LetItBe
gsuda-engine is a provenance-first self-evolving trading loop for Korean equities.
Core trick
Logs each trade with its feature vector, and an LLM-drafted suppression rule must pass a validation gate before the risk engine loads it.
Tags
Description
gsuda-engine is a provenance-first self-evolving trading loop for Korean equities.
It solves a practical problem in trading agent design: generated rules should not be deployed blindly. The system logs simulated trade recommendations with full feature vectors, tracks T+5 outcomes, clusters failed trades, drafts candidate suppression rules, and then validates each rule before Risk Guardian can load it.
The project is built from what I know: 23 years operating Korean securities systems and 13 years studying Saju / Yeokhak. Saju is not treated as fortune-telling or price prediction. Instead, original Hanja-based Saju fields are preserved as domain-informed categorical features that must be tested empirically.
In the demo, one candidate rule is promoted to skills/active after passing validation, while another is quarantined because it damages too many winning trades. The broader local research warehouse covers Korean equities from 1995 to 2026, with 11M+ enriched rows and 2,700+ stocks.
Core principle: Claude drafts. Data validates. Risk Guardian deploys.
README excerpt
# gsuda-engine Provenance-first self-evolving trading loop for Korean equities. Claude drafts candidate risk rules from clustered failed trades, but every rule must pass a validation gate before Risk Guardian loads it. The project combines 23 years of Korean securities-system experience with Saju/Yeokhak-inspired feature engineering. The claim is not that Saju predicts stocks; the claim is that domain-informed categorical features can be tested against market history. ## What Problem It Solves Trading agents can learn the wrong lesson from a few vivid failures. A language model may draft a plausible new rule, but deploying that rule directly is dangerous. gsuda-engine makes the feedback loop explicit: 1. Log every recommendation with its full feature vector. 2. Wait for the realized outcome. 3. Cluster repeated failures. 4. Draft a candidate risk rule. 5. Promote the rule only if it passes historical validation. The demo intentionally shows both outcomes: one rule is promoted to `skills/active/`, and one tempting rule is sent to `skills/quarantined/` because it would suppress too many winners. ## Research Thesis: Time Climate And Collective Behavior My working thesis comes from studying Saju/Yeokhak alongside market systems. In traditional Saju, a person's pillars are often read as a map of tendencies, temperament, timing, and life arc. Year, month, and day pillars can also be understood as the "time climate" of a moment: the heaven-and-earth context into which peopl