Open Source Agent
Reviews repositories by domain, architecture, license, reproducibility, code quality, integration value and benchmark path.
- Action: Reference / Fork / Integrate / Reject
- Output: engineering evidence and benchmark plan
Pengyi Research Development Team
AI-native research intelligence, built as a five-agent organization.
PRDT turns open-source projects, papers, professional reports and real-time information into structured evidence, Research Graph updates, contradictions, hypotheses, experiments and domain-system updates.
Evidence Pipeline
Sources
Evidence
Graph
Experiments
Open Source
Papers
Reports
Real-Time
Synthesis
Five Agents
Collectors feed the Synthesis Agent. The Synthesis Agent does not skip evidence.
Paper Agent
Moves beyond summary: paper to thesis, method, reproduction, backtest, module candidate and production decision.
- Action: reproduce or reject with reason
- Output: experiment-ready research object
Report Agent
Separates facts, opinions, forecasts, trade ideas, narratives and model output before anything enters research state.
- Action: extract systematic features
- Output: market-view evidence and risks
Real-Time Agent
Ranks events by impact and maps them to affected assets, curves, factors, models and existing conclusions.
- Action: P0 to P3 severity classification
- Output: research-state update proposal
Synthesis Agent
Fuses evidence, finds associations, detects contradictions, generates hypotheses and routes experiments to a domain system.
- Action: hypothesis to experiment queue
- Output: Research Graph and domain updates
Research Graph
Association over summarization. Contradiction is valuable.
PRDT stores papers, repos, reports, events, datasets, factors, models, risks, conclusions, hypotheses and experiments as research objects. Edges capture support, contradiction, dependency, implementation, invalidation and test relationships.
Paper
Repo
Event
Conclusion
Hypothesis
Experiment
First proving ground
PFICC consumes PRDT. PFICC does not embed PRDT internals.
PFICC Pilot
China Fixed Income multi-factor research starts with rates, yield curves, carry, roll-down, momentum, liquidity and regime controls.
Latest v0.1 Run
4 evidence objects, 1 association, 1 contradiction, 1 hypothesis and 1 proposed PFICC experiment from synthetic/internal fixtures.
Human Review
Production strategy, allocation, risk limit, model replacement, external publication and destructive repo changes remain human-gated.
v0.1 status
Runnable architecture, bounded claims.
CLIImplemented
Tests11 unit tests passing
PFICC latestWrites latest.json and latest.md
Live source qualityUNMEASURED
Market impactUNMEASURED