Every claim on this site is backed by an artifact this engine generates.

That the engine renders a complete study end to end from a single input, defines the five words that are routinely misread upward, and labels construction figures as job-years rather than positions.

Nothing about any real economy. The inputs are illustrative and the regional table is estimated rather than purchased. It does not prove forecast accuracy, and it withholds the recommendation.

That a summary can compress without becoming a lie: every headline number carries its own qualifier in the same tile, and the fiscal cell states in the cell that revenue is not net.

The same limits as the full report, and it does not stand alone as a decision document. Cell four says so rather than filling the space with a recommendation.

That every number traces to a formula, an input, an evidence tier, and a validation result, all read from the same run the report was written from, so the two documents cannot disagree.

That the arithmetic has been independently verified by a third party. It hands a reviewer what they need to do that; it does not do it for them.

That the pack names a leading candidate and refuses to overstate it in the same breath: it labels the leader provisional, calls the confidence thin, and prints in plain language the condition that could flip the ranking.

That the leading candidate is the right site. The candidates and scores are demonstration fixtures, the gap between first and second is smaller than the pack can resolve, and no parcel-level diligence is carried.

That the audit inventories the claims a reader would repeat, attaches whatever support it detected to each, runs named method gates, and turns the gaps into first-class findings rather than silence. The board memo leads with release status, then the reasons, each written as a question a board member would ask.

That the audited study is wrong. It audits the document as submitted, does not re-run the model, and prints the detector's measured limits next to the counts. Every finding is a reason to ask a question, not a finding of error.

That comparison is treated as a gated claim rather than a charting option. The pack writes the internal work and then withholds client comparison language, because rank stability was not measured. The refusal is the product behavior.

That either study is better, weaker, or more reliable than the other. The inputs are demonstration audit JSON, and the run does not supply the rank-stability evidence that exact order would require.