Skip to content

Repository files navigation

Choppa Research

Public artifacts of an independent builder-researcher working with a team of AI agents. Research, benchmarks, methodology — published honest about their own failure modes.

What this repo demonstrates, in one breath: pre-registered measurement discipline (hash first, results after) · a production multi-agent harness that compounds a fixed model across sessions · and the receipts for both, including the negative results.

SHIP LOG — dated one-liners of everything published here, newest first.

Results

finding where
Decorrelation co-failure benchmark (flagship) — how often different-lab LLMs all fail the same probe. Cross-lineage, commit-reveal pre-registered. benchmarks/decorrelation-benchmark.md · writeups/decorrelation-benchmark-v2.md
Longitudinal memory-fidelity measurements — does cross-session memory actually survive? Measured over months, not asserted. measurements/
Elevation-vs-capability boundary study — where richer instructions substitute for model capability (+28pp, pre-registered), and where they measurably don't (code transfer: FAIL under its own bar). writeups/elevation-vs-capability-boundary-2026-08.md

Discipline artifacts

what where
Pre-registered benchmark freezes — hash-committed before results exist. preregistered_benchmark_freezes.md
Weekly signal pre-registrations — same commit-reveal discipline, live cadence. preregistered_signal_freezes.md

The harness

A ~2,400-note living knowledge graph + reasoning-pattern promotion + routed specialist workers + metacognitive compile cycles: a fixed model compounding across sessions instead of restarting from zero. Core result — better scaffolding × the same model = materially better results.

  • Start here: architecture/system-overview.md
  • Build it yourself: methodology/getting-started.md
  • Current worker architecture: methodology/agent-team-design.md

Built by

An independent researcher + Claude (Anthropic), 400+ sessions. Not a reference system — it's how we work.

Landing page / more artifacts: github.com/Choppaaahh · GitHub Issues to start a conversation.

About

External memory infrastructure for AI agents — measurement frameworks, multi-agent orchestration, metacognitive self-improvement. 9 independent convergences. Open methodology + implementation scripts.

Topics

Resources

Stars

1 star

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages