| title | About |
|---|
ising-harness is a personal exploration by Ken Rollins, Chief AI
Technology Strategist in Dell Federal, into a specific architectural question:
What does the learning layer of an agent system look like — the part that makes agents get measurably smarter across engagements without retraining the underlying model?
The first project in this exploration was gemma-forge, which built a Ralph-loop + reflexion + verification harness on top of Gemma 4 31B and demonstrated it on STIG compliance remediation (90.3% PASS rate, four architecturally-distinct skills exercised). gemma-forge proved the pattern works on bounded, recurring-rule-shaped problems. It also exposed where the pattern struggles — pentest's structural failure in gemma-forge's Phase 4 (entry 41 in that project's journal) made the limits visible.
ising-harness is the second look. The framing pivots from "small model + right
harness" (gemma-forge's thesis, demonstrable on STIG, brittle on harder task
shapes) to "the cross-run causal-learning layer, built on whatever orchestration
stack the customer already trusts." It is intentionally less about model size and
more about the architectural primitives that make agent memory useful across runs:
- JSONL event-stream as canonical source of truth. Every action, decision, model output, observation, GPU snapshot recorded as a structured event. Not a log; the system's actual record of behavior. This is the substrate that makes everything else honest.
- Mechanism-bearing tips. Every learned heuristic must carry a
mechanismfield explaining the causal WHY. The distinction between "approach X failed" (a negative example, brittle, doesn't transfer) and "approach X failed because Y mechanism" (a transferable insight) is the difference between a memory system and a stamp collection. - Dream-pass causal attribution. End-of-engagement pass that, per retrieved tip, judges whether the agent's action substantively followed the tip's advice and what outcome resulted — then retires tips whose outcome signals don't justify their continued retrieval.
- Cross-run measurement protocol. Same harness configuration, different random samples of work items, before/after metrics. Makes "the agent is learning" testable rather than rhetorical.
NVIDIA released Ising in April 2026 — a family of open-source AI models for quantum control (a 35B-A3B vision-language model for qubit calibration, two small 3D CNNs for surface-code error decoding). NVIDIA also released the NeMo Agent Toolkit with memory, observability, and multi-framework support. Federal labs (Fermilab, LBNL, UK National Physical Laboratory) are already adopting Ising. The perception models exist. The orchestration toolkit exists. The missing piece is the learning layer — the four primitives above, applied to a domain (quantum calibration) where they have a clean fit.
ising-harness is that learning layer.
The JSONL event stream the four primitives sit on top of — built first for demo replayability — turned out to be a structural fit for federal agentic-AI compliance. NIST AI RMF, OMB M-24-10, NIST AI 600-1 and the direction-of-travel on agentic AI converge on a single requirement: traceable, auditable, attributable decision trails as architectural commitment, not as a deploy-time add-on.
ADR-0002 commits the JSONL substrate. ADR-0007 hardens it with a SHA-256 hash chain on every event, so tampering is detectable post-hoc, and codifies the project rule that anything observable or important goes on the stream rather than into logs. The result, in product terms:
| Requirement | How the architecture satisfies it |
|---|---|
| Every action recorded | JSONL append, fsync per event, schema-validated at write time |
| Every decision attributable | Mechanism-bearing tips + dream-pass causal attribution |
| Every model output preserved | data + model_call_metrics carry the full call substance |
| Replayable post-hoc | Single-file source of truth, walkable at any speed |
| Tamper-evident | prev_hash + event_hash per row; verify_chain flags any break |
| Open format, no vendor lock-in | JSONL is universal; jq, pandas, anything reads it |
| Customer-portable | Open source on GitHub; federal customer can run, audit, and extract |
This is not a marketing position. It is a named consequence of two ADRs. The reference materials are on GitHub for customers who want to implement the same pattern themselves.
- Not an official Dell product. There is nothing to buy. Nothing here is for sale.
- Not an NVIDIA product. This project extends NVIDIA's open-source models (Ising) and toolkit (NeMo Agent Toolkit) under their respective licenses. NVIDIA has not produced, endorsed, or reviewed this work.
- Not a Dell reference architecture. This work has not been produced, endorsed, or reviewed through any official Dell channel. Views, technical findings, and opinions represented here are the author's own and do not represent an official Dell position.
- Not a commercial agentic-AI framework.
ising-harnessdeliberately layers on open components so the whole thing can be read, reasoned about, and reused. It is reference material, not a platform. - Not a quantum-computing breakthrough. The quantum-control domain provides a clean setting for the learning-layer primitives. The contribution is architectural; the science is NVIDIA's, IBM's, IonQ's, et al.
- Not a benchmark. Numbers in the journal entries describe specific measured outcomes on specific hardware in specific configurations. They are honest, but they are not normalized comparisons across platforms.
Most agentic-AI demos are beautiful when they work and have nothing useful to say when they don't. They show the happy path, skip the recovery, and hide the source so no one can learn from the build.
gemma-forge was deliberately the opposite — every failure mode documented, every
architectural decision in a journal entry. ising-harness carries that discipline
forward but starts cleaner. It does not port gemma-forge's code; it ports the
four primitives that earned their place in gemma-forge's 41-entry journal, and
builds them fresh on top of NVIDIA's stack.
The exploration is still the product. The skill (quantum control) is the witness. The goal is the same: enable other engineers to build similar systems faster and with less surprise.
Dell hardware is referenced throughout because it is what the author works with day-to-day at Dell Federal. The techniques described apply to any platform.
Unlike gemma-forge, which was edge-first (XR7620 + 4×L4, local inference), this project is NVIDIA-Cloud-first for development, with local deployment as a secondary path:
- Primary inference: NVIDIA API (
integrate.api.nvidia.com) — Ising Calibration, orchestration models (Llama 4 Maverick, Nemotron, Mistral Large 3, etc.), embeddings. - Local fallback for sovereign-on-prem demonstrations: NIM deployment on A6000-class GPUs (Ising Calibration's minimum spec) or XR7620 + 4×L4 via tensor parallelism (untested at project start; see ADR-0005).
- Simulator: NVIDIA CUDA-Q for qubit-array simulation with realistic noise models. Runs on cloud GPU instances or local hardware.
This is a deliberate pivot from gemma-forge's edge-first stance. The edge story remains a deployment option, not the product's foundation.
ising-harness, like gemma-forge, is built in an agentic coding workflow — a
human operator paired with an AI coding assistant. The human makes every
architectural and strategic decision. The AI contributes implementation
velocity, test coverage, and documentation drafting. The journal — written in
real time as discoveries happen — is the canonical record of both what the
system did and what the human/AI pair decided.
ising-harness is released under the Apache License 2.0, matching the
license of NVIDIA Ising and NeMo Agent Toolkit. See LICENSE in the
repository root for the full text.
This is a personal project maintained by the author. Questions, suggestions, and technical discussion are welcome through GitHub issues once the repo is public. For conversations about Dell hardware, Dell Federal, or how this exploration relates to official Dell offerings, please work through your existing Dell account team — this project does not represent a commercial channel.