20 Trustworthy Foundation Models

Odyssey and Prometheus

Foundation models are usually identified with the parameters of a large pretrained network. That identification is too narrow for systems expected to support research, decisions, and persistent institutional knowledge. The model may generate a fluent candidate answer, but the answer also depends on sources, retrieval, tools, intermediate artifacts, domain assumptions, cross-document joins, and rules governing what may be retained. Trust therefore cannot be a property of the neural model alone.

Kimi K3 makes this boundary concrete. Releasing a 2.8-trillion-parameter open-weight model does not by itself make the model operational: sparse expert routing, hybrid long-context attention, low-precision realization, inference kernels, and distributed execution are part of the usable system [ Kimi Team , 2026 ] . Chapter 12 treats these interacting layers as a frontier-scale example of architectural obligations and repair. The weights are indispensable, but they are not the whole maintained artifact.

This chapter develops a complementary view. A trustworthy foundation model is a maintained foundry: a typed system of local representations whose claims can be traced to evidence, compared on overlaps, repaired when incompatible, and admitted into durable state only under an explicit contract. Odyssey supplies the governance and admission architecture for such foundries. Prometheus supplies a deep-research engine that constructs and inspects candidate world-model artifacts. Their division of labor realizes the LINCS workflow at system scale [ Mahadevan , 2026j , 2026k ] .

In this setting the distinction between a model of the world and a model of learning becomes operational. Prometheus may construct a biological, financial, or scientific world model whose mechanisms support domain questions. Odyssey models the process by which such artifacts are sourced, joined, challenged, revised, and promoted. LINCS makes that second-order system actionable: one can intervene on an extractor, cover, argument rule, refresh policy, or admission block while preserving the unaffected foundry obligations. Such an edit is structural surgery on the learning system, not automatically a causal intervention in the represented world.

A generated artifact is a proposal. It becomes maintained knowledge only after its provenance, restrictions, overlaps, arguments, and intended use have survived independent admission.

¹1. The phrase truth-preserving is deliberately local. It means preserving declared evidence and compatibility conditions under transport. It does not mean that the system possesses an oracle for external truth. ↩