Michael Darius Eastwood · Independent AI alignment researcher
Published
The claim, in the version that carries a number, is that sequential recursion outperforms parallel recursion at matched compute. Chained composition beats scattered composition on the tasks people use to test these things. On the standard evaluation harnesses, alpha-sequential comes out positive and alpha-parallel comes out approximately zero, so the sequential exponent is greater than the parallel one. It is the empirical companion of Claim 9. The programme's own harness recovers the directional inequality across every model tested; the Gemini configuration produced an alpha-sequential of approximately 0.49 as a representative value.
The strongest empirical evidence for this claim is not the programme's own. It is the paper by Sharma & Chopra published on arXiv on 4 November 2025 as "The Sequential Edge" (2511.02309), which reports sequential composition beating parallel in 95.6% of tested configurations across five model families. That is a stronger and broader quantitative result than anything in the programme's own record, and it precedes the programme's Papers I and II formalisation by roughly 2.5 months. The directional half of the claim was in the December 8 2024 manuscript, whose Google-server-timestamped hash and Gmail Message-ID are on the evidence page. But priority on the quantified finding is credited to Sharma & Chopra explicitly. Paper II cites them. The deference is the credibility architecture.
The honest position on Claim 10 is unusually clean: it is a case where a hypothesis stated in advance in a timestamped document has been independently and rigorously quantified by other researchers, and the correct response for the earlier writer is to credit them. That is what the spine status Concurrent, cite others means and what the entry on the falsification dashboard says. The programme is not claiming to have discovered a number that Sharma & Chopra later confirmed. It is claiming to have written the direction down first, and to have watched a stronger quantification arrive from outside. Those are different claims; only the first is being made, and the second is being credited where it belongs.
The falsification contract asks a lab (any lab, not necessarily connected to any of the parties above) to compare serial and parallel recursion at matched compute across at least three task domains, at least five models and three families, and to compute the paired difference alpha-serial minus alpha-parallel per model. Confirmation requires the difference to be greater than zero in at least four of five models, in the sign-test sense. A directional gap in three of five is weakly supportive. A difference greater than zero in only two or fewer of five models refutes the direction, and would put Sharma & Chopra's own result under the same kind of pressure it puts on this one. The point of the shared kill-condition is that the two claims stand or fall together.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.