# CAIN-2027 Evolution 105: learning where hidden preconditions are — learned, not yet useful

Preregistered before the sealed runs (development findings disclosed in the registration). Gate counts: {'PASS': 4, 'PARTIAL': 0, 'FAIL': 2}. Not AGI or VAGI. No external reviewer.

| test | status | evidence |
|---|---|---|
| Learn WHERE a precondition tends to be (edge-lock) | **PASS** | after 44 solved practice incidents the lift prior concentrated 33.0% of its lift on the drain column (10% of cells); 4 of the top 5 cells are on it |
| H1 use it to solve more (edge-lock, 100 sealed) | **FAIL** | 24 vs 27 of 100 solved with vs without the prior (McNemar p = 0.38) |
| H2 guardrail: no false pattern on uniform-key worlds | **FAIL** | the lift gate reported a usable prior (max lift 2.51) on worlds whose key is uniform: a false positive; the fixed 2.0 threshold needs a test against a shuffled null |
| H2 no loss on uniform-key worlds (60 sealed) | **PASS** | 25 vs 26 of 60 solved with vs without the prior (McNemar p = 1) |
| H3 no loss on other families (80 sealed) | **PASS** | 72 vs 72 of 80 solved with vs without the prior (McNemar p = 1) |
| Development findings disclosed before registration | **PASS** | raw-count prior hallucinated a pattern and hurt (10 to 7 of 16); fixed by lift before the sealed run |

## Reading
- **Knowledge was learned from experience.** Without any code that knows where the drain is, the organism's own credit assignment found the drain column.
- **It did not become better decisions.** Steering exploration toward that column did not raise success on 100 sealed incidents; the bottleneck there is the whole chain (labels, drain, lock, hidden target) within the budget, not only where to look first.
- **The guardrail needs a real null test.** A fixed lift threshold passed on uniform-key worlds once there was more data. The next version must compare against a shuffled-attribution null (a permutation test) before using any prior. It did no harm here.

Verify: `python3 -I verify_e105_offline.py .` (stdlib only). Team: Vincent Noriega (founder), Lena Noriega, Monica Diaz.
