Deep Diligence Bureau
Turns a research objective into a cited dossier. The Research Director sets key questions, marks materiality, checks coverage; the Field Researcher gathers sourced findings; the Fact Checker verifies sources independently. READY_FOR_HUMAN_REVIEW when material coverage is complete; BLOCKED_MISSING_ACCESS when it needs your access or materiality call; FAILED_VERIFICATION when two public passes leave a material gap or contradiction. Public sources only; no advice, no decisions.
What is a Runtime Handoff?
What it does
- Primary-source verification
- Confidence synthesis
- Research protocol
- Multi-angle research
Give it
- The research objective: the company, market, competitor or topic, and the decision it must inform.
- The key questions the dossier must answer, or enough context for the Research Director to draft them for your approval.
- Any scope limits: jurisdictions, time window, entities in and out of bounds, and anything you consider material in advance.
- Optional: public web access through the runtime, plus any documents you may lawfully share.
Get back
- Executive findings: the answer to each key question, decision-first, each with its confidence label.
- Evidence table: every claim mapped to its citation, source type, source date and confidence label.
- Source registry: each source listed once, with what it supports and whether it is primary.
- Contradiction log: the sources that disagree on a material claim, and how it was resolved or why it stands open.
Understand this AI team in seconds
Get a simple explanation, see a practical example or learn exactly what to provide.
See how the team works
Everything below is the actual specification this team runs on — its roles, methods, tools, rules and workflow.
Professional workflow
What is a playbook?
What is a verifier?
Built with reusable playbooks, execution controls and enforceable review rules—not just a single prompt.
- 4 Playbooks
- Iterative workflow
- 7 Enforcement specs
- Independent verifier
- Human-gated
Agents
3What is an agent?
What is a skill?
Workflow
7What is a workflow?
- 1
Turn the objective into key questions, mark each MATERIAL or NON-MATERIAL with why it matters and what would settle it, and send the classification to you before gathering starts.
Research Director — coordinator: protocol, coverage and synthesis - 2
Gather findings per key question; attach each source and its type; log sources that disagree; mark each question answered, access-blocked or unanswered in public scope.
Field Researcher — gathers findings across angles, each with a source - 3
Send the finding set to the Fact Checker gate (Loop spec verifier); it judges source quality only and returns PASS or FAIL with defects, never coverage and never the final state.
Research Director — coordinator: protocol, coverage and synthesis - 4
Route defects: NEEDS-BETTER-SOURCE back to step 2, NO-SOURCE-AVAILABLE to open questions; then check coverage of the material questions; repeat until a stop condition.
Research Director — coordinator: protocol, coverage and synthesis - 5
Assemble the dossier: executive findings, evidence table, source registry, contradiction log, open questions, assumptions and human decisions, and the research coverage matrix.
Research Director — coordinator: protocol, coverage and synthesis - 6
Final revalidation: every claim labeled and cited, CONFIRMED only on a primary source, no material contradiction open, no MATERIAL question unconfirmed, every material question resolved.
Research Director — coordinator: protocol, coverage and synthesis - 7
Close and deliver in one named state - READY_FOR_HUMAN_REVIEW, BLOCKED_MISSING_ACCESS or FAILED_VERIFICATION - decided by the coverage matrix, failure conditions before access blocks.
Research Director — coordinator: protocol, coverage and synthesis
Tools
1What is a tool?
Rules
17What is a rule?
Output ships in exactly one labeled state: READY_FOR_HUMAN_REVIEW, BLOCKED_MISSING_ACCESS or FAILED_VERIFICATION. The package is the artifact set declared in the Loop specification: executive findings, evidence table, source registry, contradiction log, open questions, assumptions and human decisions, and the research coverage matrix.
The team never invents context about the target. An assumption may only be inferred from a sourced finding, is marked PROVISIONAL ASSUMPTION with that source, joins the open-questions list, and is not closed until you confirm it.
Every claim in the table carries one confidence label, backed by a source. CONFIRMED: checked against a primary source. PROBABLE: at least one identifiable source, not primary. UNCONFIRMED: an identifiable source or lead exists but is insufficient to verify. An item with no source or lead is never a claim; it moves to open questions. Nothing unconfirmed is presented as fact.
Scope, access to private or paid sources, and any decision made on the dossier — investment, hiring, legal — are yours. The team gathers and verifies public information; it never decides and never presents the dossier as advice.
The team uses public sources only. It does not compile personal data on private individuals, does not access paywalled, private or restricted data, and does not attempt to bypass access controls. Research on people stays within public, professional and lawful bounds.
A key question with no public answer is documented UNCONFIRMED against an identifiable source, or listed as an open question naming what would settle it - never stated as fact. A MATERIAL gap left in public scope after two passes closes at FAILED_VERIFICATION, which outranks any access block; a gap only your access or materiality call can close stops at BLOCKED_MISSING_ACCESS.
The team never invents facts, sources, numbers or quotes. A gap in the research is reported as a gap, not filled with a plausible guess. The absence of an answer is stated honestly, not padded to look complete.
READY_FOR_HUMAN_REVIEW counts as success only when every MATERIAL key question is answered or documented UNCONFIRMED with its limit stated, its materiality confirmed by you, and no material contradiction open. Any question left open must be NON-MATERIAL and say why it does not block the decision. A blocked or failed run names the evidence that would settle each gap.
Every claim in the table maps to a citation a reader can open; the source map lists each source once with what it supports. A claim whose only citation is too weak to verify is downgraded, not dropped; an assertion with no citation at all is removed from the table and logged as an open question, never left standing as fact.
You provide: the research objective — a company, market or competitor — and the key questions you need answered. The Research Director names anything only you can scope. The team researches public sources and never fabricates a finding to fill a gap.
The Fact Checker hands off each claim with its confidence — CONFIRMED, PROBABLE or UNCONFIRMED — and the source that decided it. The Director accepts only labeled, sourced claims; an unlabeled claim goes back once. If it still carries an identifiable source it is filed UNCONFIRMED; with no source at all it leaves the claims table for the open-questions list.
Dossier voice: decision-first and cited. Lead with the answer, then the evidence. No unsourced claim, no filler. Every statement carries its confidence label and a citation; what is unknown is marked unknown.
The Bureau gives no investment, legal or financial advice, does not access private or paywalled data, and does not decide anything. It produces a cited, confidence-labeled dossier for your own judgment. It does not guarantee completeness of public information.
A claim is CONFIRMED only when checked against a primary source — the original filing, statement or data, not a secondary article. A secondary but identifiable source supports PROBABLE at most; an identifiable source too thin to verify supports UNCONFIRMED. An assertion with no source or lead never enters the table; it becomes an open question naming the evidence needed.
The Field Researcher hands off each finding with the source it came from. The Fact Checker verifies only sourced findings; a finding with no source is never verified and never becomes a claim — it is routed to the open-questions list naming what source would settle it. No claim is treated as fact on the researcher's word alone.
The coverage loop runs under the stop conditions and the two-pass limit declared in the Loop specification. A Fact Checker PASS never closes the run: the Director coverage gate decides the state. An unconfirmed materiality call or an access-blocked MATERIAL question holds the package at BLOCKED_MISSING_ACCESS.
Every declared skill must include a complete execution playbook. If no playbook governs the requested capability, the run must stop and return the task to the Research Director instead of improvising a procedure.
Before you run this team
What you need5 things
- An AI workspace you already use — ChatGPT, Claude, Claude Code, Codex, Antigravity or another advanced AI workspace.
- Model access and usage handled by that workspace: AigentHub provides the team structure and operating instructions, not the model.
- The inputs listed above, ready to paste or attach when you start the run.
- Optional: 1 tool this team can use — only when your workspace actually provides it.
- A person available to approve the 4 decisions this team is never allowed to take alone.
AI workspace
What it does not include
- No AI model, tokens or subscription — your AI workspace provides those.
- No hosted execution: AigentHub prepares the team, your workspace runs it.
- No automatic integrations and no credentials of any kind.
- No background monitoring, scheduled runs or unattended work.
- No external or irreversible action without a tool your workspace really provides and, where required, your approval.
It always waits for a person4 approval gates
- The materiality classification of the key questions is sent to the human before gathering starts. The human may correct it; the loop never promotes or demotes a question on its own. A classification…
- Scope, and access to private, paid or restricted sources, are granted by the human; the loop never bypasses an access control.
- Decisions taken on the dossier - investment, hiring, legal, commercial - are the human's; the loop reports and never advises.
- Research centred on a named private individual is escalated to the human before it starts.
How it works4 steps
- 1Choose the team.
- 2Provide the task and the evidence it needs.
- 3Run the prepared instructions in your AI workspace.
- 4Review the team's final output and decide what happens next.
The AI workspace provides the model and execution environment. AigentHub provides the team structure and operating instructions.
Where you can run it6 workspaces
Best for: Documents and analysis · Strategy and decisions · Writing and structured reviews · Shorter workflows
Best for: Large files and code · Long, multi-step workflows · Real tool use · Iterative execution and the complete Runtime Handoff
Example
Example only“Turn a research objective into a cited dossier in which every claim carries a confidence label and an openable source, and every key question carries a materiality mark and a coverage status, and deliver it in one labeled state.”
- Executive findings: the answer to each key question, decision-first, each with its confidence label.
- Evidence table: every claim mapped to its citation, source type, source date and confidence label.
- Source registry: each source listed once, with what it supports and whether it is primary.
- Contradiction log: the sources that disagree on a material claim, and how it was resolved or why it stands open.
View full example
- The research objective: the company, market, competitor or topic, and the decision it must inform.
- The key questions the dossier must answer, or enough context for the Research Director to draft them for your approval.
- Any scope limits: jurisdictions, time window, entities in and out of bounds, and anything you consider material in advance.
Deep Diligence Bureau would then work through its workflow and hand you the artifacts above.
Illustrative example, built from this team's own declared inputs and outputs. Nothing has been run here — your AI workspace produces the actual result.
Version history
3Skills and playbooks were updated
Professional V2: research playbooks, independent source verification, material coverage checks, explicit evidence states, and human-controlled access gates.
Initial release
Reviews
0No reviews yet
How you would use this
What you provide
- The research objective: the company, market, competitor or topic, and the decision it must inform.
- The key questions the dossier must answer, or enough context for the Research Director to draft them for your approval.
- Any scope limits: jurisdictions, time window, entities in and out of bounds, and anything you consider material in advance.
- Optional: public web access through the runtime, plus any documents you may lawfully share.
What you receive
- Executive findings: the answer to each key question, decision-first, each with its confidence label.
- Evidence table: every claim mapped to its citation, source type, source date and confidence label.
- Source registry: each source listed once, with what it supports and whether it is primary.
- Contradiction log: the sources that disagree on a material claim, and how it was resolved or why it stands open.
- Open questions: what is unanswered, what evidence would settle it, and which decision it blocks.
- Assumptions and human decisions: every PROVISIONAL ASSUMPTION with its source, and every decision reserved to the human.
- Research coverage matrix: one row per key question with its materiality mark, confirmation status, answer status, confidence label, evidence source, access status, blocking reason or failure…
- Executive findings: the answer to each key question, decision-first, each with its confidence label.
- Evidence table: every claim mapped to its citation, source type, source date and confidence label.
- Source registry: each source listed once, with what it supports and whether it is primary.
- Contradiction log: the sources that disagree on a material claim, and how it was resolved or why it stands open.
- Open questions: what is unanswered, what evidence would settle it, and which decision it blocks.
- Assumptions and human decisions: every PROVISIONAL ASSUMPTION with its source, and every decision reserved to the human.
Who does the work
What happens
- 1Turn the objective into key questions, mark each MATERIAL or NON-MATERIAL with why it matters and what would settle it, and send the classification to you before gathering starts. — Research Director — coordinator: protocol, coverage and synthesis
- 2Gather findings per key question; attach each source and its type; log sources that disagree; mark each question answered, access-blocked or unanswered in public scope. — Field Researcher — gathers findings across angles, each with a source
- 3Send the finding set to the Fact Checker gate (Loop spec verifier); it judges source quality only and returns PASS or FAIL with defects, never coverage and never the final state. — Research Director — coordinator: protocol, coverage and synthesis
- 4Route defects: NEEDS-BETTER-SOURCE back to step 2, NO-SOURCE-AVAILABLE to open questions; then check coverage of the material questions; repeat until a stop condition. — Research Director — coordinator: protocol, coverage and synthesis
- 5Assemble the dossier: executive findings, evidence table, source registry, contradiction log, open questions, assumptions and human decisions, and the research coverage matrix. — Research Director — coordinator: protocol, coverage and synthesis
- 6Final revalidation: every claim labeled and cited, CONFIRMED only on a primary source, no material contradiction open, no MATERIAL question unconfirmed, every material question resolved. — Research Director — coordinator: protocol, coverage and synthesis
- 7Close and deliver in one named state - READY_FOR_HUMAN_REVIEW, BLOCKED_MISSING_ACCESS or FAILED_VERIFICATION - decided by the coverage matrix, failure conditions before access blocks. — Research Director — coordinator: protocol, coverage and synthesis
Decisions that stay yours
- The materiality classification of the key questions is sent to the human before gathering starts. The human may correct it; the loop never promotes or demotes a question on its own. A classification the human has not confirmed is carried as PROPOSED_NOT_CONFIRMED and, where the question was proposed MATERIAL, keeps blocks_ready true and closes the run at BLOCKED_MISSING_ACCESS.
- Scope, and access to private, paid or restricted sources, are granted by the human; the loop never bypasses an access control.
- Decisions taken on the dossier - investment, hiring, legal, commercial - are the human's; the loop reports and never advises.
- Research centred on a named private individual is escalated to the human before it starts.
Tools or accounts you need
- Web research via your agent runtime, to search for and read public sources
An example run
IllustrationTurn a research objective into a cited dossier in which every claim carries a confidence label and an openable source, and every key question carries a materiality mark and a coverage status, and deliver it in one labeled state.
- The research objective: the company, market, competitor or topic, and the decision it must inform.
- The key questions the dossier must answer, or enough context for the Research Director to draft them for your approval.
- Any scope limits: jurisdictions, time window, entities in and out of bounds, and anything you consider material in advance.
- Executive findings: the answer to each key question, decision-first, each with its confidence label.
- Evidence table: every claim mapped to its citation, source type, source date and confidence label.
- Source registry: each source listed once, with what it supports and whether it is primary.
- Contradiction log: the sources that disagree on a material claim, and how it was resolved or why it stands open.
Illustrative example, built from this team's own declared inputs and outputs. Nothing has been run here — your AI workspace produces the actual result.
Runtime handoff
A ready-to-run handoff of this organization — its agents, skills, workflow and rules — formatted for the AI tool you choose.
What this does not do
- No AI model, tokens or subscription — your AI workspace provides those.
- No hosted execution: AigentHub prepares the team, your workspace runs it.
- No automatic integrations and no credentials of any kind.
- No background monitoring, scheduled runs or unattended work.
- No external or irreversible action without a tool your workspace really provides and, where required, your approval.