I build experimental infrastructure for AI agents and AI-operated systems: ways to observe runtime behaviour, evaluate it reproducibly, preserve evidence, keep authority outside the model, and verify what happened afterwards.

The recurring question is simple:

How do increasingly capable agent systems remain inspectable, evidential and governable without pretending one mechanism solves everything?

My work keeps five jobs separate:

observe  →  evidence  →  policy  →  permission  →  verify
runtime     support      standing    authority      recompute

Each project stands alone. Together they explore a division of responsibility:

agent runtime
     │
     ▼
Endophasia
observe · steer · compare
     │
     ├──── Magpie
     │     evidence · provenance · standing
     │
     └──── Deadbolt
           permission · effects · receipts

Cogitator
reproducible run verification

This is a division of responsibility, not a dependency graph. Observation is not authority, and a successful evaluation is not permission to replace a running system.

Endophasia has moved from architecture into repeated live agent studies. The current research separates three different claims:

exact replay               did the recorded session reproduce under recorded responses?
fixed-condition repeat     do repeated live runs under the same declared conditions agree?
perturbation invariance    does behaviour hold when an input that should not matter changes?

The useful outcome is not a prettier benchmark number. It is a better experimental instrument: one that can expose where sampling, environment, control flow and harness policy actually change an agent trajectory.

The runtime surface is also moving beyond one harness: a merged ACP v1 vertical slice has exercised a real OMP session through the standard protocol while keeping unsupported semantics explicit rather than treating protocol compatibility as capability equivalence.

The next layer is EVOLVE: treating prompts, cognition policies, harness components, tools, memory, models, environments and inference budgets as explicit candidate changes.

observe → propose → isolate → evaluate → compare → admit → promote

Current research directions include:

• pluggable evaluation, environment, adaptation and optional training providers
• harness and policy search
• test-time compute allocation and stopping policies
• held-out checks and explicit promotion gates
• trajectory-preserving evidence for candidate comparison

RL is one possible adaptation mechanism. Training can produce a candidate; it does not decide what the evidence establishes, nor whether that candidate may replace anything.

Languages: Rust · TypeScript/Node.js · Python · Bash/PowerShell
Agent/runtime: Pi · MCP · ACP · OpenAI-compatible APIs · local agent harnesses
Inference: llama.cpp · vLLM · SGLang · TensorRT-LLM · CUDA · GGUF · quantisation · long-context serving
Systems: Linux · Arch · Nix · systemd · containers/virtualisation · networking · low-level debugging
Delivery: GitHub Actions · reproducible environments · conformance testing · golden fixtures · cargo-dist · release verification

I use frontier and local models as a structured engineering workforce: architecture-first decomposition, bounded implementation briefs, maker/checker separation, adversarial review and human-controlled promotion.

observation  != inference
evidence     != standing
proposal     != permission
evaluation   != promotion
simulation   != real execution

These projects are experimental research software. I try to keep claims narrow, make failure states visible, test hostile cases, and document what each system does not establish.