The method
The path we can run with you.
A research job is a decision with a record. These are the five gates, and the mechanics that keep the work inspectable.
-
01
Frame
Turn the decision into a falsifiable question. Name the owner. Write the kill criteria before the work starts.
A decision frame with the owner, the question, and the result that would kill it.
-
02
Discover
Bounded parallel agents gather candidates, evidence, and counter-evidence. Every claim carries a locator.
A candidate and claim table. Claims without a source are labelled as hypotheses or removed.
-
03
Verify
An independent pass re-checks every retained claim against its source. Quantitative claims are recomputed where the data permits.
A verification record. Dropped claims are named, not softened.
-
04
Decide
The named expert writes the conclusion, the confidence, the conditions that would change it, and the cost of being wrong.
A decision memo the client's owner can accept, reject, or send to risk.
-
05
Execute and transfer
The decision enters the client's workflow. The artefacts, prompts, and method are handed over so the team can run the next question.
A handover pack and a workspace the client keeps.
In practice
What it looks like on a live question.
Start where the team already works
An analyst names the question in Slack or the AI chat the desk already uses. Nobody has to learn a new front end.
Run against approved data
Bloomberg, FactSet, internal files, git. The workspace may reach only what the firm approves.
Keep identity and policy on every call
Every job and every model or tool call resolves to the person who started it. Agents inherit that identity and the same allow-list. They do not get a broader service account.
Review before anything publishes
The first artifact is produced for human review. Nothing publishes itself. The signed record can be walked through live.
Keep the work, scope the memory
Artifacts land in workspace memory after review. Default scope is the user. Sharing to a desk, pod, or named team is explicit.
Rank what follows
Leftovers and next phases appear as a ranked queue after the work exists. The analyst picks. The system does not silently start the next job.
The workspace is devBox. Policy is mGateway. The loop is Harness.
Start here
Bring one decision.
A 2 to 4 week sprint on a live decision. You get the decision record, the verification trail, and the method your team can run next.