# E110 G6: verified cross-domain transfer of learned hidden laws

Preregistered (`G6_PREREGISTRATION.json`, peer-reviewed by the infrastructure session); every number re-derived from raw rows by `verify_e110_g6_offline.py` (stdlib only).
**Gate counts:** {"PASS": 2, "PARTIAL": 1}. Not AGI or VAGI.

Arms at a matched target budget (3 target experiments): A no knowledge · B blind transfer · C verified transfer (the source law as ONE pre-specified hypothesis, exact test, one counterexample refutes) · D learn from scratch on the same 3 experiments. D_k: scratch learning with k experiments; *experiments saved* = smallest k where D_k is no longer worse than C, minus 3.

| id | source → target | A | B | C | D | C won / lost vs A | p | C accepted | experiments saved |
|---|---|---|---|---|---|---|---|---|---|
| P1 | svc:barrier_adj → grid:barrier_adj | 76 | 99 | **99** | 76 | 23 / 0 | 2.4e-07 | 5/5 | 3 |
| P2 | svc:start_row → grid:start_row | 71 | 88 | **88** | 70 | 17 / 0 | 1.5e-05 | 5/5 | 12 |
| P3 | svc:row0 → grid:row0 | 38 | 69 | **69** | 38 | 31 / 0 | 9.3e-10 | 5/5 | 7 |
| P4 | svc:edge → grid:edge | 38 | 73 | **73** | 38 | 35 / 0 | 5.8e-11 | 5/5 | 3 |
| P5 | grid:barrier_adj → svc:barrier_adj | 60 | 68 | **63** | 60 | 3 / 0 | 2.5e-01 | 3/5 | 0 |
| N1 | svc:barrier_adj → grid:uniform | 51 | 39 | **51** | 51 | 0 / 0 | 1.0e+00 | 0/5 | 0 |
| N2 | svc:start_row → grid:uniform | 46 | 43 | **46** | 46 | 0 / 0 | 1.0e+00 | 0/5 | 0 |

**What this shows:** a law learned in one domain, carried as one hypothesis and checked against just three experiments in a new domain, solved far more unfamiliar incidents in 4 of 5 directions, with no losses — and when the law did not hold (N1, N2) the check rejected it every time, avoiding the harm blind transfer caused (N1: 39 vs 51). The reverse direction into the harder service domain (P5) gave no significant gain.

**Cost:** source knowledge came from {'barrier_adj': 250, 'start_row': 250, 'row0': 200, 'edge': 150} E108 practice worlds (service) and 150 grid practice worlds (P5); verified transfer then used 15 target experiments per pathway. Wall time 1779.3 s.

- **G6_primary** PASS: 4/5 pathways significant (Holm) with the block condition: P1 yes, P2 yes, P3 yes, P4 yes, P5 no
- **NEG_negative_transfer** PASS: N1: verified rejected 5/5, C 51 = A 51, blind 39; N2: verified rejected 5/5, C 46 = A 46, blind 43
- **G6_reported** PARTIAL: near transfer between two constructed domains (grid key-door, service lock); the spec's six real verticals are not covered

**Limits:** near transfer between two constructed domains with the same organism; the laws are coordinate laws in designed frames; same-project testing; no external reviewer. CAIN-42 Proto-VAGI is the infrastructure governor of AGI; it is not yet AGI. Team: Vincent Noriega (founder) and a three-person team.
