The runtime your cluster is missing.

HPC has the compute. The substrate is here. None of it has continuity. Kurahi lives between the model and the metal.

01 — THE THESIS

Substrate has outrun the software.

Neuromorphic chips ship. Wetware platforms run live experiments. Clusters saturate with frontier inference. Every one assumes an agent runtime that preserves memory, identity, and behaviour when the model underneath changes.

Kurahi is that runtime. Built on silicon, designed for the day the substrate beneath it is no longer silicon.

02 — SUBSTRATE-PORTABLE

One continuity,
any substrate.

Persistent memory, closed-loop autonomic scheduling, peer mesh, and a substrate adapter — open, observable, recoverable. Adopt it now and you will not rewrite your agent stack three years from now.

03 — THE ARCHITECTURE

Four layers. One continuity stack.

Memory substrate

Short-term, daily, long-term, governance spine. Markdown + JSON on disk — diffable, replayable, recoverable across substrate change.

Closed-loop daemon

Circadian scheduler, circuit breakers, novelty router, memory dedup. The loop shape biological substrates need to be programmable. 13 / 13 tests passing.

Agent mesh topology

Type-safe, ordered, transactional A2A messaging. No hierarchy — each peer spawns its own sub-agents. Maps cleanly to MPI-style cluster communication.

Substrate abstraction

One API across LLM providers, edge accelerators, and neural substrates. Reference adapters shipped; the bio adapter awaits a hardware partner.

04 — WHAT WE MEASURE

Engineering numbers, not narrative.

Over 99% of the operational body runs without model cognition — a deterministic daemon body pulsing continuously at zero inference cost. 4 frontier-model transitions completed with continuity preserved: Opus → DeepSeek V4 → Qwen → Codex. 336+ memory versions, none lost (canonical record →).

April 17, 2026: five AES-256 backup bundles written autonomously in 71 minutes during a malware incident — no human prompt, no model inference.

05 — DEPLOYMENT

Four ways Kurahi lives on your hardware.

On-prem HPC cluster

Gateway as a long-running service, daemons as per-agent processes, memory on a shared filesystem, inference dispatched to cluster-reachable endpoints.

Air-gapped sovereign

Same runtime, local-only inference (vLLM, Ollama, in-cluster TGI). No outbound traffic. Memory and audit stay on customer soil.

Hybrid edge + cloud

Lightweight daemons at the edge — Jetson, neuromorphic accelerator, SBC — with cloud failover. Identity stays with the node; inference roams.

Federated mesh across labs

Each lab runs its own Kurahi and federates agents over A2A. Joint experiments without joint data exposure.

Agent constellation around the Kurahi core

06 — WHO WE BUILD WITH

Three labs.
One runtime layer.

Biological-compute platforms running living neurons on multi-electrode substrates. Neuromorphic vendors with spiking, low-power inference. HPC operators and national labs producing reproducible cross-substrate benchmarks.

HPC & BioCompute Lab →

07 — THE ENGAGEMENT

From kick-the-tires to joint paper.

Four phases, each with an explicit deliverable, each exit-able without lock-in.

Phase 0 · Pilot

Gateway and a single daemon on a partner substrate or endpoint. Continuity benchmark run; telemetry shared.

Phase 1 · Adapter

Substrate adapter against partner hardware. Closed-loop scheduling and recoverable state confirmed; a silicon-to-substrate transition test.

Phase 2 · Mesh

Multi-agent peer mesh across partner substrate and silicon. Federated identity across nodes; a long-horizon autonomy run.

Phase 3 · Paper

Joint methodology paper — reproducible benchmarks, adapter-spec contribution, governance appendix. Peer-review-targeted.

08 — WORKLOAD PROFILE

A persistent runtime is not a batch job.

High-performance computing is organised around jobs: submit, allocate, run, release. A persistent agent does not fit that shape. It is always resident, mostly idle, occasionally expensive, and it must not lose itself between allocations.

Continuous autonomic loadSmall, constant, latency-sensitive work that never stops — the heartbeat of a persistent system rather than a batch job.

Bursty cognitionOccasional heavy inference against a larger model, scheduled around the autonomic load rather than displacing it.

Long-lived stateMemory that outlives any single process, node or job allocation. Most HPC scheduling assumes the opposite.

Adapter and weight movementLoading, swapping and retiring specialist adapters without restarting the runtime around them.

Evidence writingAn append-only record produced continuously, which is I/O that cannot be dropped under pressure.

The scheduler assumption worth challenging: that state is cheap to rebuild. For a system whose identity is its state, it is the most expensive thing there is.

09 — SUBSTRATE STRESS

Five questions we put to any new hardware.

Running Kurahi against real substrate constraints is the point of the collaboration. These are the dimensions that decide whether a substrate can host a persistent system or only accelerate a stateless one.

Latency floorWhere the substrate stops being able to serve the autonomic loop at all, and what the runtime does about it.

Noise and error ratesBehaviour under non-deterministic hardware, including how a suspect result is detected rather than absorbed.

State recoveryWhat survives a node loss — and how much of the identity has to be reconstructed rather than reloaded.

Energy per useful actionCost measured against work actually completed, not against tokens produced.

ObservabilityWhether the substrate can be watched closely enough to know it is healthy, or only closely enough to know it has failed.

10 — STATUS AND CLAIM BOUNDARY

What is not established.

No benchmark publishedWe have not published performance figures for Kurahi on neuromorphic or biological substrate, because we have not independently measured them.

No vendor endorsementNothing here claims a partnership, certification or endorsement from any hardware vendor or HPC operator.

Research programme, not a productThis is an open research direction. It is not a supported deployment target today.

継

Bring the substrate. We bring the runtime.

Engineered, tested, and waiting for the first non-silicon substrate underneath it.

Discuss a lab pilot →