Method
An orchestrator and a team of researchers.
A single prompt gives you a plausible summary. Verdict works the way an analyst actually works: divide the company into the questions that matter, investigate each one separately, cross-check the claims, then conclude. An orchestrator model runs that division of labour across a team of research agents, so the report reads like reasoning, not a paragraph.
The pipeline
One orchestrator. Ten investigations.
The orchestrator model assigns each report section to its own research agent. Each agent reasons between its tool calls, pulling sources, weighing them, and cross-checking, rather than answering in one pass. It is the same architecture Anthropic published in its multi-agent research benchmark.
The orchestrator splits the company into the report's ten dimensions and briefs an agent for each.
Each agent reasons between its tool calls, pulling sources and weighing them as it goes, not answering in a single pass.
Material claims are cited or flagged as inference. Thin signals are surfaced, not smoothed over.
A ranked risk register and a clear call land on page one, with the evidence behind it.
The lift multi-agent, interleaved-reasoning research showed over a single agent on Anthropic's published research benchmark.
Dimensions investigated separately, one research agent per report section, briefed by the orchestrator.
Of material claims sourced or flagged as inference. Where a signal is thin, the report says so.
The process
From a name to a verdict, in four steps.
What happens between the moment you send a company over and the report landing in your inbox. No call required at any point.
Intake
Tell me what you're deciding
The company name, its website, and the decision you're making. Optionally, the concerns you want prioritized and any documents you can share. Under two minutes, and the only step that needs you.
Multi-agent research
Each dimension, investigated
The orchestrator briefs a research agent per section. Each one investigates its dimension on its own, pulling sources and cross-checking claims against them as it reasons.
Human QA
The high-stakes facts, checked by hand
Before any report ships, the claims that would change your decision, funding, ownership, the top red flags, are verified by hand. The pipeline does the research; a person signs off on what matters most.
Delivery
A PDF you can act on
The finished report arrives as a PDF, verdict on page one, ranked red flags, every claim cited. Five days on Standard, or 48 hours on Express.
Accuracy & sourcing
Sourced, or flagged. Never invented.
A report is only useful if you can trust what's in it, and check it. Every material claim is either cited to a source or marked as inference, so you always know which is which. Where the public signal is thin, the report says so plainly rather than dressing a guess up as a finding. The confidence you read is the confidence that's actually there.
Different from running the AI yourself
A single prompt gives you a plausible summary. This gives you structure, sourcing, and the discipline to look for what's wrong.
Each dimension is researched separately, claims are checked against sources, and the whole thing is built to surface red flags rather than smooth them over. You get a citable artifact, not a paragraph.
Have a company you're evaluating?
Send it over. You'll have a confirmed delivery date within a few hours, and the report by the end of the week, or in 48 hours on Express.
First three reports at $149 · questions: Fabi@lucid-ai.app