System
Is it running, is it healthy, what does it cost, and how does it work?
Each run, Atlas researches the market across parallel phases — alternative data, institutional flows, macro, asset classes, and sectors — and synthesizes a daily read. Hermes then deliberates: it frames theses, screens candidates, runs per-ticker analysts and PM⇄analyst debates, and sizes risk. The result is a booked portfolio with a signed decision behind every position.
See the full graphEvery run is durable — these are the tables it writes, the source of truth the dashboard reads.
| What | Where | Notes |
|---|---|---|
| Research segments | documents | One row per segment (alt-data, macro, sectors, asset classes) |
| Daily digest | documents + daily_snapshots | The read — headline, regime bias, digest markdown |
| Analyst & deliberation notes | documents | Per-ticker analyst verdicts and PM⇄analyst debates |
| Portfolio decisions | positions + decision_log | The booked book and each signed call with its thesis |
| Run diagnostics | atlas_run_diagnostics | Cost, tokens, grounding, per-phase outcomes (this page) |
Runs are invoked from the command line. Model routing is automatic — chat phases use the configured reasoning model, web-search phases route to a grounding model.
--baselineFull pipeline — every research phase regenerated from scratch--deltaLightweight refresh — only changed segments re-run (~20–30% of baseline cost)--monthlyMonth-end synthesis across the period’s baselines and deltas