How to Turn Institutional Crawl Signals into High-Quality Trade Opportunity Windows

Estimated read time: 5 min

The question this system answers is precise: when does institutional evidence become strong enough, persistent enough, and coherent enough to justify a trade setup window? We treat this as an inference and control problem, not a pattern-spotting problem.

The signal stack is deterministic and auditable: doc_stance -> pressure_features -> setup state machine. On frozen run data (c180_tuned_01), that means every setup decision can be tied back to timestamped documents and confidence channels rather than discretionary interpretation.

At policy level, SMC selects and reweights strategy specifications at inference time while MolLEO optimizes the spec pool offline. The control plane is explicit in config:

"action_policy": {
  "name": "smc_specs",
  "params": {
    "spec_pool_run_dir": "/mnt/volume_sfo3_01/feb4_runs/run_20260206_180258",
    "spec_pool_max": 24,
    "pressure_conf_min": 0.1,
    "doc_count_min": 2,
    "ess_docs_min": 0.8
  }
}

A trade opportunity window opens only when force magnitude clears a cost-aware threshold, direction remains persistent, and certainty gates (confidence, doc_count, ESS) stay valid. It closes on force reversal, confidence collapse, regime invalidation, or time stop. That lifecycle is explicit and traceable as CANDIDATE -> VALIDATED -> ACTIVE -> EXITED/REJECTED.

The key architectural win is separation of concerns: perception builds force from evidence, setup translation produces an auditable trade plan object, and execution only follows that plan. This keeps the system research-grade and reproducible while allowing aggressive optimization over gates, decay, sizing, and invalidation logic.

Internal links

• Index: /office/
• Sibling: Set the right LLM concurrency cap
• Sibling: How CTOs increase LLM throughput (8-shard proof)
• Contact: /contact/

Scope boundary

This note covers: evidence-to-setup translation and optimization control for setup quality.

This note excludes: order execution details, broker integration specifics, and proprietary crawler implementation.