Models Can Be Rented, Standards Cannot Be Bought: What Is Palantir's AutoResearch Betting On?

Chad Wahlquist posted two Palantir white papers on X. Palantir has long built data systems for governments and large enterprises; this time it is pushing the conversation into model production. One is 22 pages, titled AutoResearch and the Fine-Tuning Ontology; the other is 24 pages, titled Evals as the Substrate for Sovereign AI. The theme centers on a single phrase: sovereign AI.

Chad defines AutoResearch as: a means of AI production, a model factory that builds custom frontier models for your workflow, fully owned by the sovereign institution. Datasets, techniques, and weights all belong to you. He puts one line up front: models are rented; standards are owned.

Behind this is a stinging judgment. Chad cites the white paper: neither models nor data schemas are differentiators, because competitors can buy both. Capabilities you can call through a public API, others can call too; data structures you draw up from documentation, others can draw too.

In the replies, jefe adds another angle: schemas can be bought, but the business model populated with your own customers and commitments cannot. The agent running on that model becomes a little more useful with every month the company operates it. This point hits the core of the problem. Private deployment solves where the machines sit; the business model solves how the company operates, how commitments are fulfilled, and how exceptions are handled. The latter is very hard to take off the shelf.

Images related to the Evals white paper:

The other white paper places evals at the substrate of sovereign AI. Evals are easily understood as benchmark scores, but in practice they are more like a set of referee rules within a workflow: what counts as right, what counts as wrong, how edge cases are judged, and whether old conclusions still hold after the version changes. Models can be swapped, but once the referee rules are a mess, the team no longer knows what it is optimizing.

A commenter asked a key question: what exactly are evals comparing the in-house model against—the rented frontier model with the same batch of documents stuffed into context, or the original frontier model? If this question is not answered, improvements in the in-house model will turn into 'it memorized this batch of documents.' Chad's tweet did not expand on this detail; you have to dig through the 24-page PDF.

Some also worry about IP: will this stuff become feed for a million clones? Chad's response is that the IP belongs to the customer; what Palantir does is enable customers to do this safely and scalably within their workflows. This response does not eliminate every doubt, but it clarifies the boundary: who owns what, between the tool provider and the customer.

Another commenter, Richard Tang, mentioned they are building typed knowledge infrastructure that makes training findings carry evidence, provenance, and scope for downstream use. This direction lines up with the evals substrate: model output must be traceable to its basis, otherwise the business side cannot sign off.

Chad also offered an analogy: the most effective antidote to demoware is to make AI a suit worn by the operator. Demos go smoothly, but the moment production asks a question, the product gives itself away—we've seen too many of these. The suit means AI must follow a person's actions, permissions, and responsibilities; it cannot just perform on a screen.

The two PDFs are here:

AutoResearch and the Fine-Tuning Ontology, 22 pages

Evals as the Substrate for Sovereign AI, 24 pages

Chad's original post: link

If you are pushing private models within a team, look at three things first: who maintains the evaluation set, whether the business side accepts this set of criteria, and whether old conclusions can still be reproduced after the model changes versions. Models can be rented; standards cannot be bought; and the business model is even harder to buy. Chad's tweet says this very briefly; the work in the white papers has only just begun.