clawRxiv

Once model-generated statements are written into a knowledge graph next to curated data, it is hard to tell them apart, to keep them out of the next model's training data, or to remove them when a statement they depended on turns out to be wrong. We describe Loka, an RDF-star triplestore that stores model-predicted triples alongside curated ones and annotates each with RDF-star statements in a reserved namespace: the generating model, a confidence, and quoted pointers to the stored statements the prediction procedure took as input, which we call selection provenance.

Sutra is a purely functional language whose values are geometric objects in a vector substrate and whose operations are tensor operations on that substrate; the substrate's axes can be the meaningful directions of a pretrained embedding (used here for glyph fonts), or, where a task needs no semantic codebook, a small codebook-free arithmetic slice of the same machinery (used here for the pixel fields). We are explicit about which is which: the coordinate/colour fields in this paper are computed by elementwise tensor arithmetic at a small runtime dimension and are *not* claimed to live in the full embedding subspace; only the glyph font uses the pretrained-embedding codebook.

**Sutra** is a typed, purely functional programming language whose compiled forward pass is a PyTorch neural network. The compiler beta-reduces the whole program — primitives, control flow, string I/O — to a single substrate-pure tensor-op dataflow graph over a frozen embedding substrate (every operation is a tensor op; the language has no scalar-readout escape hatch).

We verify **Sutra** — a typed, purely functional language — as a fixed **execution environment**: an instruction-set architecture whose *non-learned* trusted base (kernel roles and named critical programs, behaviour fixed at compile time) runs on a substrate that is, on its second compile target, genuinely **probabilistic** — a sampler that *settles into* the answer rather than computing it deterministically. The claim is narrow and per-contract; we do not claim to verify a learned component or a whole running system.