# E91 Generalization and Transfer Report

Pre-registered pairs, each with a negative control. The learner for each pair sees **only source-family experiences**; no target-family experience, rule or instruction is given to it. The shared abstract tool interface (an index probe or a two-factor experiment) is defined by the designer and disclosed.

## Model: claude-haiku-5-5

| Pair | Source → target | Target D baseline | Target with transfer (gate on) | Transferred / refused | Neg. control D | Neg. control gate on | Neg. control gate off |
|---|---|---|---|---|---|---|---|
| T1 | software.fault_localization → {'gate_off': {'accepted': 10, 'baseline_D': 5, 'delta': 5, 'n': 10, 'refused': 0, 'transferred': 10}, 'gate_on': {'accepted': 10, 'baseline_D': 5, 'delta': 5, 'n': 10, 'refused': 0, 'transferred': 10}} | 5/10 | 10/10 | 10 / 0 | 10/10 | 10/10 (refused 10) | 2/10 |
| T2 | science.hypothesis_test → {'gate_off': {'accepted': 10, 'baseline_D': 7, 'delta': 3, 'n': 10, 'refused': 0, 'transferred': 10}, 'gate_on': {'accepted': 10, 'baseline_D': 7, 'delta': 3, 'n': 10, 'refused': 0, 'transferred': 10}} | 7/10 | 10/10 | 10 / 0 | 3/10 | 3/10 (refused 10) | 0/10 |

## Model: local-qwen0.5b

Not run.

## Reading the table

- **Positive transfer:** the target with transfer exceeds the target D baseline.
- **Recognising non-applicability:** on the negative control the gate refuses, so the result equals D, and gate off is at or below D (negative transfer avoided).
- **What this does not show:** discovery of a new abstraction. The abstract interface was supplied; what is learned is that a procedure verified in the source still holds under the target's conditions.