Choppa Choppaaahh

Choppa

Hi! I'm an independent builder-researcher. Reach me on:

The Harness — system overview: cross-session AI memory + reasoning infrastructure; a fixed model compounding instead of restarting from zero.
Memory infrastructure: the ~1,900-note living knowledge graph + retrieval stack.
The autopoietic cycle: how the system maintains and improves itself.
Governance layer: commit-gate invariants, proposal queues, and why autonomy is earned.
scaffold-harness: the replicable how-to — the infrastructure layer as a build-it-yourself kit (model-agnostic, markdown-first). Made for sending to your own agents.
Getting started: build walkthrough — any LLM that reads markdown.

Elevation vs. capability: where richer instructions substitute for a stronger model (+28pp, pre-registered) — and where they measurably don't. Both failures reported.
decorr-eval: runnable benchmark + leaderboard — do your "different" models actually think differently? Judge-free, one API key, ~$1 a run. Headline: you can't infer decorrelation from the lab logo; measure it.
Decorrelation co-failure benchmark: how often different-lab LLMs all fail the same probe. Commit-reveal pre-registered.
Pre-registered benchmark freezes: hash first, results after — the discipline, in public.
Weekly signal pre-registrations: same commit-reveal discipline applied to research predictions.
Longitudinal fidelity measurements: does the memory actually survive? Measured, not asserted.

Agent team design: roles, guardrails, and three-layer governance for multi-agent work.
Commit-gate invariants: 26+ typed checks that run before anything ships.
The breadcrumb system: capture-at-the-moment memory formation.
Convergences: where our independently-built mechanisms later showed up in the literature.

Implementation scripts: sanitized, customizable versions of the harness's working parts.