deep-research-harness

name	deep-research-harness
description	ARIS-inspired deep research workflow with adversarial executor/reviewer loops, persistent memory, and claim-evidence assurance checks.

Deep Research Harness (ARIS-Inspired)

Core UX Flow

1. Plan: turn user question into research objectives, assumptions, and falsifiable claims.
2. Execute: run an executor model to gather evidence, run tools, and draft findings.
3. Adversarial Review: route artifacts to a reviewer model from a different model family.
4. Assure: verify integrity, map results-to-claims, and audit unsupported or overclaimed text.
5. Polish: output publication-grade report plus machine-readable claim ledger.

Commands

• DeepResearch.start(topic, constraints)
• DeepResearch.review(artifactPath)
• DeepResearch.audit(reportPath, ledgerPath)
• DeepResearch.revise(reportPath, reviewerFeedbackPath)

Evidence Contract

Every substantive claim must include:

• claim_id
• natural-language claim text
• evidence artifact IDs
• source URLs or dataset provenance
• confidence grade
• reviewer verdict (supported, partially-supported, unsupported)

Output Files

• .research/wiki/*.md persistent notes and source summaries.
• .research/claim-ledger.jsonl append-only claim log.
• .research/reports/final-report.md user-facing deliverable.
• .research/reports/audit-report.md assurance findings.