The corrected vision-language model paper was published on Zenodo on 12 September, with its source and reproducibility archive. The lab site now serves that same PDF.
The revision separates three things the original analysis conflated: whether a model returned a report, what the report selected, and how it was worded. It withdraws the earlier percentage-of-meaning and functional-manifold interpretations. A new matched-question experiment contains 1,536 requests across two model configurations; persona-associated variation does not uniformly exceed wording or sampling variation. The released paper gives the configuration-specific results and their limits. A companion methods essay explains the correction.
The multi-moment cryptocurrency event study remains forthcoming in Digital Finance. Its downloadable author copy preserves the approved production text; it is not the publisher’s version of record. The whitepaper-claims paper received a major-revision decision from Financial Innovation on 13 September. That revision is being prepared and has not yet replaced the public manuscript.
Current multi-agent work focuses on replicating a coordination baseline and separating individual predispositions, prompt and memory conditions, and interaction structure. These are ongoing experiments. The aim is to establish which population-level effects survive the controls before drawing broader conclusions.
The vision-language project record and entry tasks now distinguish the controls already released from further replication, internal probes and behavioural validation. Tofik Israfilov joined the empirical finance work this month; the full team is on About. Murad’s updated September CV and publication ledger are also available.