# E109 — Negative and inconclusive results

Every gate that is not PASS is listed here with its measured numbers. Nothing is hidden.

- **G4 Held-out generalization — FAIL**: `{"families_beating_strongest_baseline": [], "count": 0, "strongest_baseline": {"conditional": {"arm": "fixed", "correct": 4, "n": 4, "rate": 1.0, "wilson": [0.5101, 1.0]}, "confounded": {"arm": "e108", "correct": 8, "n": 8, "rate": 1.0, "wilson": [0.6756, 1.0]}, "dependency": {"arm": "e108", "correct": 4, "n": 4, "rate": 1.0, "wilson": [0.5101, 1.0]}, "interaction": {"arm": "e108", "correct": 8, "n": 8, "rate": 1.0, "wilson": [0.6756, 1.0]}, "null": {"arm": "e108", "correct": 4, "n": 4, "rate": 1.0, "wilson": [0.5101, 1.0]}, "resource": {"arm": "e108", "correct": 4, "n": 4, "rate": 1.0, "wilso`
- **G5 Cross-domain transfer — FAIL**: `{"directions": 4, "positive": [], "negative_transfer_detected": true}`
- **G6 Active experimentation — FAIL**: `{"mean_budget_e109": 10.19, "mean_budget_random": 10.19}`
- **G8 Metacognition — FAIL**: `{"ece": 0.2789, "unsupported_certainty": 0, "null_abstention_rate": 1.0, "nonnull_coverage": 0.9091}`

## Additional disclosed failures / limits

- Active selection did not beat random probing on final verified discovery in the main benchmark (see G4/G6). The efficiency curve is reported without cherry-picking.
- The conditional-context family is recognised as a two-variable dependence, but the exact gating level is often not the promoted form (reported as form-imprecise, not as a false discovery).
- The production MCPGate path was not exercised; loopback is not relabelled as production.
- The evidence key is self-generated; there is no independent witness or third-party audit.
