docs(run): describe cleanup coverage and preload limits Co-Authored-By: Codex <noreply@openai.com>
Run, Architecture, Decisions, C4 model, Quality, Glossary, Library: runs fork from a forkserver (#178)
Name the core calls the CLI and the server share (#180)
Studies: named collections of runs that differ only in their parameters (#157)
Run: describe the subpackage layout of sbt2.core.run (#160)
Run, CLI: launch runs uncapped by default, with the systemd scope opt-in (#153)
docs: describe the uv workspace and the sbt2 namespace packages (#145) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Run: the launcher is an ABC picked by platform
Run, Quality, C4 model, Deferred: run children are capped over systemd's D-Bus API
Run, Results, Decisions: the run child's order is JSON, the store travels as a locator Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Results, Run, Glossary: batch id and part and batch comparisons (#111)
Strategy, Spec, Results, Run, Glossary: the risk table and the drawdown guard as built (#98) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Run: sbt2 registers its pickle reducers once on import
Run: pickle reducers live in sbt2.pickling, registered by the order transport
Run, CLI, Results, Quality: run.batch and sbt2 run's memory options (#58)
Drop provenance, code identity, result identity and the data fingerprint
Run: the fresh-process cost was a one-off measurement
Run, Quality, Deferred: record the systemd-run memory cap spike (#54)
Candles: 1-minute candles stored, bars built from trades or candles per run (#49)
Liquidation off without quotes, a spec key for it, and the faster run numbers
Run pre-flight: source symbol, known gaps in the summary, new errors
Splits: train / validation / test, with only test locked The optimiser tunes on train and ranks on validation; validation runs as often as you like; the lock becomes the test lock. Walk-forward is deferred until M10. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Library: settle the open questions and bring every page in line Two parts (in-sample / out-of-sample) replace train / validation / test; Holdout by fraction or date, no embargo; only the final out-of-sample part is locked, per lineage by date overlap; every run stored; code identity replaces the git SHA; one root replaces data/ and config/. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Library: target design for sbt2 as a published package, open questions Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CLI options, reversed-position funding in Run, golden-run tests in Quality
docs: thin nautilus strategy base; asset profiles keyed by nautilus Strategies are nautilus Strategy subclasses on a thin sbt2 base that submit their own orders; the plain-Python pipeline (intents, reconciler, adapter) is gone (PR #18). Asset profiles are keyed by nautilus's AssetClass + InstrumentClass and venue profiles name them with nautilus enum names (issue #19). Results notes that annualization must be passed as nautilus's period (default 252) and that benchmark-relative stats come from nautilus. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add nautilus characterization tests: funding epoch alignment, node disposal, test-to-decision map
Move to nautilus v2: native funding, portfolio equity snapshots, run-time bars, public catalog writers, nautilus tearsheets Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drop the ingest manifest: coverage, instrument types and fingerprints come from nautilus's catalog Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Move to nautilus: run config, venue profiles, mark-price valuation, brackets, RiskEngine, bar aggregators, instrument provider, reports Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>