Research

The deeper material.

Research behind the witness log and the evaluation work, and into reasoning and supervision architectures that is not yet released. Concepts are shared as direction, not as shipping products, and each is labelled for what it is.

Evaluation methodology

Questions we had to answer to trust a result.

Grading open responses
A judge model assesses, its quotations are checked against the response, a second pass audits the assessment, and a plain deterministic step reconciles them. A contested grade is recorded as contested — not silently turned into a fail.
Twins over single records
An adversarial record alone rewards caution for the wrong reason. Paired with a twin where helping is correct, the same run separates discrimination from topic avoidance.
Scoring conventions decide conclusions
Whether an escalation counts as a catch, a non-answer or neither can reverse which of two systems looks better on identical runs. So the convention is written down before the numbers exist.
Escalation is a metric
How often a system hands a decision back — and whether what it hands back was actually hard — says as much as whether it was right.
How the catalog enforces this →How training is kept checkable →Behavioural discovery →

Witness logging

Where the witness log's format and its independent checks are worked out.

Public · Rust · early-stage

mirror-node-reconnaissance

A defensive, consent-based simulation harness built around one component: a witness log that survives a dispute. It ships the log, the seal, the witness-bundle/2 export format, a producer and signer, a Python verifier that shares no code with the producers, and a sealed lab that exercises them under load — entirely in simulation. Its README states its limits before its results.

  • Witness bundle
  • Independent verifier
  • Simulation only
GitHub →On the Witness page →

Research & concepts

Active research not yet released. Shared here as direction, not shipping products.

Research concept

topological-rag

A four-pathway topological RAG orchestrator (Rust workspace + FastAPI sidecar): an advisory-only LLM boundary with adversary (hallucination-flagging) and synthesizer pathways, backed by a witness log.

  • RAG
  • Advisory
  • Witness
Pre-release
Research concept

Gatekeeper V2

Intent-centered supervisory system for AI-to-AI and AI-to-tool interactions: hash-chained event log with drift and trust scoring. Local-first, TypeScript.

  • Governance
  • Intent
  • Trust scoring
Pre-release
Research concept

R-LLM

Dual-hemisphere split-brain LLM security classifier: two lenses reconcile a verdict, with a superposition-collapse probe. The research root beneath the split-brain harness.

  • Classifier
  • Split-brain
  • Method
Research The harness it became →

Intellectual property

Formal disclosures, published to establish position.

Prior-art disclosure

Thumb-Server

Topological Photonic Thumb-Server — a formal 35 U.S.C. §102 prior-art disclosure (inventor Christopher Howard). Published to establish IP position, not a shipping product.

  • Photonics
  • Hardware
  • IP
Disclosure →

Collaborate

Research collaborations are on the table.

Start a conversation →