Integration and Workflow Capabilities

Contract

This matrix distinguishes source implementation, source-run evidence, source GUI/worker acceptance, and clean distribution acceptance. Nine migrated research implementations have been exercised from the source checkout, including the five GUI courses, the 19-figure report, cache, and MCP paths. This does not certify a clean wheel installation or optional runtimes in every user’s environment. “Discoverable” means a capability can be listed, not that a calculation succeeded.

All example results are RESEARCH-ONLY with pit_guarantee=false. Inputs are explicit user-owned CSV, Parquet, or completed manifests; no raw J-Quants cache, full report, model artifact, credential, or Paper ledger is bundled. The accepted reference workflow is described in the five GUI research courses; screenshot provenance and scope are in the public evidence summary.

Status dimension	Data	Indicator	Model	Backtest	Broker
Implemented in source	J-Quants connector code plus explicit local-cache loading, validation, selected-bar freezing, and identity checks.	TA-Lib and native pandas-ta providers run in isolated workers; Native MA is a strategy path, not a general indicator provider.	Fixed LightGBM and CatBoost research recipes with chronological split and purging.	Factor/model score strategies, Native/VectorBT/Backtrader fixed-order replay, dashboard, sensitivity diagnostics, and isolated historical Paper replay.	Offline order mapping and separately invoked read-only cash/positions/orders diagnostics; submit/cancel actions are disabled.
Discoverable	Local-input workflow and source formats are documented; GUI capability depends on the installed build.	Provider/runtime metadata can be listed; availability is determined by the isolated worker, not main-process imports.	Model workflow and fixed learners are listed; optional runtime resolution is separate from discovery.	Strategy, engine, report, and historical-Paper workflows are listed separately; listing is not a successful run receipt.	Mapping and read-only capabilities are distinct; a credential reference is not a connected terminal.
Runnable source evidence	The reference uses an existing local J-Quants cache: 4,392 raw-price rows over 1,464 common observed dates, 2017-01-04 through 2022-12-30. This does not verify a live J-Quants request.	Both providers produced 1,464-date reference series; outputs were independently checked against indicator calculations.	Two models each produced 2,190 historical predictions; checkpoint reload reproduced predictions with zero reported difference.	Factor research kept 4,209 feature rows separate from 4,392 forward-label/evaluation rows. Factor strategy: 1,464 dates. Model strategies: 489 dates. Four engine candidates matched on the fixed 20/60 comparison. Historical Paper: 1,464 dates, 156 fills, 11 skips. Nine fee/delay scenarios matched independent financial-path accounting.	Source supports offline mapping. Read-only source path is implemented; a real broker terminal was not verified.
Source GUI/worker acceptance	Local cache selection, validation, and freezing were exercised; no online J-Quants request is claimed.	Both isolated TA-Lib and pandas-ta workers were exercised over the reference data and their GUI plots checked.	The fixed LightGBM/CatBoost GUI paths were exercised; six model-specific checks passed with no skips.	Five courses, the 19-figure report, factor/model/engine paths, Paper, and nine sensitivity scenarios were run and reconciled from source; cache/MCP paths were also checked.	Source GUI paths for offline mapping and explicit read-only diagnostics were checked; missing reference failed with zero GETs. A real terminal remains unverified.
Clean wheel/package acceptance	Gate: install the exact wheel identified by SHA-256 and version in an isolated -I environment, record the installed module origin, and run the local-cache workflow. Source GUI evidence alone is not installed-package evidence.	Gate: use the exact wheel SHA-256/version and an isolated -I install, record module origins, and run the selected indicator provider with its declared optional runtime. Source GUI evidence alone is not installed-package evidence.	Gate: bind the run to the exact wheel SHA-256/version, isolated -I environment, and installed module origins; run the selected learner and retain its result. Source checks alone do not certify the installed worker.	Gate: bind the selected backtest operation to the exact wheel SHA-256/version and isolated -I origins, then run it and retain its result. Source GUI evidence alone does not certify installed worker startup.	Gate: bind offline mapping and any explicitly requested read-only operation to the exact wheel SHA-256/version and isolated -I origins, then retain operation evidence. This does not verify a real terminal or enable orders.
Real terminal verification	Not applicable.	Not applicable.	Not applicable.	Not applicable.	Not verified. An explicit read-only request requires a configured reference; an offline mapping preview makes no network request. No order is submitted or cancelled.
Evidence and limits	The reference uses a local J-Quants cache and a fixed actual TOPIX closing-price index, with no dividends and exact NAV-date matching. It does not establish historical availability.	SMA/RSI/MACD/ATR are provider outputs with documented initialization and warm-up; no signal or order is implied by an indicator.	Split: train 2017–2019, validation 2020, historical diagnostics 2021–2022, seed 42. The test period is already observed history, not fresh OOS.	Portfolio quantities, fees, skips, and prices are recalculated from fixed recipes; matching account paths do not imply identical fills for every engine. Paper is historical continuous-share replay, not forward Paper.	Secrets are referenced, not stored by the mapping workflow. Historical-PIT, approval-bound live trading, terminal operations, and real submissions are not provided.

Reading the capability states

Source-checkout GUI/worker paths have been exercised, but that is separate from clean wheel/package acceptance. Across four optional runtime groups, 31 source workflow tests passed with no skips; these were local checks, not GitHub Actions runs. A discovered provider or listed engine is not proof of a successful calculation. Broker read-only diagnostics are explicit and separate from the zero-network offline preview; neither means a real terminal was validated. Nothing here upgrades the research outputs to strict PIT, investable, PAPER-READY, or live-trading evidence.

Next work

The roadmap tracks the remaining clean wheel/install gate and longer-horizon research requirements. The most important research gaps are reliable point-in-time availability, exchange-calendar and corporate-action handling, dividends, capacity and realistic execution costs, and fresh forward observation. Public Regime remains Off; its private state-machine bridge is not included.

Evidence

The accepted source-checkout GUI demonstrations and their limits are summarized in the GUI course guide and its linked public evidence summary. They contain screenshots and provenance summaries, not raw bars, complete reports, model checkpoints, or Paper ledgers. Source GUI acceptance does not mean a clean wheel was installed; it also does not prove a real broker terminal was connected or live trading enabled.