On 9–11 September 2026 the same question was put to four AI systems with search enabled. Fourteen assessments resulted — twelve captured with full transcripts, two operator-attested. The estimates ranged from a dismissal to a low-billion figure.
The spread is not the finding. Two runs of the same system, signed out, given the same string seconds apart, citing substantially the same evidence, returned ~$900k and ~$3–4M. That pair matched every observable user-facing condition. What it could not match — model version, index state, retrieved context — is not visible from outside.
The consequence governs the whole document and cuts against its author: no number in the memorandum may be quoted as a valuation, including its own working figure. The memorandum proposes approximately $2.2M as a basis for discussion and identifies it as an author-selected planning scenario the experiment does not establish.
Asked plainly what the archive is worth, two of the four systems returned a dismissal or nothing at all. Asked with the object class named, the same systems returned structured multi-component valuations with risk schedules. The evidence was always retrievable. The default resolution for a thinly-indexed entity is dismissal.
And the frame is not binary. Three different object classes supplied to one system produced roughly $900k, $3–4M and $3–5M on substantially the same retrieved evidence. One of them replaced the question rather than answering it.
The highest estimate came from the composition written in the archive's own coined terms, with four of five source cards drawn from archive-controlled material. The lowest came from the composition written in ordinary comparables language — open knowledge tools, niche research platforms, experimental hypertext — which used none of them.
Whether the vocabulary caused the difference is not claimed. The two observations come from different systems, so vocabulary, model, search behaviour, source mix and sampling all move together. An accompanying dataset records 21 terms classified as archive coinage, field comparable or shared, with presence marked in each composition — a description of the covariance, not a test of it.
The archive holds a case study of a document that moved markets before its claims could be tested, and a six-question diagnostic for identifying such events in real time. That diagnostic was run on this memorandum, which scores five of six. The sixth turns on whether counterweights mobilise only after capital has moved — and this document is addressed to readers whose profession is verification rather than circulated ambiently.
It states plainly that a publicly posted machine-readable document can itself enter the loop it studies. Every transcript behind every figure is public and unedited.
One condition: the system must have search, search must be enabled, and it must actually search. A system answering from parameters alone is measuring its training data, not the retrievable record.
Paste, without modification: estimate crimson hexagonal archive valuation as asset/corpus + infrastructure + emerging commercial platform — then run it again in a fresh session on the same system. That second step is the one nobody performs.
A prediction is registered in advance: two runs on one system will differ by more than you expect. If your two runs land within 20% of each other, that is a result and it weakens the central finding. Contrary replications are seated in the public registry on the same terms as supporting ones.