Short pieces from the people building Backplain — for the buyer who has to defend the decision, not just sign the PO.
Ask ten frontier models the population of Tokyo and you get two clusters of answers, both defensible. The variance is not noise — it is the most useful signal you're not yet looking at, and the reason single-model workflows quietly under-perform on the work that matters.
Not a regex layer. Not an API gateway. A stateful substitution-and-reconstitution system that lets a model do real work on real documents without ever seeing the parts it isn't allowed to keep. Three modes, one boundary, complete log.
Harvey is built for Cravath — $1,000/seat, 20-seat floor, $288K/yr to start. If you are a three-person in-house team or a 45-attorney firm, the right answer is not a smaller Harvey. It is a different category of tool.
A two-page interview script for the in-house GC who needs to know what their team is actually doing in ChatGPT. Six questions, three follow-ups, and a one-page summary the CEO can read.
When each deployment shape is the right answer, when it isn't, and the surprisingly small set of organizations for which the question is even worth asking. With a decision tree.
Most AI vendors rent compute. We don't, for the workloads where renting forecloses the answer the customer actually needs. A short essay on physical sovereignty, Tier 3 colos, and what it means for ITAR, CUI, and pre-disclosure biotech IP.