Dr. Nib research agent

Dr. Nib turns a question into a cited report. You describe what you want to know, answer a few scoping questions, set a budget cap in USDC, approve the plan — and it searches the web, reads sources, scores them for relevance and trust, and writes the report. Whatever it doesn’t spend comes back.

Open it at nibgate.xyz/dr-nib/research (wallet or Google sign-in required).

The flow, end to end

composer → intake (batches ≤5) → configure → review → approve → run → report
  topic      scoping Q&A          depth/cap    plan +     escrow?    stages     cited,
  sent       JEV stop/reframe     formats      estimate   then run   SSE live   exports

1. Ask — type the research question and hit Send. The thinking screen opens instantly with your message shown as sent; questions stream in when generated. A project is created straight away, so you can leave and resume later (?project=). Nothing spends yet.
2. Intake, in batches — up to 5 questions per screen (easy choices first), progress shown (“3 of 5 done”), prior batches collapsed with your answers labeled. Answer everything, submit once: one thinking per answer streams back, then you press Next when you’ve read. Just plan it skips the rest any time — unanswered questions stay unanswered, nothing is fabricated.
3. JEV runs the stop rule — after each batch a calibrated decision picks proceed (brief is whole), ask_more (another batch, max 5), or reframe (the angle is dead — the next batch opens a different one, and the rejected direction stays on the record). The model proposes, JEV disposes; fallbacks are labeled, never silent.
4. Configure — depth (quick/standard/deep), budget cap, live-web toggle, report length. Planning is free.
5. Review — sub-questions, uncertainties, cost estimate. Optionally fund an onchain escrow first; approve is then gated on it being Funded.
6. Run — the command-center page: report and sources on the left, the live reasoning feed on the right (tool calls with their results, queries, verdicts, your Q&A transcript), guidance box or mid-run question card pinned at the bottom. Pause/resume any time.
7. Report — versioned, cited, exportable (Markdown, JSON, BibTeX, PDF, Word, Excel, PowerPoint). Finished runs can run again with a new prompt (reprompt creates version N+1, original kept).

How a run spends

Every stage draws from the run’s budget plus a 1% platform fee. The run pauses — never overspends — at the cap; raise it or end it. Search fans sub-questions (keyword-ified) across 11 free indexes plus keyed providers when configured; fetch opens ranked pages (PDFs and Office docs parsed, paywalls and bot-checks reported not worked around); a data stage runs the full agent loop (model proposes, JEV judges, tools execute) reaching for primary APIs, sandbox compute, or paid sources when retrieval is thin; JEV scores every source and decides continue/write/stop per round.

How the agent decides

The data stage is a propose → judge → execute loop with termination enforced in code, not in prompt prose: step/deadline/token/spend budgets, duplicate and stagnation detectors, an evidence checklist the run must cover, and a validated finish tool. Free reads run on deterministic policy; anything moving money always passes JEV, and a silent judge stops the run instead of running judgeless. Full ruleset: Agent spending. Enforcement map: Spending gates.

Machine access

The same runs are drivable by agents over MCP (answer_question, answer_questions, approve_plan, reprompt_run, reconcile_run, …): signed USDC payments for tips, paid unlocks, and x402 resources come from the run budget inside per-call ceilings, and every tool call is logged with what it cost and what came back. See the spending rules below and Budgets & refunds.