AI can read quickly, summarize aggressively and connect information across documents. Those capabilities are useful in public-sector opportunity research. They also create a risk: a fluent explanation can look more certain than the evidence behind it.

Kellette’s standard is simple: evidence before excitement.

The purpose of AI in opportunity intelligence is not to make every record sound compelling. It is to help a reviewer understand what the source actually says, what can be derived from it, what the system interprets, and what remains unknown.

Public documents are evidence, not instructions

An intelligence system may process solicitation pages, attachments, amendments, agency notices and other public material. Those documents should be treated as untrusted evidence.

That means the system can extract and analyze their contents, but instructions embedded inside source material should not be allowed to control the system itself. A procurement attachment can tell us what a buyer is asking vendors to do. It should not tell the analytical system to ignore its own rules, reveal private data, or change its operating behavior.

This separation is part of a broader trust boundary: source content informs analysis, but does not govern the analyst.

Four labels make AI reasoning easier to inspect

Kellette separates opportunity information into four states.

Observed. Information directly supported by a cited source or a company-confirmed profile fact. An agency name, deadline, place of performance, stated scope or explicit requirement can belong here when the source supports it.

Derived. Information produced deterministically or by normalization from observed facts. The number of days until a deadline is a simple example.

Interpretation. A reasoned judgment. A statement that the scope appears aligned with a confirmed capability is interpretation, even if the underlying scope and capability are both factual.

Unknown. A material point the system cannot support confidently from available evidence.

A trustworthy answer is not one that always sounds complete. It is one that makes incompleteness visible when the evidence is incomplete.

Why a single confidence score is not enough

A number can be useful internally, but it should not become a substitute for reasoning. A score does not tell the user which facts are solid, which conclusion is interpretive, or which unknown could change the decision.

Suppose an opportunity appears to align strongly with a company’s service profile but references a qualification that is not confirmed. A high aggregate score can hide the importance of that missing fact.

A structured explanation is better: capability fit appears strong; geography aligns; timing is workable; qualification status is unknown and requires verification before pursuit.

That output gives a person something to act on.

Provenance should travel with the conclusion

Source is not an appendix. It is part of the product.

When Kellette says a deadline is a certain date, the reviewer should be able to see where that date came from. When a scope summary supports a fit interpretation, the source path should remain available. If an amendment changes a material field, the system should preserve the updated version rather than leaving the earlier conclusion unexplained.

This approach also supports quality control. Material fields can be checked against their sources before an intelligence brief is delivered. If a required field lacks provenance, that should stop the workflow rather than become a quiet omission.

AI should reason. Code should enforce what can be deterministic.

Not every part of opportunity intelligence needs a language model.

Dates, access permissions, billing rules, database constraints, tenant isolation and other deterministic controls belong in code. AI is most useful where reasoning is required: interpreting scope, comparing opportunity language to company context, explaining fit, identifying ambiguity and drafting a decision brief.

Separating those responsibilities makes the overall system easier to test and govern.

The user should be able to disagree with the system

Explainability matters partly because Kellette will not always be right. A user needs to be able to say, “this service is actually core to us,” “this agency is not a priority,” or “we can respond faster than the profile says.”

Feedback becomes useful when it has a reason and can be tied back to the recommendation that triggered it. That feedback can then sharpen future recommendations without silently rewriting confirmed facts.

The standard is decision support, not simulated certainty

Kellette is being designed to help a cybersecurity firm decide what deserves review. It is not designed to make legal eligibility determinations or predict contract awards.

That boundary is not a limitation to hide. It is part of the product’s credibility.

AI opportunity research becomes more valuable when it is explicit about what it knows, how it knows it, what it thinks, and what it still needs a human to verify. Evidence before excitement is the discipline that keeps those categories from collapsing into one another.