Technical

The lab notebook. This section is the original research site — preserved at its existing URLs — covering the methodology, the experiments, every numbered learning, and the live state of the orchestrator. If you want the numbers, the leakage tests, and the negative results in full, you're in the right place.

For the plain-language version, see The journey. For the product the technical work is now feeding into, see OpenBracket — for investors.

Map of the technical site

Guide

The project's goal, methodology, harness evolution, and the autograd-style multi-agent orchestrator. Start here if you want to understand how the experiments are run, not just what they found.

• Project goal — what would have counted as success, what the realistic Sharpe ceiling looks like (L23).
• Methodology — walk-forward, leakage tests, bit-exact reproducibility.
• Harness evolution — what changed in the leakage harness between v0.1 and v0.6, and why.
• Orchestrator — five Claude Code agents running tmux teams on prod, claiming experiment nodes from a shared sqlite state.

Experiments

Every phase that produced a documented learning, in chronological order. Phase 5 onwards.

Learnings

All L1 through L54 — the carry-forward learnings every later phase has to respect. The most load-bearing ones are L23 (Sharpe ceiling), L24 (cheat-test scale mismatch), L47 (per-stock z-score identity leak), L57 (per_stock_zscore leak), L58 (look-ahead cheat preprocessing sensitivity), L59 (n_seeds=5 is undersized).

Results

• Best candidates — the two leakage-clean noise-band results (Sharpe -0.35 and -0.51). What "no measurable edge" actually looks like at this scale.
• What didn't work — the artifact-positive strategies. Sharpe +5 to +6 but failed window_robustness or static_features.

Live

Auto-regenerated every 10 minutes from the production sqlite database.

• Orchestrator state — current frontier, finished nodes, in-flight nodes.
• Recent commits — git log of the active branch.

Status

v0.6 — methodology release-ready. After 12 phases, 367 orchestrator nodes, and 54 carry-forward learnings, the honest result on Indian large-caps at both 5×3y and 46×3y windows is no measurable time-edge. The methodology and the negative finding are the deliverable on the technical side.

The technical infrastructure now powers OpenBracket — the Phase 2 pre-open range forecaster. Phase 2 reuses the leakage harness, the walk-forward eval, and the orchestrator infrastructure unchanged.