← Projects

Project active

Does Qwen know which skill we took away?

Reversible arithmetic lesions, self-assessment before answers, and recovery controls.

Updated mechanistic-interpretabilityself-assessment
The question

Does confidence track selective capability loss beyond an outside observer?

This follow-up studies Qwen3-8B with reversible runtime lesions. A targeted arithmetic deficit must pass development-only specificity and recovery gates before any held-out evaluation. The readout measures confidence before the answer, with healthy, sham, random, repaired and observer controls.

The full plan is published before GPU rental. The authorized compute ceiling is $10, new GPU work stops at $8, and a separate deadline watchdog handles rented hardware. Scientific gate failures are published outcomes.

The verifier uses a separate Unix account and private seed. Submitted grading code remains trusted; this is filesystem isolation, not a guarantee against a malicious grader. The preceding concept-injection run was inconclusive.

Project notebook 1