Generate signals
alphinity-aisys @ 7f78814 · 2026-10-16
Runs AISys's backtest with its own code (the same as its daily job
run_backtest.ipynb) over a time window, and writes the signals that are missing.
Pick a strategy, a window, and press ▶ Run — nothing is written before that.
- 1 · Strategy
- 2 · Window
- 3 · Run
- 4 · Check in Databricks
1 · Strategy
start from a saved strategy and change what you need — the fields marked differ from AISys's defaultA date already written for the same (agent, series) —or a poem at the same version— is skipped. To generate a new series, change its name.
Saving is optional — ▶ Run uses what is on screen. Saved strategies live on this server
(config/aisys/); on the deployed site a redeploy resets them.
YAML what ▶ Run uses · the form writes here · edit it to add or remove agents and poems
default is AISys's own params.yaml and is never overwritten.
2 · Window
dates in UTCWrites to:
3 · Run
writes to DatabricksCalls AISys's own run(catalog, mode="backtest", start, end) — the function its daily
job runs — with the strategy of step 1. For every trigger that is missing something:
- Reads the market at that minute from
production:silver.btc_market_*,bronze.orderbook_data,silver.vw_iv_surface,gold.vw_economic_calendar… - Agents — each give a signal (rules, no LLM).
- Poems — write the market context.
- CIO takes prompt
cioml— from the registry and asks the LLM (openai/gpt-5-mini, OpenRouter) for the final signal — the only LLM call. Voting takes the sign of the agents' average. - Writes all of it to —.
Anything already written is skipped, so running the same window twice is safe.
▶ Run goes to the background — follow it in Runs → log. The preview writes nothing: it counts how many rows ▶ Run would write and how many LLM calls it would make.
| Piece | Written as | Already there (skipped) | Run will generate |
|---|
Verify these numbers in Databricks
Same pieces, same window: triggers, already_there,
run_will_generate. Sum the last column = rows ▶ Run writes; the cio row = LLM calls.
4 · Check in Databricks
read-only SQL for exactly this strategy's series, catalog and window — paste it into a Databricks SQL editorPick a valid strategy and a window to get the queries.
Runs
| Started (UTC) | Params | Catalog | Window | CIO | State |
|---|
The log shows, per component, how much was missing before and after the run. AISys's
errors land in <catalog>.gold.btc_system_monitoring_aisys.