Papers with their evidence attached: pre-registered experiments, published data, and results anyone can re-run.
The parent study's winning arm still missed a fifth of the corpus. This pre-registered bake-off shows most of that ceiling was a ranking problem (BM25 +21.7 pp on localization) — and retires the shipped fusion default, which failed its own frozen prediction. Verdict: revised, both outcomes at equal prominence; independently validated before writing.
Read on Zenodo →Treats deliberately incurred technical debt as a call option on the validated product: repayment is the strike, paid only if the hypothesis survives, and abandonment is a costless default. Gives the strategic/toxic boundary a checkable cost-loading test, proves a debt-overhang theorem, and derives the refactoring burst practitioners report at product-market fit as optimal exercise. Two primary tests are pre-registered with frozen analyzers before any data contact. With Mohammad Reza Azarang Esfandiari.
Read on arXiv →In preparation, not yet submitted. Listed here so the claim is on the record with its actual status rather than implied by a citation. Manuscripts are shared on request.
A five-plane decomposition and a conditional invariant-preservation result
Separates proposal authority from effect authority in an agent harness, then proves that under complete mediation of a declared effect boundary and a sound commit gate, every reachable protected state preserves the declared invariant — regardless of what the model proposes. The corollary is the useful part: model stochasticity stops mattering to the safety argument once proposals carry no independent commit authority. Study pre-registered before analyzer code was written.
Captured knowledge is almost never reused in a production AI-agent system
The agent-memory literature measures how well agents use prior context on evaluations where the answer-bearing memory is guaranteed present. This measures the step upstream: in a live system, does captured knowledge get surfaced and reused at all? Over an append-only, hash-chained ledger, return is effectively zero. Decomposes that into a funnel, localizes the stage the instrument could not previously see, and pre-registers the intervention.