The Second Pass: Why Good AI Research Should Be Challenged

A source is not the same thing as certainty

Seeing citations is reassuring, but a citation badge alone does not make a claim true.

There are at least four separate questions worth asking:

1. Does the source really exist?
2. Is it the right source for this subject?
3. Does it actually support the particular claim being made?
4. Is there another credible source that contradicts it?

Opis already treats source provenance as part of the research object rather than decorative footnotes. Findings can carry source references, the system retains the grounded sources that were actually available to the run, and the output QA layer checks for problems such as invalid references and suspicious source patterns.

For you as the operator, the behaviour should be equally deliberate.

When a finding matters, open the source.

Do not ask “does this report have citations?”

Ask “would I be comfortable defending this claim to a client using this evidence?”

That is a much higher standard.

The four uncertainty signals worth learning

Opis separates several kinds of uncertainty because they require different responses.

Assumption

An assumption is something the analysis had to treat as provisionally true to move forward.

Example:

“The business appears to target SMEs based on the language and packages displayed on its website.”

That may be a reasonable interpretation, but perhaps you know that 70% of its revenue actually comes from enterprise contracts.

That is not a web-search problem. It is a context problem.

Ambiguity

An ambiguity exists when the available material leaves more than one reasonable interpretation.

For example, a supplier says “UK delivery” on its homepage but a shipping page only names mainland England and Wales.

The information exists. Its meaning is unclear.

Contradiction

A contradiction is stronger. Two pieces of evidence point in incompatible directions.

One company page may call a product “cancel anytime” while current terms describe a 12-month commitment.

That should not be silently averaged into a confident conclusion.

Gap

A gap means the research does not currently contain enough evidence to answer something important.

Perhaps wholesale minimum order quantity is never disclosed. Perhaps no credible source establishes how long a supplier has been trading. Perhaps the competitor's current pricing simply is not public.

A gap is not a failed report.

It is a map of what needs to happen next.

ICC: correct the interpretation without starting again

Interactive Contextual Correction—ICC—is the part of Opis designed for moments when the public evidence is not the whole story.

The important idea is that your intervention has a type.

Add context

Use context when you know a relevant fact or constraint that Opis could not reasonably know.

“Our sales team confirmed that this supplier will manufacture private-label orders from 500 units.”

Add an instruction

Use an instruction when you want the analysis to prioritise or handle the existing information differently.

“For this decision, delivery reliability matters more than headline unit price.”

Ask a question

Use a question when you want to interrogate what is already present.

“Does the current evidence establish that the warranty applies to commercial customers?”

Questions are particularly important because Opis should not manufacture an answer merely because you asked.

If the existing material cannot answer the question, the correct outcome is a gap.

That gap can then become a research target.

ICC and new research are deliberately different operations

This distinction is subtle, but it is fundamental to how the system is intended to behave.

Suppose your first competitor analysis says:

“Competitor A primarily serves consumers.”

You know from a recent sales conversation that the company pivoted six months ago and now focuses on office contracts.

You add that through ICC as authoritative context.

Opis can revise the interpretation using what you supplied. It does not need to pretend it independently discovered your private knowledge on the web.

Now suppose you ask:

“What percentage of its business is corporate?”

Neither the original evidence nor your clarification contains that number.

The honest answer is not a generated percentage.

It is a gap.

Opis's gap-filling path can take unresolved gaps and perform targeted grounding searches for new evidence. Newly found material is appended to the research rather than being silently confused with the information from the first pass.

Think of the two loops like this:

ICC asks: “Did the system understand the situation correctly?”

Gap filling asks: “Do we need more evidence?”

You often need both.

A practical example: supplier research

Imagine you are sourcing a product for a client.

The first pass identifies four suppliers and assesses areas such as price, delivery time, reviews and time in business. Sources are attached to the relevant findings where available.

Supplier B ranks highly.

Before recommending it, you notice two issues:

• the apparent price is taken from a retail product page;
• minimum wholesale quantity is unverified.

Instead of accepting the ranking, you interact with it.

First, add context:

“We require wholesale supply for recurring orders of approximately 2,000 units per quarter. Retail pricing is not decision-useful.”

Then ask:

“Do the current sources prove Supplier B offers wholesale terms at this volume?”

If they do not, that becomes a gap worth researching.

Once better evidence or authoritative user context changes a criterion used by a scored research mode, the deterministic scoring layer can recompute the result.

That is a very different workflow from asking a chatbot to “have another go.”

You know what changed and why.

Why the refinement trail matters

Business research changes.

People learn new facts. Websites update. Clients correct briefs. Assumptions that looked reasonable on Monday become obviously wrong on Thursday.

A refinement trail gives those changes a history.

Instead of treating the newest prose as if it appeared from nowhere, Opis can retain the sequence of clarification and refinement that produced it.

For consultants and agencies, that is useful operationally.

Imagine a client asks:

“Why is this recommendation different from the draft you showed us yesterday?”

A poor answer is:

“The AI changed it.”

A professional answer is:

“Yesterday the analysis assumed their service area stopped at West Yorkshire. You confirmed this morning that they now serve the North West as well. We incorporated that information and recalculated the relevant comparison.”

That is what a research process sounds like.

Don't correct everything

ICC is not an invitation to force the research to agree with you.

If a result challenges your prior belief and the evidence is strong, changing the context until you receive the answer you wanted defeats the purpose.

A useful discipline is to distinguish three states:

• I know this is wrong — correct it and state the basis for your knowledge.
• I suspect this is wrong — ask a question or create a research gap.
• I dislike this conclusion — leave it alone until you have evidence.

That boundary is especially important in consulting work. Human expertise improves AI research only when the human is also willing to be challenged.

Your second-pass exercise

Take one completed Opis report and do not rerun it immediately.

Instead:

1. Pick one conclusion that would materially affect a decision.
2. Open the sources attached to it.
3. Read the assumptions, ambiguities, contradictions and gaps around the report.
4. Add one piece of context you know to be true.
5. Ask one question whose answer you genuinely do not know.
6. If the current research cannot answer it, allow it to become a gap.
7. Use targeted gap research where appropriate.
8. Compare the revised result with the original.

The objective is not to produce more text.

It is to improve the quality of the decision.

The principle behind the feature

Good AI should not demand that humans surrender judgement.

And good human judgement should not demand that evidence surrender to opinion.

The most useful arrangement is a controlled conversation between the two.

Opis provides the research structure. Public evidence provides an external view. Your experience provides context the web may never contain. The uncertainty layer tells you where the joins are weak.

The result can then improve instead of merely regenerate.

That is the difference between asking AI for an answer and working with an intelligence process.

Next in the series: how individual research modes become much more valuable when evidence and context travel through a workflow.