A human-led agentic system for design research. It turns a brief, selected evidence, and structured exercises into reviewed claims and traceable framing reports for later ideation and prototyping.
- Bounded agents. One analyst interprets each exercise. An optional critic audits evidence, reasoning, and scope without creating or approving claims. Models have no filesystem, shell, web, or ledger-write tools.
- Host-controlled execution. Python handles source selection, typed output validation, budgets, bounded retries, checkpoints, and resume. Configuration lives in config/project.yaml.
- Explicit context. Runs freeze the brief, exercise, assigned evidence, and selected reviewed claims. Models cite passage IDs; Python resolves exact quotations and locations. Evidence directories are never scanned.
- Human judgment before persistence. An independent human frame is frozen before framing runs and withheld from AI. Review records comparison, decisions, disagreement, and the next investigation. Acceptance is separate from factual support.
- Append-only memory. Human review is the only path into the claim ledger. Corrections preserve history; invalid premises block dependent claims. Replay and report compilation are deterministic.
- Separate evaluation layers. Citation checks, coverage, and cost are measured separately from human-assessed grounding and reasoning quality. Structured logs record public rationale and decisions, not hidden reasoning.
brief + exercise + selected evidence/context → analyst → optional critic
↓
independent human frame ─────────────────────→ comparison + human review
↓
append-only claim ledger
↓
reports + next inquiry
- 13 ordered exercises covering framing, audiences and stakeholders, field and secondary research, constraints, opportunities, and project planning.
- CLI workflow:
validate,frame,run,review,eval,compile, and a zero-cost fictionaltutorial. Runs support request previews, optional live calls, checkpoint reuse, and later evidence-based corrections. - Research outputs: comparison packets, review worksheets, evaluation JSON, a design research brief, provenance report, and synthesis index. Factual report sections require human-accepted, evidence-supported claims and dependencies; human framing decisions are labeled separately.
| Component | Current state |
|---|---|
| Core Python runtime | Implemented, with offline unit and integration coverage. |
| Research website | Flask prototype: private projects, drafts, exercise navigation, evidence uploads, review, and report downloads. |
| Cloudflare adapter | Locally tested Workers/Durable Objects/R2 pilot. Production sign-in and paid model calls remain unverified. |
The Cloudflare pilot runs one analysis at a time with a 12 MiB working-data cap across the cohort. Backup/restore and representative load checks remain open; see deployment status. Research quality and learning benefits have not been established by the fictional tutorial or structural checks.
Currently accessible to selected testers at: https://agentic-reframe.drvr-web.workers.dev/