OptiVis Labs

Research / Explainers / 4 of 6

Teacher stamps, student copies, a held-out deck grades the copy

Jev is the judge model OptiVis runs on its own copy and decisions. A student classifier was trained on Jev's labels and then graded on held-out labels. This illustration shows the three steps and the recorded agreement in two conditions: labels from knowledge bases the student saw in training, and labels from knowledge bases it did not. The baseline is Laya, the local decision model, with no training on those labels.

  • StatusIllustration
  • Published2026-10-06
  • Example textInvented
  • Embedded onThis page only
  • State in the address?kb=
  • Codelib/explainers/jev.ts, lib/readings.ts
  • Protocolv1.0

IllustrationExplainer

Teacher stamps, student copies, a held-out deck grades the copy

Agreement with the judge, not understanding.

Teacher stampsJev labels each itemStudent copiesa small classifier trains on the labelsHeld-out decklabels the student never trained onStudent, in-KB89.0%Laya, zero-shot52.1%agreement with Jev, one cell per point,recorded 2026-10-05
Held-out condition

Student agrees with Jev89.0%

Laya, zero-shot baseline52.1%

in-KB: held-out labels from knowledge bases the student saw in training.

  • Recorded
  • Trained on 68,828 deduplicated, decontaminated passages labelled by Jev; evaluated on 4,387 in-KB and 1,301 unseen-KB passages
  • Agreement with the teacher judge, not accuracy against ground truth
  • Receipts not yet published
  • Protocol v1.0
Recorded run, 2026-10-05, trained on 68,828 deduplicated, decontaminated passages labelled by Jev; evaluated on 4,387 in-KB and 1,301 unseen-KB passages. Agreement with the judge, not understanding. Each cell is one percentage point, rounded to whole cells; the exact figure is printed beside the deck. Receipts not yet published.

The sentence it teaches

Agreement with the judge, not understanding.

Method

  1. Teacher: Jev labels each item. The labels are deduplicated and decontaminated before use.
  2. Student: a small classifier is trained on those labels and runs without Jev afterward.
  3. Held-out deck: labels the student never trained on. Agreement is the share of held-out items where the student's label equals Jev's.
  4. Two conditions: in-KB labels come from knowledge bases the student saw in training; unseen-KB labels come from knowledge bases it did not.
  5. Baseline: Laya, the local decision model, labels the same held-out items zero-shot, with no training on those labels.

What this does not show

  • It does not show how either model performs on anything outside those labels, and it does not compare against human judgment.

State in the address

The figure keeps its state in the query string, so a particular view can be linked and the page still renders completely without it.

  • ?kb= the condition shown: in (in-KB) or unseen (unseen-KB)

Cite this explainer

OptiVis Labs. Teacher stamps, student copies, a held-out deck grades the copy (explainer). 2026. https://optivisai.org/research/explainers/jev-distillation

@misc{optivislabs2026explainerjevdistillation,
  title = {Teacher stamps, student copies, a held-out deck grades the copy (explainer)},
  author = {{OptiVis Labs}},
  year = {2026},
  month = oct,
  howpublished = {\url{https://optivisai.org/research/explainers/jev-distillation}},
  note = {Agreement with the judge, not understanding. Illustration with invented example text.}
}