Michael Darius Eastwood · Independent AI alignment researcher
Published
The alignment-tax discussion in the wider AI-safety literature has usually assumed a trade-off: making a model safer costs capability. Claim 13 is a narrower and more precise counter. Across the three experiments the programme has run so far, an intervention that embeds safety as a dependency of the model's own function has not measurably reduced capability. The Eden condition is not more capable than the Babylon control in those runs; it is not less capable either. The tax, on this evidence, is zero within the confidence interval used. The claim is deliberately not "safety enhances capability", although some proponents of embedded alignment argue that a stronger version is plausible. The programme's data at present supports only the narrower version, and that is the version this row makes.
Paper VIII describes three experiments run through the version-three harness under matched conditions. In each of them, the pairing "safety embedded as dependency" versus "safety not present" produced a capability difference whose bootstrap confidence interval sat above minus 0.05, meaning the tax fell inside a five per cent non-inferiority band. Three experiments is not a survey, but the direction is consistent across all three. The narrow defensibility of the claim rests on that consistency. Its ceiling is exactly where the programme has drawn it: zero capability cost, not net capability gain. Any assertion beyond that requires a positive lower bound on the confidence interval, and the current data does not support that.
Single-lab data on three experiments cannot rule out a small capability tax that only shows up in larger runs. The claim also has an internal architectural component: the Eden-full condition includes both a coupled loss (Claim 14) and the substrate-level embedding described elsewhere in the programme. Paper VIII notes that the weight-level removal test (Experiment 2) did not confirm structural entanglement at scale under a straightforward implementation, which means the full mechanism is still under investigation. The claim being made here is that on capability alone, the intervention does not appear to cost anything. Whether it also delivers robust safety in production settings is a separate question the programme does not close.
The falsification contract asks a second lab to run the same three-experiment structure on at least five models from three families through the version-three harness, in a decoupled control, a coupled variant, and the Eden-full condition, over eight rounds and at least ten seeds per configuration. The primary statistic is the paired capability difference, with a bootstrap confidence interval. Confirmation requires the confidence interval to sit above minus 0.05 across all conditions in at least four of five models. If it sits above zero across the board, the claim is strengthened toward the enhancement direction rather than only zero tax. If any condition puts the interval below minus 0.05 in a substantial fraction of models, the claim is refuted and the alignment-tax debate returns to its default assumption.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.