Yes, because the README explicitly places mined factor and strategy results in a candidate library and requires confirmation before publication, rather than automatically connecting them to monitoring or execution.

• The factor platform supports DSL, IC/IR, layered returns, Newey-West tests, factor portfolios, and nested out-of-sample search.
• The Mining page re-estimates factor direction in the training period, removes correlated factors, searches multifactor ranking combinations, and provides a user strategy as a control track.
• The README states that candidates are stored and published only after explicit confirmation, “never automatically brought online,” matching your required publication boundary.
• Backtesting also includes parameter sensitivity and rolling out-of-sample validation, with candidate saving and one-click re-testing.

However, providing validation workflows does not automatically eliminate data snooping, survivorship bias, or overfitting; the README offers no statistical guarantee for research results.

Yes, provided that you accept a continuously running local service, provider-dependent data capability, and position it as a research and monitoring tool rather than an execution system.

• The README defines TSP as a self-hosted A-share screening, monitoring, and backtesting workbench, with all data stored in local Parquet files.
• Quick Start requires capability detection and manually running the after-hours pipeline; in None/Free mode, same-day data is generally available 1–2 hours after the market closes.
• The trading-day probe handles adjusted holidays and stops real-time polling and minute increments on holidays, which fits persistent local use.
• Monitor supports auction, intraday, and deviation events, with pop-up alerts, voice announcements, Feishu notifications, and persisted trigger records.

The limitation is explicit: it is not positioned as a replacement for Tonghuashun, Tongda Xin, or an automated trading system; sleep, network failures, and rate limits may cause missed alerts.

Yes, because the AI assistant is designed as a local-data analysis extension rather than an automated recommendation or order-execution component.

• The README’s “AI Assistant” lists 18 read-only tools covering stocks, the market, sectors, watchlists, holdings, signals, strategies, and factors.
• Each call exposes the tool name, parameters, latency, and result summary, while every response includes risk and data-definition notices, matching your auditability requirement.
• The project explicitly converts trading-decision questions into technical status, financial status, key levels, risk factors, and conditional scenarios instead of issuing buy or sell instructions.
• Without AI_API_KEY, or with a provider that lacks tool calling, it fails closed rather than silently generating an answer.

However, the model may still misunderstand natural language or summarize incompletely; the README also does not define privacy and transfer boundaries when third-party model services are used.

Yes, because it connects synchronization, indicator calculation, screening, and backtesting through a shared enriched-data pipeline instead of implementing data handling separately in every page.

• The README section “Why TSP” states that screening, backtesting, monitoring, and review share one enriched data definition.
• “Core Features” lists about 68 indicator and signal columns, 25 built-in strategies, and one execution pool for daily and minute strategies.
• Backtesting includes T+1, fees, slippage, stop-loss, trade details, and SSE progress; minute backtests replay signals day by day.
• The architecture uses Polars for scanning, Parquet for local storage, and spawn worker processes for isolated backtest execution.

It is not a ready-made Python strategy library, however: custom strategies, field contracts, and data-source integration still require understanding the project’s data model.

Yes, because the project defines separate extension boundaries for backend routes, frontend pages, data fields, and the AI assistant.

• The README’s “Technical Architecture” says app/custom/ routes are auto-discovered and registered at startup, while src/custom/*/extension.tsx pages are mounted into frontend slots at build time.
• Incompatible versions or setup failures are isolated and skipped; deleting the directory removes the extension without modifying core files.
• “Data and Extensions” allows third-party or extension fields to become top-level menus and stores historical snapshots by day.
• The AI assistant is explicitly a reference extension: its backend and frontend are separate, and it fails closed when AI_API_KEY is absent or the model lacks tool-calling support.

This is friendly to React/TypeScript developers, but the README does not provide the complete extension API, type definitions, or compatibility matrix, so source and documentation review are still required.

It depends: the architecture is suitable for reducing provider coupling, but coverage, permissions, and field quality can still limit practical interchangeability.

• The README’s “Technical Architecture” says six dataset classes are routed independently by provider capability, so sources can be switched while indicator and backtest definitions remain consistent.
• Capability tiers include None, Free, Starter, Pro, and Expert; undeclared critical fields fail closed, such as rejecting an undeclared pct_unit.
• Financial data is PIT-merged by reporting period and announcement date, remaining null before publication; this helps avoid contaminating backtests by filling missing values with zero.
• The README also states that provider coverage, latency, adjustment methods, and permission tiers may differ.

Therefore, it suits researchers who need a replaceable data layer, but does not guarantee equivalent history or real-time capability across providers.