Six areas of work, clearly bounded

01

Governed AI & agent authorization

The problem. What must happen before an AI agent is permitted to act — and what happens when the permission mechanism fails? Most harnesses fail open and document the action afterward.

We research and prototype AI systems with explicit permissions, pre-action controls, human review points, and records that distinguish a proposed action from an executed result. The commercial proposition is not that an AI can always be trusted; it is that the conditions under which it may act become explicit, testable, and reviewable.

Evidence: COSMIC-ALLOW (published RFC) · T2Helix (instrument, in use) · Project Epistemic Bound (instrument, release blocked)

Deliverables available for scoping

• Permission architecture
• Bounded automation prototype
• Adversarial test plan
• Implementation review
• Event and evidence schema; replayable evidence bundle

02

Persistent knowledge & research infrastructure

The problem. Preserving usable knowledge across changes in people, models, sessions, and software — so a successor can recover an experiment’s assumptions, decisions, evidence, and open questions, and tell an authoritative record from an AI-generated summary.

We design software architectures for preserving research context, tracing claims to evidence, recording corrections that supersede earlier claims rather than silently replacing them, and supporting continuity across technical workflows.

Evidence: Sovereign Stack (instrument, in use) · public chronicle mirror · temple-harness (read-only access layer for local models)

Deliverables available for scoping

• Claim and evidence record design
• Continuity architecture review
• Read-only access layer for local models
• Correction and supersession workflow
• Handoff and recovery documentation

03

AI evaluation & human–AI workflows

The problem. Most of what shapes a deployed model’s output is application-layer material that is inspectable but never inspected. And human direction, machine contribution, approval, attribution, and correction interact in ways that are rarely documented.

We build research tools and assessment workflows for examining model inputs, system behavior, evidence quality, human oversight, and the limits of an evaluation’s conclusions. Corrections are retained: one readable example of a claim that changed after testing says more than a paragraph about integrity.

Evidence: The Black Box Boundary (study, bounded n=1) · governed co-creation case study and retrospective · Master Operations Agent (research workbench)

Deliverables available for scoping

• Evaluation harness
• Input-provenance audit
• Reproducible comparison
• Failure analysis
• Documented human-review workflow

04

Local-first AI & architecture R&D

The problem. Indiscriminate scaling is not the only route to useful AI. Which architectural choices actually change what a model computes, and what can run well on appropriately sized local hardware?

We investigate experimental AI architectures, locally operated inference workflows, and comparative studies of model behavior under practical computing constraints — measuring task quality, latency, memory, and energy where instrumented.

Evidence: Phenomenological Compass (study + instrument, supported within scope) · Phase-Modulated Attention (study, supported within scope, standing on a published negative result)

Deliverables available for scoping

• Architecture comparison
• Local inference prototype
• Ablation study
• Task-specific test harness
• Deployment runbook or feasibility report

05

Scientific modeling & simulation

The problem. Investigating how a complex system behaves — a process unit, an oscillator network, a thermodynamic hypothesis — without using the live system as the experiment, and keeping what the investigation actually showed.

We build computational models, interactive scientific instruments, synthetic scenarios, and experimental software for examining complex system behavior. The industrial simulator is one instance of this capability, not the company’s identity.

Evidence: Entropy as a Tunable Equilibrium (study, bounded null) · operator training simulator (independent training aid) · Temple Lab (chemistry workbench, shipping) · Kuramoto Teaching Instrument

Deliverables available for scoping

• Model implementation
• Parameter or sensitivity study
• Visualization or teaching instrument
• Synthetic training scenario or instructor workflow
• Simulation-based evaluation environment

06

Computational biology & pharmacology research

The problem. Turning a mechanistic question about cellular decision-making into predictions that a laboratory can actually test — and stating plainly which have been tested.

We do mechanistic modeling, literature-linked hypothesis development, and computational analysis in support of experimentally testable biological questions. The contribution is a model, evidence map, analysis workflow, or experiment-forward research package — with the next validation step named.

Evidence: VDAC1 Pharmacology Atlas (hypothesis set, untested in vivo) · Gate-Jamming Score (candidate biomarker) · CBD dual-pathway mechanism (literature concordance, not validation)

Deliverables available for scoping

• Mechanistic model
• Evidence map linked to literature
• Cohort or transcriptomic analysis workflow
• Preregistered prediction set
• Experiment-forward research package