Vosti's deterministic LLM inference specification and verification approach gains adoption in agent harnesses for reproducible execution.
state: seedheat: lowuncertainty: mediumconvergesscott: mediumdeterministic-inference llm-verification reproducible-execution agent-harnessesVosti authorsmatt_d
What is this?
Vosti is a new arXiv paper (2609.38981, Oct 1 2026) and research prototype from QDelta that formalizes a system-level specification for deterministic LLM inference — requiring bitwise-identical logits across batching, chunking, prefill/decode splits, and KV-cache reuse. The engine is implemented in Rust, verified with Verus, and uses annotated Triton kernels checked by a relational kernel verifier. The web results show the paper and GitHub repo but no evidence yet of adoption in agent harnesses; the hypothesis projects forward from the paper's relevance to reproducible agent execution.
Why it matters to Scott
Formal-methods researchers are independently building what Scott's Agent Provenance Stack assumes is missing: a machine-checked deterministic execution layer — Vosti's Verus-verified Rust engine with bitwise-identical logits across batching/chunking/KV-reuse would turn his Execution Attestation receipts from records into replayable proofs. Dated-receipts angle if it pans out, but relevance is capped at medium because the supplied evidence shows only a paper and prototype from an unfamiliar group (QDelta) with zero adoption in agent harnesses yet — the hypothesis's adoption claim is projection, not evidence.
ip:concept.execution-attestationip:framework.agent-provenance-stackdev:concept.deterministic-agent-control-planedev:concept.hardware-aware-local-inferenceradar:concept.formal-verificationradar:mi300x-h100-byte-identical-inferenceradar:contract-verifier-llm-gpu-kernelsradar:concept.llm-serving
queries asked of Scott's wikis
- deterministic inference specification verification agent harness
- reproducible execution agent systems kernel verification
- Verus Rust verification LLM serving Triton kernels
- bitwise identical logits batching chunking KV cache agent evaluation
- formal specification LLM inference harness reliability
Measured heat
now 0 pts/hpeak 1 pts/hcomments 0/hpeers p16momentum: steady2 platformsage 67h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
How the heat travelled
pace: p10 vs 1204 stories at the 48h mark (now 67h old) — behind 3jsbench-llm-3d-generation-benchmark (0.5x)
Evidence (2) — ⭐ canonical anchor
Interpretation history
2026-10-09T01:03:51Z
grounded: converges/medium — Formal-methods researchers are independently building what Scott's Agent Provenance Stack assumes is missing: a machine-checked deterministic execution layer —
2026-10-09T00:53:05Z
case created — Arxiv paper proposing framework for specifying and verifying deterministic LLM inference, relevant to agent harness reliability.
Decision trace
- 10-09 13:18attention_routeThe editor compared this story and chose to keep watching.
- 10-09 13:11attention_candidatecreate
- 10-09 12:03groundFormal-methods researchers are independently building what Scott's Agent Provenance Stack assumes is missing: a machine-checked deterministic execution layer — Vosti's Verus-verified Rust en
- 10-09 11:53createArxiv paper proposing framework for specifying and verifying deterministic LLM inference, relevant to agent harness reliability.