Robust to Wording, Sensitive to Meaning

Understanding a sentence requires two properties that pull against each other.

The first is stability. "Transition 600 acres with financing support" and "move approximately 600 acres through a transition, with financing needed to make it possible" should yield substantially the same understanding. Wellzai's evaluation harness calls this paraphrase stability, and it is why the harness grades structured meaning rather than exact wording. A related expectation is resistance to irrelevant context: a grandfather who farmed the land for forty years should not change who needs the financing.

The second is sensitivity. "I have 600 acres" and "I need 600 acres" share almost every word and mean nearly opposite things. So do own, lease, manage, want to buy, could transition and can make available. The harness organises its cases into families of such sentences precisely because they expose superficial keyword matching, and it requires that small wording changes which alter meaning must alter the result.

Keyword matching fails both tests at once. It treats different words for the same thing as different, and the same words in different roles as the same.

The implication is that neither property can be tested on its own. A system that is only stable may be ignoring the words; a system that is only sensitive may simply be reacting to them.

The open question concerns the tests themselves. Every family of contrasts is finite, and a system can learn the families rather than the meaning. Wellzai's semantic architecture is designed to answer part of this with blind holdouts, sentences that never shaped the implementation, but a holdout is still a sample. Can one test suite demand both stability and sensitivity without rewarding a different brittleness, an interpreter tuned to the test's own sentences?