feat: add autonomous paper-trading and calibration pipeline

2026-08-03 14:03:27 +01:00
parent 5a9a2e4c6d
commit c4028cc394
46 changed files with 2246 additions and 115 deletions
@@ -0,0 +1,53 @@
+# Duriin autonomy runtime
+
+The autonomy runtime is additive to the existing archive and intelligence
+tables. It is deliberately disabled until the paper-trading cutover is
+approved.
+
+## Services
+
+- `api`: HTTP only; no background scheduler when `DURIIN_RUN_SCHEDULER=false`.
+- `autonomy`: bounded archive reconciliation and live/historical job creation.
+- `coordinator`: LLM proposal extraction only. It cannot create an order.
+- `autonomy-outcomes`: resolves matured predictions against market and benchmark returns.
+- `calibration`: creates empirical calibration snapshots and deterministic decisions.
+- `intelligence`: legacy worker, available only under the `legacy` Compose profile.
+
+Start only the API by default:
+
+```bash
+docker compose up -d api
+```
+
+The new runtime is explicitly opt-in:
+
+```bash
+docker compose --profile autonomy up -d autonomy coordinator autonomy-outcomes calibration
+```
+
+## Authority rules
+
+The coordinator may return categorical, evidence-backed proposals. It may not
+return probabilities, returns, position sizes or trade actions. Proposals must
+reference existing archive article IDs and instruments in
+`autonomy_instruments` with `active=1` and `tradable=1`.
+
+Calibration is computed from resolved market-relative outcomes. The policy
+engine emits `BUY`, `SELL`, `HOLD` or `ABSTAIN`; an LLM is not in this path.
+
+Paper order intents use deterministic client IDs and are bounded by the
+execution validator. A broker adapter must be explicitly configured; the
+included simulator is the default test adapter.
+
+## Database initialization
+
+The autonomy schema is initialized additively by the autonomy workers and
+maintenance scripts in the intelligence database. To create the initial
+reconciliation job without starting workers:
+
+```bash
+npm run autonomy:init
+```
+
+Legacy predictions remain legacy records and are not silently included in
+calibration. New prospective predictions are the trusted learning set.