Research

Perslis Research

The research arm of Perslis — proof-first work on legacy-system control, data recovery, and verified synthesis. Every claim is labelled by how it was earned; every number carries its provenance.

Papers

Win95 USB Lab: from USB assumptions to testable evidence

Experimental previewVM and physical qualification pending

Why we built it, what the investigation revealed, and what still needs to run before we claim support. Includes the complete white paper, 20 recorded host checks, and the plan for VM and physical testing.

Proof Is the Product — the Perslis white paper

PILOT-READY wine tier: measured VM tier: prototype

Models are heads in jars — brilliant and helpless. Part I argues that the binding constraint on operational AI has moved from intelligence to embodiment: the harness, shell, and loop around a model — and describes Perslis, a model-agnostic operational runtime with a symbolic guardrail layer (the Captain), a persistent-memory evolution system, and a verification-first contract for "done." Part II proves the thesis on the hardest possible substrate — dead legacy binaries — with a self-evolving compatibility engine that routes, repairs, and verifies from the binary itself, with every claim tagged measured, prototype, or illustrative.

Read the paper PDF Typeset LaTeX · 4 pp · same source of truth as the page

Verified Program Synthesis with a Symbolic Knowledge Graph and an Un-gameable Compiler Oracle

Preprint · working draft

A no-bake, self-extending COBOL system: knowledge is attached, not baked into weights; an external compiler is the un-gameable judge; every answer is proven or refused. The system couples symbolic, provenance-tracked knowledge retrieval, deterministic program composition, executable compiler-and-runtime verification, and explicit refusal. It generates behaviourally correct GnuCOBOL across three structurally distinct program families, serves each answer in ≈486 ms on one CPU core with no GPU, and never emits unverified code — and because the verifier is an external deterministic oracle rather than a model grading itself, self-improvement cannot be reward-hacked.

Read the paper PDF Working draft — under active review

Project pages and practical guides

Read the Jira Monkey project story and terminal guide, explore legacy-data recovery with Liberate, or start with the Windows and USB testing page.