Research / Evidence-first engineering

Evidence before narrative.

NH Applied research is organised around working systems, explicit success criteria, preserved faults and a clear distinction between what was demonstrated, what remains provisional and what has not yet been tested.

Research discipline

The aim is not to make an interesting result sound stronger. It is to make the result easier to inspect. Each project therefore tries to preserve the machine state, configuration, measurements, faults, corrective changes and claim boundary that produced the result.

Publication rule: an interesting result is not made stronger by removing its limitations. The limitation is part of the result.

Typical evidence path

Question

Define a falsifiable engineering question and an external success boundary.

Build

Construct the smallest route that can answer it without redefining success.

Measure

Capture results, failures, hashes, environment and relevant telemetry.

Freeze

Separate demonstrated claims from limitations and future work, then preserve the evidence.

Current research areas

Local and on-premise large-model execution on constrained consumer hardware.
User-controlled AI agents, Model Context Protocol infrastructure and safe local tool authority.
Engineering automation, autonomous research tooling and evidence-driven optimisation.
Manufacturing, quality systems and practical engineering workflows.

Publications

Research programmes, papers and supporting evidence.

Primary publications are grouped with their reproducibility and evidence records so the relationship between a result and the material supporting it is explicit.

QES / SINGLE-NODE LARGE-MODEL EXECUTION

QES

Bounded-memory execution and systems characterisation for very large Mixture-of-Experts workloads on a single 128 GB system.

2 public records

Primary publication12 September 2026 · v2.7

QES: Bounded-Memory Execution of Large Mixture-of-Experts Models on a Single 128 GB System — From Qwen3-235B to Kimi K2.5

Main QES technical report covering the bounded-memory architecture, cross-family validation, GPU-authoritative execution and measured Kimi K2.5 result.

View on ZenodoDOI 10.5281/zenodo.22730031

Companion paper14 September 2026 · v1.0

QES Characterization: Data-Path, Cache Locality and Energy in Single-Node Trillion-Parameter MoE Inference

Complementary systems-characterisation paper measuring generated expert traffic, routing locality, cache effectiveness, sustained decode behaviour and package energy.

View on ZenodoDOI 10.5281/zenodo.22755791

ACE K2 / GOVERNED AUTONOMOUS LOCAL AI

ACE K2

Governed autonomous local engineering in which the working model can research, plan and use capabilities while independent controls retain authority over qualification, validation and acceptance.

4 public records

Capability-era technical note31 October 2026 - v1.0

ACE K2 - The Second Major Capability Era

Documents the transition from the September governed-autonomy proof to a long-running governed engineering organisation. Most capability growth is recorded as emerging from governed autonomous operation on real projects, without case-by-case human direction or approval, while ACE's independent evaluator/controller mechanisms retain qualification and terminal acceptance authority.

View on ZenodoDOI 10.5281/zenodo.23836829

ACE K2 project page

Current capabilities white paperSeptember 2026 - v1.2

ACE K2 White Paper v1.2 - Current Capabilities Edition

Published current-capabilities baseline covering governed autonomy, qualification and validation, bounded recovery, cloud escalation and demonstrated local engineering workflows.

View on ZenodoDOI 10.5281/zenodo.22869831

ACE K2 project page

Causal validation report19 October 2026 - v1.0

Microsoft WSL #40988 - AMD ROCm / WSLC Causal Validation

Independent physical A/B validation of AMD ROCm GPU access for WSL containers on Strix Halo gfx1151. With matched WSL-aware ROCm userspace, --gpus all produced successful HIP execution in 3/3 runs while the no-projection control failed HIP initialisation in 3/3. With /dev/dxg present, enabling the DXG backend succeeded in 3/3 while disabling DXG backend detection failed in 3/3. The validated path is /dev/dxg with ROCDXG/HSA/HIP userspace; no Microsoft source patch was required by the tested configuration.

Scope is limited to the tested Strix Halo gfx1151 environment and does not claim universal AMD GPU or workload compatibility.

View on ZenodoDOI 10.5281/zenodo.23469296

Upstream issue #40988

Validation report9 October 2026 - REV 03

ACE Validation of a Source-Level Fix for ROCm Issue #12571: Windows PAL/SDMA Stream-Ordering Fault on gfx1151

Independent source-build and physical causal A/B validation of a durable-compute-event correction for ROCm/rocm-systems #12571 on Windows gfx1151. The pristine build reproduced STALE 50/50; the patched build returned all six test cells clean 50/50 with SDMA enabled. The claim remains bounded to the tested matched stack.

View on ZenodoDOI 10.5281/zenodo.23267518

Upstream issue #12571

AMD STRIX HALO / MIOPEN CONV3D

Windows gfx1151 Conv3d failure investigation

Independent reproduction and isolation of a size-dependent MIOpen Conv3d failure on Ryzen AI MAX+ 395, including a scoped experimental application-level workaround and evidence for the upstream investigation.

1 public record

Technical white paper5 October 2026 · v1.0

MIOpen Conv3d Failure on AMD Strix Halo under Windows: Reproduction, Isolation and a Scoped Workaround

Documents the reduced stage-3 reproduction, tile-256/tile-512 failure boundary, MIOpen-path isolation and an experimental workaround candidate. The mitigation remains unconfirmed pending repeated runs and numerical-equivalence validation.

View on ZenodoDOI 10.5281/zenodo.23158752

Upstream issue #5845

AMD GAIA 0.23 / INDEPENDENT EVALUATION

Persistent local AI agent infrastructure

Independent engineering evaluation of AMD GAIA 0.23.0 on a Windows Ryzen AI MAX+ 395 system, accompanied by reproducibility and public evidence records.

3 public records

Primary technical note14 August 2026 · v1

Evaluating AMD GAIA 0.23 as Infrastructure for Persistent Local AI Agents

Primary independent study of model-slot arbitration, context boundaries, supervised sidecars, crash recovery, local authentication, agent scaffolding, packaging, discovery and rollback.

View on ZenodoDOI 10.5281/zenodo.21935035

Reproducibility pack14 August 2026 · v1

NH Applied GAIA 0.23 Independent Evaluation - Reproducibility Pack v1.0

Sanitized procedures, environment notes, machine-readable results, evidence ledger and compact PowerShell reproductions supporting the independent evaluation.

View on ZenodoDOI 10.5281/zenodo.21935403

Public evidence archive14 August 2026 · v1

NH Applied GAIA 0.23 Independent Evaluation - Public Evidence Archive v1.0

Privacy-redacted evidence archive with original byte counts and SHA-256 hashes preserved in the accompanying ledgers and explicit redaction manifest.

View on ZenodoDOI 10.5281/zenodo.21935521

NH APPLIED MCP / WORKING RESEARCH

User-controlled AI-to-computer infrastructure

Engineering record for the Agent/relay architecture, adversarial qualification and deployment boundaries.

Not yet a public research release