Agentic Engineering System · recorded verification run
Which findings reached a person, and why.
An AI analysis of interviews produced seven findings. Before any of them could be trusted, two independent AI auditors checked each one against the interview quotes it cites. Where the auditors agreed, the finding was accepted without a person. Where they disagreed, it was sent to a person with both opinions attached. Nothing was rejected by a model alone.
Loading recorded run…Findings from the Qualitative Analysis Workbench's analysis of a public interview corpus, each citing exact passages.
Software re-hashed every cited passage against the source. A small model investigated each finding; a larger model reviewed it independently.
Disagreement is treated as a reason for human judgment, not settled by a vote or by the bigger model.
Result
Escalate judgment, not activity.
The seven findings
Each finding, both verdicts, and the quotes the auditors saw.
Auditors saw at most a few quotes per finding. That is why "overstated" sends a finding to a person instead of rejecting it: the finding may be right on evidence the auditor was not shown.
Why this needed custom work
One "check this claim" call is not verification.
In three repeated runs of the same audit, the same finding was accepted once and corrected twice. A single model pass can wave a false finding through. So this system re-verifies every quote outside the model, uses two auditors with different models, and treats their disagreement as the signal that a person is needed. The orchestration, routing, and durable human inbox are the engineering; the auditors are replaceable.
Provenance
Where these numbers come from.
Everything on this page is read from one recorded run record. The system's repository is private; the identifiers below let a reviewer match this page to that record.