for weeks before verified recovery begins. A score of 5 indicates errors are detected, isolated, and verified-corrected within short, consistent windows, with documented recovery protocols that function across all critical workflows.
What is the difference between detecting an error and correcting it?
Detection is the first signal that something is wrong. Correction Latency begins at detection and ends only when shared operational reality is verified as restored — not when a fix is deployed, not when a ticket is closed. Many organisations have good detection but poor correction infrastructure: they know something is wrong but cannot stabilise the downstream effects within a recoverable window. This is the gap Correction Latency measures.
What high and low Correction Latency looks like
High Correction Latency (SAND signal)
- — A pricing error appears in an AI-generated quote. Three weeks pass before it is identified as a systematic error rather than a one-off. Another week to verify the full blast radius. Recovery takes a further month.
- — An AI-assisted performance summary contains a factual error about a team member. The individual is managed based on it for two quarters before a human audits the source data.
- — A model drift in a customer-facing recommendation engine compounds for sixty days before a downstream metric signals that something is wrong.
Low Correction Latency (GLASS signal)
- — Anomaly detection flags an AI output divergence within 24 hours. A named human confirms the error, isolates affected decisions, and verifies recovery within 72 hours.
- — A correction protocol exists before the error occurs: who owns detection, who confirms divergence, who approves recovery, who verifies restoration.
- — Recovery metrics are tracked — not just incident count, but mean time to verified recovery — and are reviewed at senior leadership level monthly.
How to measure Correction Latency in your organisation
Step 1 — Define your detection boundary
Latency begins at verified detection — not when someone suspects a problem, but when a divergence from shared operational reality is confirmed. Define what constitutes a verified detection event in your context. Without this, latency measurement starts at different points for different incidents.
Step 2 — Define your recovery boundary
Latency ends at verified recovery — not when a fix is deployed, but when shared operational reality is confirmed as restored. Identify who verifies recovery, what evidence they examine, and what constitutes confirmation. "It looks fixed" is not a recovery boundary.
Step 3 — Log the last five significant errors
For each: when was detection confirmed? When was recovery verified? What was the gap? If you cannot answer these questions for the last five errors, your correction latency is unmeasured — which is itself a SAND signal.
Step 4 — Score against the EI Scale
A d5 score of 1 means errors take weeks; a score of 5 means consistent recovery within short, documented windows. The 25-question EI Diagnostic scores this dimension independently and places it in the context of your other four dimensions.
Measure your organisation's Correction Latency
Correction Latency (d5) is scored as part of the full 25-question Execution Integrity Diagnostic. A weak d5 score is the clearest leading indicator of execution survivability risk.
12-Question Screening Tool
Get a signal on your correction infrastructure in 4 minutes. Identifies structural risk patterns. Free.
aiadoptionperformance.com →
25-Question Full Diagnostic
Scores Correction Latency (d5) separately — the single highest-value dimension for organisations approaching SAND territory.
aiadoptionperformance.com →
Learning pathway