Humans set the rules.
The system grades itself.

We built the system as layers we call the RI series. A value-judgment layer, a market-structure layer and a stock-lifecycle layer each produce an output. A judgment core combines the three, issues the final call and grades itself. A derivative for young and small companies reuses the same kernel under a stricter charter: no growth or earnings forecasts. Every model is pre-registered before scoring, 563 so far. No statistic reaches the screen before a point-in-time backtest. Rejected hypotheses stay published beside the ones that passed. Coverage is Korea's KOSPI and KOSDAQ, and US equities.

In the volatility layer, ten pre-registered models race in real time. Each stamps an 80% interval forecast at 30-minute and 60-minute horizons, and a record is never edited once written. Scoring covered 27 Korean symbols and 2.95 million observations. We calibrate so that out-of-sample coverage converges to the nominal 80%. A band that is too narrow hides risk; one that is too wide is useless. The results are written up in a working paper. We do not claim to call direction. We measure the size of uncertainty and publish how well that measurement held.

AI agents write the code. Humans write the rules. Agents with separate build, review and exploration roles produce code under those rules, and an agent from a different model reviews it independently. Before anything runs, automatic guards step in to block access to protected paths, leaks of secrets and dangerous commands. 227 research modules and 474 test files have accumulated this way since July 2026. We treat the operating method itself as part of the research. Deciding what an agent may do is governance.