A correction can remain true and still stop being sufficient.
“The file was saved” may accurately describe what happened. But if the user needed to see the result and continue working, saving the file answers only part of the question. The implementation may have passed its check while the check missed the intended outcome.
That distinction became central while we developed Trust Without Perfection, a working paper about correction in human–AI workflows.
We began with a familiar question: how do people and AI systems catch and repair mistakes? The more useful question became what happens after the repair.
Remembering the correction isn’t the whole job
A handoff might preserve:
We checked this. It works.
But what was checked? Under which conditions? What does that result authorize us to conclude now?
Without that scope, a valid observation can become an overly broad assurance. The words survive; the limits disappear.
The opposite failure is possible too. A system can repeatedly reopen settled questions, spending time and attention without improving the decision. More checking is not automatically better judgment.
Useful continuity has to support both:
- Reusing a correction when its conditions still hold.
- Reconsidering it when the intended use or relevant conditions change.
Sometimes the test needs correcting
Sometimes the intended use changes after verification. Sometimes it stays exactly the same, and we discover that our original checks never covered it adequately. The first calls for assessing a new requirement; the second calls for repairing an incomplete test.
When a result disappoints, it is tempting to assume the implementation failed.
Sometimes it did exactly what the acceptance criterion required. We simply required too little.
“Saved,” “visible,” and “ready for the next action” are different claims. Evidence for one does not automatically establish the others.
This matters for evaluation as well as development. A benchmark can reward a system for satisfying an incomplete definition of success. Correcting the system without examining that definition may make it better at passing the wrong test.
What we have—and haven’t—established
Our development observations helped us formulate these distinctions. They do not establish that our proposed corrective workflow outperforms a capable human–AI pair, or that structured correction records cause better later decisions.
The working paper proposes ways to investigate those questions. Those experiments remain unrun.
The next comparison we want to make is deliberately narrow: give fresh workers the same verified correction, conveyed either through a scoped correction record or a length-matched ordinary handoff. Then examine both routine reuse and cases where the correction’s original scope no longer suffices.
That would help distinguish discovering a repair from carrying it forward usefully.
For now, the practical lesson is smaller:
Preserve what a correction establishes, the conditions under which it holds, and what it leaves unresolved.
A correction should help the next decision—not merely give it something reassuring to quote.
This field note continues the question raised in Remembering Is Not Enough. It reports no new experimental results. The working paper remains unpublished while we prioritize evidence.