EA-VALUATION-MEMO-01, deposit #1607 of the Crimson Hexagonal Archive (10 September 2026), is an investor-facing valuation memorandum whose evidence is eight machine valuations of the archive produced within one day, and whose strongest finding limits what the memorandum itself may claim.
On 9-10 September 2026 one string was put to four AI systems with search enabled. The resulting estimates span roughly three orders of magnitude, from a dismissal to a low-billion figure. Two of the runs are the same system, signed out, given the same string seconds apart and citing substantially the same evidence; they differ by about 3.5x. The memorandum takes that pair as its governing constraint: a machine valuation is a draw from a distribution, and reporting one as a figure states a sample as though it were an estimate. It applies that rule to itself, stating that no number in the document should be quoted as a valuation, including its own working figure.
Two further findings are ranked below it and marked as weaker. Supplying the object class in the prompt is the difference between a structured valuation and nothing, and the frame is not binary: three frames on one system produced three distinct figures on the same retrieved evidence, one of them replacing the question rather than answering it. Valuation also covaried with the vocabulary of composition, the highest figure coming from the composition written in the archive's own coinages. The memorandum explicitly declines to call that causal, because nothing in it controls vocabulary.
The working figure of approximately $2.2M is identified as an author-selected planning and negotiating scenario which the experiment does not establish, drawn from the one assessment that inspected the live assets and enumerated its discount schedule — and flagged as an informed follow-up rather than an independent check, since that assessment was supplied the memorandum containing the earlier valuations.
The document preregisters a natural experiment around a first transaction, specifying prompts, systems, run counts, dispersion measures and seven possible outcomes including three unflattering ones, and states the confound it cannot remove. It also names and refuses the pathological version of its own mechanism, committing to publish any transaction as a structured event rather than as a valuation claim: the event is evidence, the valuation is the reader's.