Guides / Software / Assessment errors

Assessment errors

Every model conclusion about a tender is guilty until proven cited.

Language models assessing tenders invent clauses, skip mandatories, contradict themselves, and deliver guesses in the tone of facts. Four checks catch each failure; run on every output, every time. Part of our guide on bid and tender software.

Published: 16 September 2026 · By Bill · Reviewed by Andre · selrob

Takeaways

  • Demand a clause citation for every conclusion. No citation, no conclusion.
  • The four checks: evidence, omission, contradiction, false certainty.
  • Examples below are explicitly constructed to show the failure shapes, not model test results.

What are the four checks?

Evidence: every conclusion must point to a pack location (document, section, clause). A finding without a pointer is a guess with formatting. Omission: verify the model found every mandatory, every date, every insurance figure; models under-report gates because gates hide in schedules. Cross-check against a human mandatory hunt on mandatory criteria. Contradiction: ask the same question twice, phrased differently, and compare, plus check the answer against a second pack section that should agree. False certainty: flag every “must,” “requires,” and number in the output and confirm each against the pack text; confident tone is not evidence.

The broad limits of model assessment are on AI tender assessment; the chatbot question on can you just use ChatGPT. This page is the protocol that makes either usable without trusting either.

What does each failure look like?

Explicitly constructed examples, invented to show the shapes, not outputs of any tested model, version, or prompt:

Failure Constructed example Check that catches it
Invented clause “The contract caps liability at $2m”: no such clause exists in the pack Evidence: no pointer, no conclusion
Missed mandatory Summary omits the ISO 27001 condition buried in schedule 4 Omission: human mandatory hunt in parallel
Contradiction “Insurance required at lodgement” vs “insurance required at mobilisation” in two answers Contradiction: ask twice, compare, check the schedule
False certainty “Payment is 30 days” stated flatly; the schedule says 30 days from accepted invoice False certainty: confirm every number against pack text

What does this mean for the bid/no-bid call?

Model output enters the gate as a draft finding, never as evidence; every cited clause gets human eyes before the call. Budget the verification time in the pilot: the 30-day pilot measures reviewer hours as a first-class cost. A tool whose output needs full re-reading has automated nothing; a tool that survives this protocol has earned its place in week one.

Keep reading

Elsewhere in this guide.

Bid and tender softwareThe landscape this protocol sits in. AI tender assessmentWhat models can and cannot read. A 30-day bid-software pilotMeasure reviewer time as a cost.

Book a call

Bring the last tender you regretted.

In 30 minutes we’ll show you the call selrob would have made, and the working behind it.

Book a call