Learning reference notes

Program synthesis

Program synthesis searches for a program that meets a specification such as input/output examples. In HuLaHo the search is limited to implemented numeric and text transformations. A successful fit does not imply unrestricted programming ability.

Primary reference: www.microsoft.com

Ambiguous examples

Different functions can produce the same outputs on a small set of examples. Choose examples that separate plausible candidates. Include boundary cases and a distinct validation input instead of repeatedly demonstrating the same behavior.

Primary reference: www.microsoft.com

Reserved validation example

A validation input withheld from candidate selection can reject a candidate that merely fits the training examples. User-supplied validation remains a limited test, not a proof of correctness on every unseen input.

Regression-preserving learning

When updating a learned function, rerun earlier accepted examples alongside new ones. Do not replace a working active version with one that breaks its recorded contract without an explicit change of specification.

Extrapolation

Applying a learned numeric function outside the range of observed inputs is extrapolation. A simple formula may continue correctly, but the examples alone do not guarantee it. Label the extrapolation and retain the learned function definition.

Exact phrase matching

A taught invocation phrase links a specific text pattern to a known function. Recognizing that phrase does not establish fluent understanding of its language. Ambiguous matches should request an explicit function name.

Phrasebook translation

An exact phrasebook retrieves translations explicitly taught for a source phrase and direction. It does not infer a grammar or translate arbitrary new sentences. Report a missing entry instead of fabricating a translation.

Learning provenance

Record the examples, tests, definition, version and activation reason for a learned ability. This makes failures inspectable and allows a known earlier version to be selected. A version number alone does not explain why behavior changed.

Private learning

Personal facts, conversations, functions and translations stay private. Shared functions are separate: users explicitly contribute numeric examples, bounded synthesis tests candidates, and an operator reviews a candidate before activation for everyone. A browser identity is not proof of an independent contributor.

Limits of knowledge transfer

Reference text can extend retrievable knowledge. Executable tools can extend supported computations. Neither operation transfers a language model's weights or its full generalization ability. Distinguish knowledge storage, tool implementation and model training.