Skip to content
Pondara Pondara docs

Evidence and verdicts

Every engagement produces evidence — observations about how a learner performed on the dimensions the scenario exercised. Pondara keeps two layers of evidence on every record: what the AI proposed after watching the dialogue, and what the trainer decided. The trainer is the canonical voice. The AI is visible alongside.

AI signal

After an engagement closes, the AI extracts what it noticed from the dialogue and proposes an AI signal for each affected dimension: a suggested direction (the learner appears stronger, weaker, or unchanged on this dimension), a confidence level, and the turns that support it.

The signal is visible to both the learner and the trainer. It is not an assessment. It does not move the learner’s record. It is a structured first read so the trainer does not have to scrub the whole dialogue cold.

Low-confidence signals are not hidden. They land lower in the trainer’s queue, but they are still on record. Suppressing diagnostics would defeat the point of having them.

Trainer verdict

The trainer reviews each signal and chooses one of three outcomes:

  • Accept. The trainer agrees with the AI’s read. One click.
  • Adjust. The trainer keeps the dimension but changes the magnitude or the direction, optionally with a written rationale.
  • Reject. The trainer disagrees. The signal is recorded as dismissed but stays visible.

Only the verdict updates the learner’s competency record. Until a verdict exists, the AI signal sits on the engagement and the record does not move.

A trainer can also write evidence the AI missed, or evidence from parts of the dialogue the AI did not surface. That goes through the same shape: a verdict with no AI signal underneath.

Status

Each evidence row carries one of four statuses:

  • Pending — the AI proposed a signal; the trainer has not acted.
  • Accepted — the trainer accepted the signal as written.
  • Adjusted — the trainer kept the dimension but changed the call.
  • Rejected — the trainer disagreed; the signal is on file but does not move the record.

The learner sees the canonical line on their radar: what their trainer concluded, with the AI signal collapsible underneath if they want to read what the AI thought. Trainers see the full queue with pending items at the top.

Disagreement

Trainers and AI will disagree. That is fine — and useful. When the trainer’s verdict diverges substantially from the AI signal (the dimension reverses, or the magnitude changes by a wide band), the record is flagged as a disagreement.

Disagreements show up in two places:

  • The trainer’s own reflection log, so they can spot patterns in their overrides.
  • An aggregated view your org admin can see, so repeated disagreements on a single scenario or a single learner trigger a conversation. A scenario whose AI signals keep getting overridden may need a rewrite. A learner whose AI signals keep getting overridden may need a different cohort assignment.

Disagreements do not show up to the learner as a flag. They see both lines without a “your trainer overruled the AI” stamp; that adds noise without insight.

Where to look next