LAB NOTE 00 / OPERATING RULE
Evidence, inference, narrative
Every technical page should make it possible to tell which layer a statement belongs to. Evidence is what was measured, observed, sourced, or reproduced. Inference is what we conclude from that evidence. Narrative is how we frame the material so a human being remains awake long enough to understand it.
Those layers can coexist. They just should not impersonate one another. This sounds obvious right up until a confident caption, marketing sentence, or elegant diagram quietly upgrades an inference into a fact.
LAB NOTE 01 / MULTI-PERSPECTIVE AI
Benchmark the chamber, not the costumes
The core research question is whether structured multi-perspective reasoning, with persistent differentiated cognitive roles, selective routing, adversarial interaction, and evidence-aware convergence, improves reasoning coverage, error detection, calibration, or decision quality relative to strong simpler workflows.
The public protocol now defines four core conditions: conventional single pass, self-critique, independent multi-sample aggregation, and structured differentiated roles. It also defines budget matching, error-detection metrics, calibration, unsupported-claim rate, coverage, human usefulness, latency, token cost, and five architectural ablations.
The most important part is the failure rule. If gains disappear against strong baselines or after cost normalization, the architecture does not get to declare victory because the interface has better dialogue.
LAB NOTE 02 / ROBOTICS
Human-machine work starts on the bench
The long-term cybernetics program should climb capability in stages rather than jumping directly to the most dramatic interface. A credible sequence begins with myoelectric or other non-invasive intent signals, robotic actuation, and deterministic safety limits. Then adaptive intent prediction can be added and benchmarked against the simple controller.
Haptic or tactile feedback comes after the actuation stack is understood well enough to know what information is worth returning to the operator. More invasive interfaces belong much later and only inside appropriate clinical, ethical, and regulatory structures. The early engineering work can still be ambitious without pretending a benchtop result is a medical device.
Before any human-contact prototype becomes interesting, characterize force limits, pinch hazards, thermal rise, electrical isolation, runaway behavior, emergency stop logic, loss-of-signal state, sensor failure, and what the machine does when software becomes confused. Graceful failure is a feature.
LAB NOTE 03 / FIELD ENGINEERING
Serviceability is performance
Engineering teams often optimize the object and accidentally punish the person who has to keep it alive. Field systems deserve different priorities: access, standard fasteners, visible diagnostics, replaceable modules, known consumables, mechanical overrides, documentation that can be used with dirty hands, and enough telemetry to distinguish a bad sensor from a bad assumption.
Resource constraints can improve design when they force clarity. They become dangerous when improvisation is treated as evidence that margins no longer matter. Salvage, repair, and repurposing are powerful engineering methods when the unknown history of a component is incorporated into the test plan instead of wished away.
LAB NOTE 04 / SITE INFRASTRUCTURE
The website is already part of the experiment
The public site is intentionally lightweight: static HTML and CSS, private source control, automatic production deployment, explicit security headers, no remote font dependency, crawler blocking during review, and a custom domain behind HTTPS. That architecture is not glamorous. It is inspectable and reversible, which is a better starting point.
Future complexity should have to earn itself. A content system becomes useful when editorial volume exceeds what static pages can manage. Accounts become useful when there is something worth authenticating. Databases become useful when data has a lifecycle. The stack should grow because the work demands it, not because software developers experience emotional distress in the presence of an empty dependency file.