| name | paper-reproduction |
| description | Implement or maintain research-paper reproduction in DaxAlgo Terminal: paper ingestion, queued sandbox jobs, provenance-preserving signal bridges, confidence scoring, and backtest integration. Use for Windows backend work under src/windows/Core/TradingTerminal.Core/Research/ or src/windows/Pipeline/TradingTerminal.Infrastructure/Research/, or for the separate Linux PaperLab UI. Always load paper-ingestion and untrusted-execution; never assume the Windows and Linux trees share UI or implementations. |
Paper Reproduction
Turn a research paper into a queued, sandboxed reproduction whose validated outputs can feed the
backtest engine. Bridge signals and provenance; never vendor or execute paper code inside the host.
Choose the tree first
- Windows backend:
src/windows/Core/TradingTerminal.Core/Research/ and
src/windows/Pipeline/TradingTerminal.Infrastructure/Research/.
- Linux backend: the equivalent
src/linux/Core/ and src/linux/Pipeline/ paths; verify independently.
- PaperLab UI:
src/linux/AI/TradingTerminal.Ai.PaperLab/ only, composed by the Avalonia shell.
There is no Windows PaperLab project. A Windows backend change does not authorize Linux UI changes,
and a Linux UI change must use the Linux context layer and Avalonia contracts.
Pipeline
- Ingest a paper URL into
PaperRef and pinned RepoRef candidates through
IPaperIngestClient; keep sidecars on 127.0.0.1.
- Resolve the environment and run a minimal reproduction only through
ISandboxRunner; enforce
quotas, cancellation and deny-by-default egress.
- Validate declared artifacts and map them to
InstrumentId-keyed ReproducedSignals through
IReproSignalBridge.
- Score confidence, retain paper/repository/environment provenance, and expose a paper-tagged
backtest option through the existing strategy/backtest seams.
Layer placement
- Put records, enums, options and interfaces in that tree's Core Research/Configuration folders.
- Put stores, clients, orchestrators, sandbox runners, bridges and scorers in that tree's
Infrastructure Research folder.
- Keep Python/native environment logic in
tools/python-ml/; C# orchestrates over process or
loopback HTTP boundaries.
- Keep Linux PaperLab as a thin client over those seams. Do not move UI types into Core or Pipeline.
Reliability and scale
- Queue jobs asynchronously; stream status; make cancellation idempotent; requeue unfinished work
from
IReproJobStore after restart.
- Cache by paper identity, pinned repository commit and configuration hash before launching a runner.
- Run a minimal reproduction before a budget-approved full replication.
- Preserve paper ID, commit, environment hash, artifacts and confidence on results and signals.
- Fold expected external failures into explicit failed/unavailable results rather than crashing the UI.
Security invariants
- Execute untrusted code only in the configured container/VM/WSL boundary.
- Never mount credentials, the canonical store, user profiles or broad host paths.
- Apply CPU, memory, process, disk and wall-clock limits; kill the complete process tree.
- Accept output only through declared, validated artifacts.
- Never add a live order path; reproduced output is data/signals for backtesting.
When cloning process-lifetime behavior, verify the implementation in the selected tree first.
