The machines
Apple Silicon / Mac Studio
NVIDIA DGX Spark
Local compute for development, inference and experiments. Hardware capacity is an input; useful accepted work is the outcome.
INDEPENDENT APPLIED RESEARCH / ABOUT SW30 LABS
Human judgment.
Agent momentum.
Shared evidence.
One researcher. A growing system of agents. A research practice built to turn questions into small, inspectable experiments.
Nic sets direction and makes consequential decisions. Agents contribute research, implementation, verification and continuity. Together, we keep what works, preserve failures and decide what deserves another experiment.
Build a chapel, not a cathedral.
And no preaching.
And no preaching.
01 / THE RESEARCH LOOPSeparate projects.
Separate projects.
A shared purpose.
Memory informs. Strategy directs. Agents work.
Evidence returns to human judgment.
Select a role to explore its connections.
Implemented connection Proposed handoff
Architecture map, not live telemetry. Solid links represent the implemented priority-to-Nightshift path and its scope/review records. Dashed links show the wider intended feedback loop. Last reviewed 6 September 2026.
02 / THE CADENCEA rhythm for work.
A rhythm for work.
Room for judgment.
Daily operations. Weekly direction.
Strategy changes when the evidence warrants it.
01 / DAILY
Every morning
Delivered evidence, failures, accept / reject / defer, and the effort it took to review and repair.
DECIDE WHAT TO KEEP ↗02 / WEEKLY
Reconnect to priorities
Achievements versus intent. Relevant Singularity Atlas themes. Blockers, review burden and the next allocations.
CHOOSE WHAT COMES NEXT ↗03 / MONTHLY
Revisit the forces
Refresh the full Porter assessment: rivalry, suppliers, buyers, new entrants and substitutes. Reconsider the strategic implications.
CHALLENGE THE DIRECTION ↗04 / EVENT-DRIVEN
When something changes
A material finding, dependency change or external development triggers an immediate reassessment of the affected project.
RESPOND TO THE EVIDENCE ↗03 / A CONCRETE EXAMPLEFrom a photonic idea
From a photonic idea
to a falsifiable question.
RPC-H16 connects the research loop to a frontier question: can a three-dimensional optical architecture compute usefully while degrading gracefully under faults and intermittent energy?
Explore the frontier research ↗- 01
Approve the assumptions
Define physical bodies, channel mappings and clearance rules before geometric conclusions.
HUMAN DECISION - 02
Run a small simulation
Deterministic fixtures, explicit edge decisions and a reproducible report. No language model in the numerical path.
NEXT BOUNDED EXPERIMENT - 03
Let measurements answer
Future optical calibration and a physical gate could test the model. Hardware requires separate preparation and approval.
FUTURE / NO OPTICAL PROTOTYPE CLAIMED
04 / MY TECH STACKLocal compute.
Local compute.
Connected instruments.
The tools behind the research practice.
Each layer has a job to do.
The model layer
oMLX / MLX
Local Spark serving · Ollama
Local model serving for agent workloads and private-content processing. Model selection follows the job and available capacity.
The working language
Python · LangGraph
Deterministic scripts and tests
Graphs where stateful agent work needs them. Small standard-library experiments where a simpler tool answers the question.
Memory and context
TwinOps · MCP
Neo4j · Snapshot data
Retrieve relevant background with provenance. Live graph queries and saved snapshots have different freshness; neither overrides current decisions.
Observe and assess
Loopscope · DriftLab
OSCAL Skills Guardrails
Execution traces, behavioral-drift research and skill-admission evidence. Complementary instruments, with different scopes and integration maturity.
Deliver and decide
GitHub · Nightshift
GrokBot / Morning Prayers
Versioned artifacts and reviewable branches, followed by human decisions. Accepted evidence feeds the next portfolio review.
A working stack, not a claim that every tool is connected to every other tool. The map above distinguishes implemented connections from intended handoffs. Specific models and configurations change with the experiment.
More useful work accepted.
Less effort to get it there.
Accepted outcomes
alongsideInspect the artifact. Assess the evidence. Record what actually advanced the question.
Specification + review + repair
Count the human effort, failed attempts and maintenance that made the outcome possible.
Track both. Do not hide the burden behind commit counts or combine unlike units into a productivity score.