Gary Engineering Lab
Open engineering records for FPR’s local-first AI stack: self-hosted agents, benchmarks, service reliability, recovery tests, and modular open-source software.
Freedom’s Purchase Ranch / Engineering Program
Real systems. Real failures. Verifiable fixes.
Meet Gary Engineering Lab: the open engineering journal behind FPR’s local-first, modular AI infrastructure. We publish benchmarks, agent experiments, technical incident reports, reproducibility notes, and explicit limits—not polished claims without evidence.
From a living lab to open-source infrastructure
FPR’s pilot infrastructure is a working laboratory for software that should survive real failures: intermittently reachable services, model context mismatches, agent orchestration problems, file storage limitations, and the human need for simple recovery.
FPR Mechanic
Local-first observability, evidence-driven diagnosis, backups, and guarded recovery. An early modular prototype exists; it is not yet production-qualified.
Gary Model Lab
Reproducible model comparisons: correct output, first-token latency, memory use, context, failure modes, and independent scoring.
Agents in the open
Hermes, coding agents, orchestration, local services, and workflows. Every published capability should have a dated, inspectable acceptance test.
Featured investigation: why a connected Hermes agent failed
A live Telegram gateway answered with an error although its connection was healthy. Our log documents the 32K/64K context incompatibility, independent GPU-model tests, a 65,536-token profile, and what still needed end-to-end verification.
What our lab reports include
- Methods: exact tasks, evaluation rules, model and software versions, and hardware constraints.
- Results: measured pass/fail/blocked outcomes, full defensible latency metrics, and known confounding factors.
- Evidence: checksums and sanitized reproducibility instructions without exposing private records or security credentials.
- Recovery: implementation scope, required approval, rollbacks, and explicit acceptance gates.
Negative results stay in the record. Private data and unreviewed ranch footage do not belong in public engineering posts. Public code releases will include license, modular installers, regression tests, restore verification, and a security review before claiming production readiness.
See the latest work: Gary Engineering Lab archive · FPR mission