The test he hadn’t seen before.

Frozen primary examiner

The v111.3 Blind Examiner was not edited. Its hidden semantic programs, realiser and scorer remained frozen. The new candidate saw only each current prompt and learner-visible entity descriptions; examination caused zero model changes.

The historical pre-v111.4 seed-771113 language result was 5/17. The replicated post-development range is 13–14/17.

Independent frozen hand-authored set

This second set is defined in a separate module that imports neither the Mentor nor learner internals. It contains 40 fixed cases and was not used for teaching.

Controls scored 13/40 for text-only and 6/40 for shuffled context.

Failure categories

Full-arm failures comprised:

• 9 semantic-role/generalisation failures;
• 6 insufficient-language-evidence results;
• 1 negation-scope failure;
• 2 unjustified-certainty failures.

The principal boundary is not lexical recall. It is deciding role orientation and knowing when weak overlapping evidence is insufficient.

Leakage and mutation checks

• Examiner modules do not import developmental Mentor modules.
• Mentor source does not import either blind examiner.
• Exact blind prompts and hidden expected structures are never passed to development.
• Blind candidate fingerprints were identical before and after examination.
• The primary comparison continued to use the unchanged v111.3 examiner.

Classification

PARTIAL DEVELOPMENTAL TRANSFER PROVEN IN CONTROLLED OWEN WORLD.

Not proven: human English mastery, open-world linguistic understanding, calibrated ambiguity, robust unfamiliar pragmatics, or general noisy-speech competence.