Every worthwhile expedition begins with a single question. Mine has never changed —

"How do you know your system actually works?"

Beyond the safe, well-lit demos lies rougher country: where a beautiful backtest is a booby-trap, an agent's confident "looks good to me" is a mirage shimmering on the sand, and a forecast with no error bars is a treasure map drawn from rumor. I go there on purpose — build the machine, then try to break it. Preregistered claims, executable gates, adversarial review, and an honest logbook of every run that didn't make it back.

Torches up. Here's what the expeditions have brought home.

• The Validation Crisis in AGI Capability Forecasting — I carried quant finance's hardest-won tools (deflated Sharpe ratio, probability of backtest overfitting, walk-forward) into the temple of AGI timeline forecasting, and came back with the Deflated Capability Forecast.  read →  ·  web edition →
• An Executable-Gate Multi-Agent Research Organization — a 39-role expedition party where no member's word is trusted on faith; every finding must clear a gate before it counts. Shipped with a pre-registered study that tests the rules instead of assuming them.  Zenodo · DOI →

• worldclass-research-org — the research party above: 39 roles, executable gates, adversarial review, pre-registered calibration.
• QuantBrief — a 5-model Mistral pipeline that reads 256K-token SEC scrolls, listens for the market's rumors, and returns with spoken briefings in five tongues. Mistral AI Hackathon 2026.
• neo-triage-hack — because some treasure falls from the sky: Bayesian triage of Near-Earth Objects with Claude Opus, cross-checked against JPL Sentry-II and ESA NEOCC. Anthropic × Cerebral Valley.  live →
• claude-code-skills — the toolbelt: Claude Code skills for Python, software engineering, ML, and quantitative finance.
• X_Quant — a lone crew sailing live markets: 20 AI agents, an Investment Council, and real IBKR execution.  overview →

_{More from the vault — an ESG compliance agent, a document analyzer, and Kelly-criterion prediction-market sizing — at kacpersaks.dev/projects.}

Executable gates over vibes · preregister before the result is known · the artifact is the deliverable · report the failures out loud.

In a hurry? Three relics tell the whole story — the Validation Crisis methodology, the research org that enforces it, and X_Quant sailing live markets. CV & references on request.

kacpersaks.dev  ·  LinkedIn  ·  LeetCode  ·  kacpersaks@proton.me