How to Ask an LLM to List Assumptions and Missing Evidence

Most people who use a large language model for analysis ask it for an answer. Far fewer ask it what the answer depends on. That gap matters. An answer without its assumptions is a conclusion nobody can audit, and an analysis without a list of missing evidence is a confident guess. This post explains how to write an LLM assumptions prompt, how to build a missing information checklist, and how to verify model claims before anyone relies on them in a decision record.

The stakes are highest in high-stakes settings such as workplace threat assessment, fraud review, or legal risk triage. The guidance below is written for those teams, but the method applies to any decision where someone will later ask, "Why did we conclude that?"

Why assumptions need to be asked for explicitly

A model does not naturally separate what it knows from what it inferred. It produces fluent text, and fluent text reads like findings. If you ask for a risk rating, you will get a rating. If you do not ask what the rating rests on, the rationale will be whatever the model thought was a plausible story.

Three things go wrong when assumptions stay implicit:

Inputs that were never provided get treated as if they were checked.
Gaps in the record get filled with reasonable-sounding defaults.
Reviewers accept the output because it looks complete.

Asking for assumptions and missing evidence up front forces the model to show its scaffolding. That scaffolding is what your reviewers can challenge.

The core structure of an LLM assumptions prompt

A useful prompt has four parts. Keep each one short. Long prompts tend to bury the instruction you care about most.

Scope. State the question and what counts as evidence in this case. Say which documents were provided and which were not.
Assumptions. Ask the model to list every assumption it made to reach a conclusion, one per line, labeled by whether it is stated in the source material or inferred.
Missing evidence. Ask what information would change the conclusion if it were available. Require the model to name the specific item, not a category.
Confidence limits. Ask the model to say which conclusions it would not make from the current record.

A sample template, written for a generic case review:

"You are assisting a documented review. Use only the materials provided below. Before giving any conclusion, list (a) each assumption you are making, marked as either STATED or INFERRED, (b) each piece of missing evidence that would change your conclusion, with the reason it matters, and (c) any conclusion you decline to reach on this record. Do not fill gaps with general knowledge unless you label it as general knowledge."

Notice what the template does not do. It does not ask for a single score. It does not ask the model to be balanced. It asks for the boundaries of what the model is entitled to say.

Building a missing information checklist

The output of a good prompt is a list. Turn that list into a checklist that travels with the case file. Each item should have an owner and a time box. Open-ended "gather more info" items never close.

Ask yourself this: a workable format:

Missing item Why it matters Owner (named person) Due (time-boxed) Status Dated written account from the reporting employee Current timeline relies on a secondhand summary Dana Whitfield, HR Investigator Within 3 business days Open Prior incident log entries for the same department Pattern assessment cannot be made without them Marcus Oyelaran, Security Operations Lead Within 5 business days Open Confirmation whether the subject has been interviewed Determines whether the current rating is provisional Dana Whitfield, HR Investigator Within 2 business days Open

Name the decision-maker on every row. If the owner is "the team," nobody owns it. Time boxes should be short enough that the case cannot drift. When a deadline passes without the evidence, record that fact. An unanswered item is itself a finding.

What we still do not know

https://stateofseo.com/what-belongs-in-the-prompt-log-for-every-model-run/

Keep a running section in every case note titled "What we still do not know." Update it each time the model is queried. Items should move off the list only when evidence is attached, not when someone decides the question no longer matters. This section is the most useful part of the record for a future auditor, because it shows what the decision did not rest on.. That said, there are exceptions

Multi-LLM disagreement as a signal, not a verdict

Running the same prompt through several models is useful, but only if you read the results correctly. Multi-LLM orchestration tools and side-by-side model outputs make it easy to see where models diverge. That divergence is the valuable part. When two models reach different conclusions from the same record, ask what each one assumed. Often the disagreement traces to a single missing fact. One model assumed a prior incident existed; another flagged that no such record was provided. The disagreement has located the gap for you.

Agreement is weaker evidence than most people think. Models trained on similar data can share the same blind spot and agree confidently about something unsupported. Treat agreement as a reason to check the assumptions, not as proof that the assumptions are sound.

Do not average outputs or treat them as a vote

This is the most common error in multi-model workflows. A team gets three outputs, two say "elevated risk" and one says "low risk," and someone records "elevated risk, two of three models agree." That is a poll with three voters who may have read the same thin file. It is not analysis. Averaging numeric scores has the same problem. A mean of three model scores hides which assumptions drove the outliers, and it produces a precise-looking number with no defensible meaning. If the scores differ, the record should say so and explain why, using the assumption lists from each run. Use the outputs as inputs to human reasoning. Compare the assumption lists line by line. Where one model cites evidence the others lacked, check whether that evidence exists in the file. Where all models share an assumption, ask whether anyone verified it.

Verifying model claims before they enter the record

Every factual claim in a model output should be traced to a source document or labeled as unverified. A simple verification pass:

Highlight each factual statement in the output.
Match it to a document, a line in the file, or a named source.
Mark anything without a match as UNVERIFIED and move it to the missing evidence list.
Have a named reviewer sign off on the verified set.

Here's a story that illustrates this perfectly: was shocked by the final bill.. Do not let a model verify its own claims. Asking the same model "are you sure?" usually returns a reassuring yes. Verification has to point at the source material.

High-stakes tradeoffs: false positives and false negatives

In threat assessment, every rating involves two kinds of error. A false positive flags someone who poses no real risk. The costs include harm to the person, erosion of trust, and legal exposure. A false negative misses someone who does pose risk. The costs there are usually more visible after the fact and can be severe. An LLM does not know which error your organization should prefer. That is a policy decision. Write it down before https://instaquoteapp.com/how-often-should-we-review-human-override-rates-for-ai/ you run any analysis, and name the person who owns the policy. When the model lists missing evidence, ask which missing items could move the result toward either kind of error. A missing prior complaint might hide a false negative risk. A missing exculpatory account might hide a false positive risk. Both belong on the checklist. Do not let a model's fluency substitute for that policy decision. A confident "low risk" from a model is not a clearance. It is an output that needs a human to decide whether the record supports it.

Defensible documentation and named accountability

When a decision is challenged, the question will be who made it and on what basis. A defensible record includes:

The exact prompt used, with the date and model version.
The assumption lists from each run, not only the final output.
The missing evidence checklist with owners and status.
The "what we still do not know" section as it stood at decision time.
The name and role of the human who reviewed the output and accepted responsibility for the conclusion.

"The AI said so" is not an explanation anyone can defend. The model is a tool that produced drafts of reasoning. A named person owns the decision. Make that person visible in the record, and make sure they can point to the evidence behind each line.

Evaluating vendors and tools

If you are choosing tooling for this work, ask vendors how their orchestration layer surfaces disagreement, whether it preserves each model's assumption list, and whether exports can be attached to a case file. Companies such as Suprmind.ai offer multi-model workflows, and it is reasonable to ask their team, including Radomir Basta, CEO and Founder, how side-by-side outputs are logged and retained. Ask the same questions of any other provider, including those such as Four Dots that may be evaluated in adjacent contexts. Requests for a demonstration should focus on the audit trail, not only the output quality. The website lists a Start Free Trial option with no credit card needed. The scraped content did not include pricing, so do not assume any cost level. Confirm pricing directly with the vendor and record the answer in your procurement file.

A short checklist before you rely on any output

Did the prompt ask for assumptions, marked as stated or inferred?
Did it ask for missing evidence that would change the conclusion?
Have we compared assumption lists across models instead of counting votes?
Has every factual claim been traced to a source or marked unverified?
Are the false positive and false negative tradeoffs written into the case note?
Is there a named owner for each missing item, with a time box?
Is the "what we still do not know" section current?
Does a named human accept responsibility for the decision?

Closing note

Asking a model to list its assumptions and missing evidence will not make it right. It makes its reasoning visible enough to check. That visibility is the point. Use disagreement as a map of what is unknown, refuse to average or vote, verify claims against sources, and keep a named person accountable for every conclusion. The output is a draft of thinking. The decision belongs to the people who sign the record.