GOVERNED RESEARCH SYSTEM
A multi-agent research platform where probabilistic models can propose, critique and rank research candidates while deterministic validation and explicit human authority retain control.
CONTEXT
Quantitative research with multiple AI agents operating under explicit scientific and operational constraints.
Engineering Challenge
Large language models can generate hypotheses, critique ideas and assist research, but probabilistic output must not automatically become scientific truth, empirical evidence or financial authority.
STRUCTURE
The system separates AI reasoning from deterministic scientific execution. Local inference is the default path; an external provider is an explicitly controlled route for independent scientific adjudication when required.
OWNERSHIP
Hands-on architecture and implementation across the boundary between model reasoning and empirical execution.
JUDGMENT
WHY Research generation should not depend on one model family or one reasoning path.
TRADE-OFF / EFFECT Multiple local roles add orchestration complexity while creating a more explicit comparison boundary.
WHY Free-form output and unstable terminal behavior cannot safely mutate research state.
TRADE-OFF / EFFECT The provider path moved behind a Python interface with structured JSON contracts, reducing parsing ambiguity.
WHY Empirical results must remain reproducible and independent of a model deciding how execution occurs.
TRADE-OFF / EFFECT AI contributes research reasoning, while deterministic code owns validation and empirical execution.
WHY Local inference is the default, but independent scientific adjudication may require a separately routed provider.
TRADE-OFF / EFFECT External inference is deliberate and subject to authorization and cost guardrails rather than being an implicit dependency.
WHY Model output may express intent, but it cannot create legitimate approval state.
TRADE-OFF / EFFECT Approval metadata is created and validated by deterministic authorization logic.
AUTHORITY
The system implements explicit boundaries between model intent and system authority.
Propose, critique, rank, interpret and generate research candidates or request allowed actions.
Validation, backtesting, permission enforcement, persistence, approval metadata and controlled state transitions.
Scientific review and high-consequence research progression remain separate, explicit authority layers.
Self-authorization, broker or live authority, capital deployment and uncontrolled consequential state mutation.
RESPONSE
A local model produced a syntactically plausible authorization identifier inside generated tool arguments.
Model-provided identifiers are treated as untrusted input and removed before approval creation. Valid authority remains system-issued metadata controlled by the authorization layer.
An early command-line provider path produced terminal noise, repetition and unstable JSON.
The provider path moved to a Python interface using structured JSON responses, making model integration explicit rather than dependent on general terminal parsing.
A model may express intent or recommendation. It must not mint the authority that authorizes its own action.
EVIDENCE
Evidenced implementation detail is shown selectively. Exact versions not captured by the evidence are intentionally qualified rather than inferred.
| Technology | Version / Model | Role | Evidence basis |
|---|---|---|---|
| Python | Exact runtime version not captured | Orchestration and deterministic system logic | Version qualified |
| Ollama | Exact library/server version not captured | Default local model provider | Version qualified |
| qwen3.5 | qwen3.5:9b | Independent local proposer | Evidenced |
| DeepSeek R1 | deepseek-r1:8b | Independent local proposer | Evidenced |
| Llama 3.1 | llama3.1:8b | Preliminary critic / screening role | Evidenced |
| Granite 3.3 | granite3.3:8b | Research-priority ranking only | Evidenced |
| OpenAI | Provider integration | Controlled external adjudication route | Evidenced |
| GPT-5.4 mini | gpt-5.4-mini | Independent scientific adjudicator | Evidenced |
PROOF
This case study uses a sanitized public boundary. No private project tree, research artifact, strategy result or proprietary logic is linked here.
BOUNDARY
Multi-agent architecture, model roles, local and external provider routing, structured outputs, permission design, deterministic validation, human scientific review, technology identities and abstract failure-response decisions.
Research hypotheses, strategy logic, signal definitions, feature details, formulas, parameters, thresholds, portfolio construction, execution parameters, internal state and private prompts.
Autonomous trading, autonomous capital deployment, broker authority, air-gapped operation, audited model training or fine-tuning, guaranteed performance and production-scale brokerage execution.