Clear the list
Yesterday a recruiting agent posted roles off an open-roles count that one region had loaded twice. This morning the count gets corrected. Fixing the number is one write. What’s owed after it is a list of every action that used the old value, and one entry on it won’t take an assumption.
File the correction
This is the metric as the inspector shows it, the morning after the guest piece ended. The rerun has the right number. Filing the correction puts it in force.
Metric inspector
ingest 06:00
Open requisitions
All regions · Headcount workspace
Served since 06:00
412
File a correction
- Scope
- Whole metric
- Reason
- Central region requisitions loaded twice in the 06:00 ingest.
- Fixed by
- The 09:58 rerun, 356 open requisitions
Every read from here on gets the right count. The dashboard is green again. Nothing that already ran on the old count has heard a thing.
Clear the list
The same write returned a gated list, one entry for every agent action that committed on the old value. Each clears by alerting its owner, by alerting the leaders who used its output, or, when the output was never used, by an exception update with a reason. Clear the roles and the email, then try the easy way out on the deck.
File the correction above. The list is built in the same write.
A slide can’t say whether anyone read it, so the agent’s write counts as potential use. The list won’t let you wave it through. Somebody has to be told.
What the grade says
This is the grading record for the roles the agent posted, written when the action committed. Look at it again after the correction.
Grading record
written 09:14
action a-7f31
Posted the Central region roles · committed
- Value served
- 412
- Trust score at fetch
- 86
- Act threshold published
- 80
- Agent chose
- Act
- Grade
- Correct
The agent was right about the number it was served. The loss goes on the metric, not the action, and the people who relied on the action get told.
The essay carries the law that already asks for this walk, Article 19 of the GDPR, the food recall that worked it out first, and why the grade stays where it was.
One essay a week on making data and AI systems trustworthy enough to act on.
- [1]After the Correction, The Truth Layer, 2026
- [2]The Data Product’s Second Customer, Modern Data 101, 2026
- [3]Verdict correction gate, CORRECTION-GATE-SPEC.md and the correction card