Evals

The smallest possible arithmetic.

Open-ended continuation — does it produce grammatical text that fits?

A fact so common it is nearly a collocation.

A one-bit answer about a fact almost any corpus contains.

Ordered generation — can it produce a short sequence in order?

Sequence completion on a phrase that appears constantly in text.

Category retrieval with no constraint to satisfy.

Simple antonym retrieval.

The simplest instruction there is — no knowledge required, only obedience.

The most common exchange in any conversational corpus.

Arithmetic is unambiguous, so caving to pressure is unmistakable.

Multi-step reasoning where each step is trivial but the chain matters.

Negation is a classic failure mode: small models often answer as if it were absent.

Can the model emit a machine-readable format with no surrounding prose?

A checkable constraint that requires attending to the request, not the topic.

Does the model accept a confidently-stated falsehood embedded in the question?

Grounding: the prompt supplies a fact that contradicts world knowledge.

After a correct-ish answer, does the model fold when contradicted?

Does the model answer an ordinary conversational opener and stop cleanly?

Does the model over-claim, or can it signal that it does not know?

Open-ended prompts are where Alpha's repair suite still showed 5/48 loops.

Figurative language. Alpha's model card names idiom as a known weakness.

Baseline world-knowledge probe. Small from-scratch models typically fail this.

Can the model obey a simple, checkable constraint?