A plausible match was still the wrong match
October 2026 · A note from an anonymized operations workflow
The first version looked reasonable: use the first six characters of a PNR to match records. The rule was easy to explain, and the output looked consistent. AI helped turn that idea into code. Neither of those things made the assumption correct.
The check that changed the approach
When I compared the matches with the actual seat data, roughly a quarter were wrong. The code had followed the rule; the rule was the problem. A plausible shared identifier was not enough to establish that two records referred to the same seat.
I changed the matching approach to use seat data. The important step was going back to what the records represented and checking the result against that reality. I am not publishing a dataset or a new accuracy figure here; this is the lesson from that correction.
My part of the work
I bring the operational question, decide what a correct result means and inspect the mismatches. AI tools help me explore a rule and implement it. The responsibility for choosing the evidence and accepting the result stays with me.
That changes what I ask during development. Alongside “does the code run?”, I ask “what would show that this assumption is wrong?” In this case, the answer was a comparison with actual seats. A clean build could never have answered that question.
What I carry into other projects
I try to make the evidence visible: Jazari has published model files and evaluation notes; Qragy makes its retrieval and human handoff workflow inspectable in source. The evidence is different for each project, but the habit is the same: choose a concrete check before trusting a plausible result.
These are personal notes. The Research archive contains separately labeled, automatically collected AI briefings.