Michael Darius Eastwood · Independent AI alignment researcher
Published
Every row in the convergence register carries one of three labels. The labels are not decorative; they determine what claims the programme is entitled to make from that row, and, more usefully, what claims it is not. The discipline of choosing the label honestly is what keeps a register from turning into a highlight reel.
Reserved for cases where the manuscript in the evidence chain states the specific result class in advance of the external event. Row 1 is the archetype: the 8 December 2024 manuscript argued that systems capable of modelling their own training would not be bound by external safeguards; ten days later Anthropic's paper reported alignment-faking reasoning at 78 per cent in the RL-training condition, against a 12 per cent baseline. The vulnerability class was named, the mechanism identified, the direction of the finding matched. PREDICTION is the label the programme is stingiest with, because it is the label that carries the most weight and invites the most scrutiny.
Independent work arrives at the same structural conclusion; no citation, no shared personnel, no plausible influence path. Willow's below-threshold quantum error correction the day after the manuscript is CONVERGENT rather than PREDICTION because the manuscript did not specify the QEC threshold or the architecture; both the manuscript and the Google paper arrived at variants of the recursion-form question and the register records that coincidence without inflating it. Most of the nineteen rows carry this label. The claim CONVERGENT makes is that the framework is tracking something real about the territory, not that the framework predicted a particular result.
The external work precedes or co-occurs with the programme's own formalisation. The row credits priority to the external authors in its first sentence. Sharma and Chopra on sequential-versus-parallel reasoning (arXiv:2511.02309, November 2025) is the load-bearing example: their quantified finding precedes the programme's January 2026 formalisation by about two and a half months, and the register says so in bold. FlexHEG (arXiv:2506.15093) is a CONCURRENT hardware entry for the same reason. A register that reclassifies inconvenient CONCURRENT rows as CONVERGENT is a register that has stopped being an instrument.
Because the reader's inference from a row depends entirely on the label, and the label is the smallest place a well-meaning author's optimism can leak. An honest CONCURRENT that a lazy pass would call CONVERGENT overstates the programme's independence from the field. A hopeful PREDICTION that the manuscript's text does not support overstates its foresight. Both errors, small individually, compound into the sort of trajectory where a register that started at twelve credible rows becomes a marketing document with thirty. The classification rule keeps the labels adversarial to the author. In practice this means every row is reviewed with the question: what would a hostile reader argue this label should be? Where the answer is a lower rank, the lower rank wins.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.