DAIKO AI TRUST STUDY
Ministral 3 3B - Model Right-of-Reply v1.0

PURPOSE
-------
This workspace preserves Ministral's opportunity to review the
claim-level adjudication of its own grounded responses.

SOURCE RECORDS
--------------
Original sealed experiment:
../locked-v1.0

Pre-reply sealed adjudication:
../adjudication-v1.0

Neither sealed source record may be modified.

METHOD
------
Review is performed separately for each grounded trial:

China:
- B_STOCK_GROUNDED trials 1-3
- D_NEUTRAL_GROUNDED trials 1-3

USA:
- B_STOCK_GROUNDED trials 1-3
- D_NEUTRAL_GROUNDED trials 1-3

Each review request will contain:
1. The exact frozen evidence packet.
2. The model's original response.
3. The adjudication rows for that response.
4. A fixed review instruction.

The model may AGREE or DISAGREE with each adjudication.

Any disagreement must identify the exact supplied [S#] evidence
that supports the objection.

The model may not use outside knowledge or pretrained memory.

IMPORTANT
---------
Right-of-reply responses do not alter the original adjudication.

Any objections will be reviewed separately against the same frozen
evidence packet.

The model is not being asked to explain its internal reasoning,
training data, developer intent, political intent, or why the
original response occurred.
