EVIDENCE-CONTROLLED RESEARCH AGENT
Let the model choose the path without waiving the evidence rules.
Agent Systems Evidence Scout separates model discretion from deterministic evidence enforcement.
Problem
Research agents can produce fluent answers while citing papers they never actually verified.
Engineering question
How can a model choose a useful research path without controlling whether its own evidence is accepted?
What I built
A bounded arXiv research workflow using smolagents and Hugging Face inference. The model selects research tools; deterministic Python validates canonical arXiv identities, records verified papers for the current run, and withholds drafts that fail the citation contract.
Architecture
Search returns at most five papers. Comparison accepts two to five unique papers. Retrieved titles and abstracts are treated as untrusted research data. A per-run registry separates successful arXiv verification from model assertions, and the final citation gate runs outside the LLM.
Evaluation
The accepted release has 39 deterministic tests, green Python 3.11 and 3.12 CI, source-compilation checks, Chromium smoke coverage, and a successful real Hugging Face plus arXiv integration run.
Failure evidence
The first live run exposed a model-tool parameter mismatch, a blocked JSON import, malformed comparison input, and step-budget exhaustion. The original offline suite did not reveal these issues. The interfaces were corrected and regression coverage expanded without increasing the six-step budget.
Measured result
After the integration repair, the same live scenario completed in four of six available agent steps. This validates the bounded workflow and interface path; it does not establish the truth of every research conclusion.
Current limitations
Evidence comes from arXiv metadata and abstracts, not full-paper analysis. The registry is in memory and scoped to one run. There is no production authentication, persistence, hosted tracing, or deployment. This is a compact work sample, not a production research platform.