What gets logged
Whenever AI assists evaluation — criterion-level evidence extraction, a draft score suggestion, a transcription, a transcript evaluation, or a candidate summary — the output is logged with model and version provenance. You always know which model, at which version, produced a given suggestion, so an output can be traced back to its source even long after the fact. AI-drafted library items follow a different guarantee: they are marked as drafts precisely so a person reviews them before they can score anyone.
Where provenance appears
Provenance shows up where it matters most. Transcript evaluations and candidate summaries display the model that produced them right in the interface, so an evaluator sees what they are working from; every assisted evaluation output is additionally logged server-side with model, version, and usage in the platform's AI generation log.
The record itself stays human: the audit pack — the self-contained decision file — documents the AI policy and shows every score under its named evaluator with a timestamp, while the model and version log for each assisted step is kept in the platform's AI generation log.
The confirm-or-override rule
No AI output enters the decision record on its own. A named person confirms or overrides every AI suggestion: an AI-drafted score only becomes a score when an evaluator accepts it under their own name, and AI-generated items are drafts you review before use. In every case, the record shows a person, not a model, as the decision-maker.
Why this matters for defensibility
When a candidate challenges an outcome or a client audits your process, "the AI suggested it" is not an answer — but "this named evaluator confirmed this score, assisted by this model at this version, on this date" is. Provenance logging turns AI assistance from a liability into a documented, reviewable step in the decision.
It also makes AI's role bounded and inspectable: because every assisted evaluation output is logged with its model and version and confirmed by a named person, you can demonstrate to a candidate, a client, or an auditor that AI supported the decision without ever making it.
At a glance
- AI model and version are logged for every assisted output used in evaluation — evidence extraction, score suggestions, transcriptions, and summaries.
- Transcript evaluations and candidate summaries show the model in the interface; the full log lives server-side.
- Every AI suggestion is confirmed or overridden by a named person before it counts.
- Every score in the record carries a named evaluator and a timestamp, whether AI-assisted or not.
- The trail exists so decisions can be explained and defended later.