Systematic strategies that survive multiple-testing correction. Multi-agent LLM systems that ship with a runnable eval harness. Same gate stack for both — declared before any number is published.
process/
Same five-step loop for every multi-agent build. The eval harness at the end is what keeps the other four honest.
Write the agent charter and acceptance criteria before any model is called. Output contract: JSON Schema strict.
Orchestrator-worker topology. Eleven agents. Each with explicit inputs, outputs, and failure modes.
MCP for structured calls (backtests, file I/O, retriever). No tool calls without a logged schema.
G1–G31 mechanical validity + statistical nulls + LLM-as-judge on samples. Frozen spec, no mid-run edits.
Runnable repo + readme + memo + eval-harness output. Scorecard attached, not promised.
Lane 1 · Quant Researcher ⟵ primary
Nine reproducible public-data research projects. Each candidate survives deflated Sharpe, CSCV-based PBO, walk-forward, and frozen-spec evaluation before the number is published.
Best-of-160 BTC rule: IS Sharpe 1.14 is only 1.24× the pure-noise expectation — DSR = 0.70 (fail).
IS Sharpe 0.91 → OOS −0.03 — an honest decay; bootstrap CI straddles zero.
Vol-targeting lifts Sharpe 0.27 → 0.39 and halves max DD (−62% → −30%).
Lane 2 · AI Engineer ⟵ secondary
Six OSS projects. Agent charters, MCP servers, eval harnesses, and LLM-as-judge gates — everything that says "shipped" also has a scorecard attached.
RAG service that proves its own retrieval — recall@3 = 0.886, MRR@3 = 0.805 with offline stdlib TF-IDF retriever.
ReAct-style tool-calling agent with OTel traces, fault injection, and 100% tool/arg correctness.
LLM-as-judge pipeline validated against human raters — Cohen's κ = 0.58 with bootstrap CI and position-bias measured.
The method · G1–G31
The same gate stack gates LLM outputs and systematic strategies. Six families, 31 gates, one contract. See the full gate taxonomy on the methodology page.
Public-data reproducible projects. Runnable eval harnesses. A 31-gate statistical validation stack. NDA-clean by construction.
