Superintelligent Software Engineering: Governance That Scales with Capability

The paper’s definitional work is rigorous. The three-regime taxonomy—assisted, delegated, and superseded—clearly separates human-verified, capped-delegation, and outpaced-throughput scenarios, and the escape model in Eq. (2) formalizes why capability alone cannot lower the stale-built escape rate once $r > c$. The author is admirably transparent: the research agenda states falsifiable claims with refutation conditions, every opinion and practice heuristic is explicitly bracketed, and the GitHub snapshot follows a pre-specified protocol with released scripts and sensitivity audits.

The central problem is that OSTEON is proposed but never tested. Section VI-H candidly notes that OSTEON’s own development did not apply the discipline to itself—there are no decision records, no bone repository, and capture stopped weeks before the last commit. The escape model assumes independent staleness and fixed review capacity, assumptions the author admits are violated in practice, and it understates the problem by holding $s$ fixed when agent changes themselves raise staleness. While the 9.5%–22% dangling-reference rate in context files is an interesting symptom, unresolved paths are not proven defects; the leap from unresolvable paths to harmful agent behavior remains speculative. Because the seven requirements were co-designed with OSTEON rather than derived independently, they risk rationalizing the proposed solution rather than grounding it in comparative evidence.

The positioning against related work is fair and precise. The paper distinguishes itself from SE 3.0 and agentic software engineering by focusing on the superseded regime where humans cannot read every change, and it correctly cites Gloaguen et al. to support the claim that ungoverned context files do not generally improve task success while raising inference cost by over a fifth. The comparison to ContextNest and Deliberative Curation is nuanced, acknowledging complementarity. However, there is no comparative evidence that OSTEON outperforms simpler baselines—such as better-maintained context files, stricter review caps, or lightweight linting—and while the paper frames Gloaguen et al. as supporting the currency hypothesis, that causal interpretation is explicitly marked as the author’s own addition rather than the study’s finding.

Reproducibility is unusually strong for the exploratory measurement: the protocol, collection scripts, dataset, and sensitivity audit are all released as ancillary files, and the per-file statistics can be recomputed from the released artifacts. OSTEON itself is archived on Zenodo with a DOI and an open license. However, the core claims about governance scaling cannot be independently reproduced because no evaluation protocol, benchmark results, or effect sizes are provided; the paper explicitly reports no measured effect of OSTEON. The chief barrier is the absence of an empirical evaluation against the falsifiable questions Q1–Q6, which the paper honestly states as future work.