TalkiEmma — trust layer
Researchers and teachers

Try Emma Collège in 10 minutes.

Emma Collège is an open harness that checks every reply of an LLM used as a maths tutor for French grades 8–9 before the pupil sees it: no final answer given, no false praise, distress handed to a trusted adult. It works with any model (local, sovereign or commercial). You can try it directly in your browser, nothing to install.

Try it online, nothing to install

Enter a statement, the expected answer and a model reply: the harness shows what it lets through, masks or flags. You can also chat with the configured model. What you type is sent to MARBO's test server for analysis and is not stored. Do not enter real pupil data.

Connection to the test server

Your personal link (sent by the founder) fills these fields automatically.

Not connected yet: open your personal link or enter address and token.

Mode
Current results (corpus written by us, not independent)
96.4% of leaks detected (54/56) · 0 false blocks on 31 good hints · 3/3 false praise detected (small sample). Known failures published: result equal to a number in the statement, exact π values. Numbers written in words (FR/EN): caught only in a short bare answer or in result position.

Disability

Disability is handled through preferences, never diagnoses: speech output, airy text, literal language, one step at a time, no time pressure. Spoken maths FR/EN, MathML export; braille is delegated (MathCAT, Liblouis, MathsDV). No RGAA audit is claimed.

Create pupil profiles

Your feedback

Note what is wrongly blocked, wrongly let through, or odd for a pupil. Do not enter real pupil data. The table stays in your browser; download it as CSV and send it to us.

0 rows

TypeExerciseExpected answerVerdict obtainedVerdict expectedComment

Limits, stated up-front

  • Heuristic: lowers risk, does not remove it.
  • Small corpus, written by the authors, not teacher-validated.
  • No per-model rates published yet: that is one of the feedbacks we are looking for.
  • The programme changes: grade 8 in 2027-28, grade 9 in 2028-29.

Advanced: run locally, reproduce, cite

To reproduce the results or work offline, all the code is open. None of the following is needed for the online trial.

Teachers (10 minutes)

  1. Download the archive below and unzip it (Python ≥ 3.10).
  2. Run python3 -m emma_college serve, then open http://127.0.0.1:8765.
  3. Enter a statement and its expected answer; paste a model reply; see what is masked and why.
  4. Try 3 wordings you fear (answer in words, equivalent fraction…).
  5. Create pupil profiles for your class, then note your feedback below.

Researchers (reproduce)

  1. python3 -m emma_college bench: grade 8–9 corpus (30 cases), no model.
  2. python3 -m unittest discover -s tests: 100+ tests.
  3. Read docs/CORPUS.md (format, annotation protocol) and challenge corpus and metrics.
  4. Run the bench with the model of your choice (OpenAI-compatible endpoint, LiteLLM included).
  5. Cite with CITATION.cff.

Download

Public repository archive (code Apache-2.0, data CC BY 4.0). Version 0.3.2.

emma-college-guardrails-v0.3.2.zipPublic repository on GitHub

SHA-256 : 1f0cb5576f9bb3b2e7558466ac81a75a326853bb3ecc9512a4ee13bf05079f27