Answers are cheap. Evidence is not. A working briefing on turning a hard question into an investigation that keeps a record every claim can be traced to.
Ask a hard question and a confident paragraph arrives in seconds. What does not arrive is what the model weighed, what it skipped, or what it found that pointed the other way.
You cannot see the evidence it used, or the evidence it passed over.
Nothing was committed in advance, so the answer cannot be proven wrong.
A reviewer who was not there is left with only the prose to trust.
Agents are being handed real analysis and real decisions. The better the prose gets, the easier it is to trust an answer that nothing stands behind.
A well-written answer feels right, whether or not it holds up.
Agents now touch research, diligence, and decisions that have to stand up later.
A board, a customer, or a reviewer eventually asks for the basis. Prose is not a basis.
The fix is not a better prompt. It is a different shape of work, one that keeps a record as it goes. One tool built for exactly this is Zorp, a local research agent.
Fluent in seconds. It will not tell you whether to believe it, or what it found that pointed the other way.
Every claim in the answer traces back to the record. It is checkable by someone who was not there.
Before Zorp gathers anything, it commits the hypothesis, the metric, and a kill threshold to a human-readable, hash-verified record. A run cannot quietly rewrite what it set out to test.
A number, set in advance, that says what would prove the hypothesis wrong.
The agent never proposes the threshold. A human does, and only a human can move it. Every attempt is recorded, and when a run crosses the line, the record says why.
Each investigation writes to a six-part evidence record. Measurements are typed values, not narrative, so a claim in the report resolves to a stored row rather than to a paragraph the model wrote about itself.
Validate, investigate, co-write, and deliver. A human checkpoint sits between every step, because a research checkpoint has no safe default. The models can run entirely local, so nothing leaves the machine.
Is the question worth investigating. Scored on redundancy and feasibility, each must cite evidence or score zero.
Gather evidence through staged, pre-registered attempts. One attempt per step, so each surfaces at a checkpoint.
Draft the artifact from the record. The model is handed only what was recorded, and cites only those figures.
Match the finished artifact to the right format and audience.
The model drafts, but a human approves every step and signs the result. The human is always the author of record, never the agent.
Point the loop at a model on your own hardware and the question, the sources, and the record never leave. The same loop runs a lab question and a market question, because it reasons about evidence, not about the subject.
The question, the sources, and the evidence record all stay on the machine you run it on.
It checks evidence, redundancy, and feasibility. Those hold in any field, so one loop fits them all.
Because nothing is sent out, you can point it at the questions you could never hand to an API.
The record turns a confident paragraph into something a reviewer, a board, or a customer can check line by line, long after the run is done.
Every claim traces to a stored row, not a sentence the model wrote about itself.
Someone who was not there can follow the record and reach the same place.
The registered intent and the attempts re-run, so the result is not a one-off.
A killed run is kept, with the threshold it crossed and the reason why.
Do not boil the ocean. Pick a single question where being wrong is expensive, and put a record behind the answer before anyone acts on it.
One where someone will later ask how you know, and the honest answer today is a shrug.
With a human, before any evidence, say what would prove it wrong. Then let the agent look.
Let every claim resolve to a row, and keep the attempts that failed alongside the one that held.
The first investigation shows you how much of your current answer was confidence, and how much was evidence.
A local research agent that turns a question into a preregistered investigation and an evidence record.