IProgramNoveltyJudge
Interfaces · AiDotNet.Interfaces
Decides whether a candidate program differs meaningfully from the incumbent it most resembles.
Known implementations: LlmProgramNoveltyJudge<T>
For Beginners
Two programs can share almost every word and still do genuinely different things, or look different and be the same idea renamed. When the cheap checks cannot tell, this is what gets asked. Because asking costs a model call, a good gate reaches it rarely.
How It Works
This is the last and most expensive rung of a novelty gate, reached only when the cheap structural check and the embedding check have both failed to separate the two programs. An implementation typically asks a language model, but nothing here requires one: a rules engine, a compiler-based comparison, or a human-in-the-loop queue satisfies the contract equally well, and a test double satisfies it with no I/O at all.
An implementation must return Unavailable rather than throwing when it cannot
produce an answer, and must never let untrusted program text reach a log or an exception message unbounded and
unredacted. Choosing what an unavailable verdict means — admit the candidate or discard it — belongs to the
calling policy, not here.
Properties
| Property | Summary |
|---|---|
string Id { get; } |
Gets a stable judge identifier. |
Methods
| Method | Summary |
|---|---|
ValueTask<ProgramNoveltyVerdict> JudgeAsync(ProgramGenome candidate, ProgramGenome incumbent, CancellationToken cancellationToken) |
Judges a candidate against the incumbent it most resembles. |