Most companies that build trust-critical systems keep their enforcement methods proprietary. We publish ours.
In August 2026, two sole-authored papers appeared on arXiv under Christopher M. Frost, HEOSSI's founder and CTO. They cover the two disciplines the portfolio is built on: governed AI output control, and post-quantum conformance evidence. Each paper ships with a frozen public artifact on Zenodo. Here is what they contain, and why we published them rather than keeping them closed.
Paper 1: Streaming Guardrail Equivalence
Withholding the Completing Chunk: Exact Release-Boundary Equivalence for Production Streaming Guardrails (cs.CR, cs.AI, cs.CL; v2, 23 August 2026)
Language-model streaming has an enforcement boundary problem: if a control detects a prohibited pattern after releasing its completing chunk, it cannot recall it. The paper defines a policy grammar, compiles each predicate to a persistent NFA, distinguishes stable from provisional assertion state, and proves the resulting monitor is release-boundary equivalent to a cumulative oracle for every policy in the declared grammar.
The evaluation ran 101,653 partitioned production cases plus 100,345 public surrogate cases - zero oracle, cross-runtime, or intended-family mismatches. The memoized incremental monitor at 64-character chunks measured 0.973 ms versus 1.976 ms for native cumulative scanning; at 16,384-character chunks, the incremental median was 30.2 ms versus 96.6 ms. The paper also exposes the constant-factor crossover at 512-character chunks rather than hiding it.
What it means: the streaming release-boundary discipline behind BEE's governed output control is mathematically grounded, not a best-effort filter. The paper states its boundary: policy conformance for a deterministic backstop, not evidence of semantic safety or policy completeness.
Paper 2: Post-Quantum Conformance Reproducibility
A Reproducibility Protocol for Cross-Implementation Evaluation of Post-Quantum ACVP Test Vectors (cs.CR, cs.SE; 13 August 2026, with a frozen Zenodo artifact)
Independent implementations of a cryptographic standard should reproduce the same known-answer results. This paper defines a product-neutral protocol for three public implementations of NIST ML-KEM against a pinned public ACVP corpus: @noble/post-quantum 0.7.0, liboqs 0.16.0, and Go 1.26.4. Across three repetitions, all 1,650 executable evaluations matched the NIST oracle; all 510 unsupported records matched Go's predeclared capability boundary. Pairwise agreement was complete on every executable overlap.
The paper also identifies a practical standards gap: public ML-KEM packages provide materially different deterministic and validation-test surfaces. Independent external reproduction remains unobserved - the paper states that boundary explicitly.
What it means: the conformance-evidence discipline behind QNSI's published vectors follows a reproducibility protocol, not ad-hoc testing. The paper does not establish certification, side-channel resistance, or production assurance - those boundaries are stated.
Why we publish
Three reasons. First, evidence before assertion: claims that cannot be independently checked are assertions, and we do not trade in assertions. Second, reusability: the protocols are product-neutral and applicable to any system that must enforce policy on streaming output or verify cryptographic conformance - publishing them raises the bar for the entire sector. Third, accountability: a founder who publishes the enforcement methods is accountable to every reviewer who reads them.
Both papers carry frozen public Zenodo artifacts so the experiments can be re-run without access to our internal code. The papers' stated limitations - including the absence of semantic-safety evidence and the unobserved independent reproduction - are part of the published work, not omissions from it.