Research workflow
Every run follows the same deliberation architecture: clarify the brief, optionally ground it in live web context, generate independent council answers, blind peer-review them with a multi-criteria scorecard, then synthesize one chairman decision brief. This page maps the full path — from your question to the artifact you act on.
Click any step to jump to the detailed explanation. Dashed boxes are optional; solid stages always run when prior steps succeed.
start
Ask your question
Sign in and dispatch one decision-grade research run.
choice
Free or Premium
Pick entitlement and model lineup.
Each run writes progress to Firestore so the UI updates live.
Stage 2 scores every anonymized draft on these dimensions (1–5). Weights are configurable in admin; defaults favor correctness and actionability.
Correctness
Sound reasoning and defensible claims
Completeness
Covers constraints and material angles
Actionability
Clear next steps someone can execute
Risk coverage
Failure modes and edge cases named
Grounding
Uses available evidence; avoids invention
Step 01
Sign in with Google and write one difficult question — a real decision with trade-offs, constraints, and consequences. LLM Peers creates a research conversation: your message plus an assistant placeholder that fills in stage by stage as the council runs.
Step 02
There is no subscription tier. Free Research uses welcome credits, then one run every 24 hours with an efficient multi-lab council. Premium Research draws from a prepaid wallet and unlocks research panels, custom peers, frontier models, and deeper token budgets.
Step 03
Before the council spends tokens, an intake agent may ask 1–5 focused clarifying questions — single choice, multi-select, or short text. Answers are woven into the deliberation brief. If your question is already specific enough, intake is skipped.
Step 04
When web research is enabled, a dedicated search agent collects a current briefing from public sources — market context, dates, disagreements across sources, and markdown source links. That briefing is attached to every councilor's prompt so models deliberate from shared external context.
Step 05
Each councilor model answers alone — no peeking at peer drafts. Prompts ask for decision-grade depth: recommendation, analysis, risks, alternatives, and next steps. Each draft ends with a structured extract (recommendation, key claims, assumptions, risks, actions) used later for agreement mapping.
Step 06
Successful drafts are anonymized (Response A, B, C…). Each councilor reviews every draft against shared criteria, scores them 1–5 on five rubric dimensions, names strongest assumptions and missed risks, then submits an ordinal ranking. Self-votes are excluded from the aggregate by default.
Step 07
A chairman agent receives all council drafts, peer rankings, the deliberation scorecard, and the agreement/dissent map. It produces one structured brief with verdict, why this won, ranked options, where peers agreed and disagreed, assumptions you are accepting, risks, next action, and open questions.
Step 08
The conversation title is generated, status moves to complete, and Premium wallet usage is settled with a receipt. Inspect every stage in the workflow inspector, publish a public research link, or ask a follow-up that builds on prior context.
Progressive persistence — each phase writes to your assistant message as it completes.
| Phase | What persists |
|---|---|
| Intake | questions, answers, skipped |
| Web | briefing, sources[], model |
| Stage 1 | per-model response, structured extract |
| Stage 2 | ranking prose, peer_scores[], parsed_ranking |
| Metadata | scorecard, agreement_map, aggregate_rankings, label_to_model |
| Stage 3 | chairman model, decision brief |
| Billing | cost report, wallet receipt (Premium) |
Email when ready
Leave during a long run — you get a link to the completed research conversation.
Public research share
Publish a read-only case with decision, drafts, scorecard, and sources — without re-running the board.