Skip to main content
← Back to iBacalao

Model transparency

How iBacalao's assessment AI works

iBacalao uses a language model fine-tuned for IB coursework assessment and revision, combined with subject-, criterion-, level- and session-specific context. This page separates what is technically true, what has been internally tested, and what we do not claim.

Last transparency review: 22 September 2026. Publishing and correction standards are described in our editorial policy.

What “fine-tuned” means here

For this page, a fine-tuned language model means a language model whose model weights or dedicated adapter/checkpoint have been modified through additional training for the IB assessment task. That is different from using only prompts, retrieval, or supplying IB context to an otherwise unchanged general-purpose model.

The review workflow also applies configured rubric, subject, level and assessment-session context. Fine-tuning by itself is not presented as proof that a score is correct.

What has been evaluated

Internal QA has tested whether feedback is specific, uses appropriate IB vocabulary, respects student level, scales with draft sophistication, and gives actionable next steps. We also compare language paths and inspect criterion-selection errors and factual mistakes.

Some internal QA uses another language model as a judge. That is an engineering test, not an independent examiner study. It can reveal regressions and inconsistencies, but it does not establish agreement with IB examiners or predict a final moderated grade.

What we do not claim

  • We do not claim that an iBacalao score is an official IB grade or predicted final grade.
  • We do not claim examiner-equivalent scoring accuracy.
  • We do not currently publish a blinded human-examiner benchmark with agreement, error or calibration statistics.
  • We do not treat fine-tuning as evidence of quality on its own; product quality must be evaluated separately.

Historical “first fine-tuned” claim

Evidence review last performed: 19 August 2026

In August 2026, we reviewed publicly available student-facing IB AI products, technical disclosures and product documentation. In that review, we did not identify an earlier publicly documented student-facing IB product demonstrating a dedicated language model fine-tuned specifically for IB assessment under the definition above.

This is a limited public-evidence claim, not proof that no earlier private, unpublished or undisclosed system existed, and it is not evidence that iBacalao is more accurate than another product. We no longer use it as the headline description of the model.

If you know of an earlier publicly documented system meeting the same definition, send the source to support@ibacalao.com and we will review the claim.

Independence Notice

iBacalao is an independent product and is not affiliated with, endorsed by, sponsored by or approved by the International Baccalaureate Organization.