Zenodo (CERN European Organization for Nuclear Research)· 2026
Mirror Observerhood Lab V: Recursive Reliability Under Estimator Corruption
- 23citations
- 2026year
Short summary
This study introduces recursive reliability, a system that estimates its own reliability, and finds it outperforms simpler methods under specific failure conditions but not always. The core finding is that recursive reliability is only beneficial when lower-order reliability failures are critical and detecting them is cost-effective.
AI-generated from the title and abstract; the full text is not read.
Key points
- Recursive reliability agents outperform first-order agents under conditions of reliability blindness and estimator corruption.
- Performance degrades when first-order reliability is already truthful or meta-diagnostic costs are high.
- The benefit of recursive reliability is conditional on the viability-relevant nature of lower-order failures and the cost of detection.
- 39,200 deterministic episodes were used to compare four system architectures.
AI-generated from the title and abstract; the full text is not read.
Abstract
This release forms part of the Computational Observerhood Labs of Mirror Programme, Volume I: Observerhood. Lab V tests recursive reliability: what happens when a system’s first-order reliability estimate itself becomes unreliable. It follows Mirror Observerhood Labs I-IV, which established that self-model reliability can matter, that decomposition alone is insufficient, that reliability must be actionable and cost-sensitive, and that reliability-gated repair has a bounded positive threshold region. Across 39,200 deterministic episodes, four architectures are compared: a no-reliability baseline, a first-order reliability agent, a recursive reliability agent and an oracle upper bound. The recursive reliability agent outperforms first-order reliability under reliability blindness, estimator corruption and false-alarm reliability, but underperforms when first-order reliability is already truthful or when meta-diagnostic cost is too high. The central finding is conditional: recursive reliability is not infinite self-reference and it is not always useful. It becomes observer-relevant only where lower-order reliability can fail in viability-relevant ways and where the cost of detecting that failure is lower than the avoided loss. The release includes a standalone paper and a reproducibility package containing the Python implementation, fixed-seed outputs, summary data, recursive-advantage data, figures, requirements, citation metadata and licences.
The authors' abstract, as published at the source. Zenodo (CERN European Organization for Nuclear Research), 2026 · DOI ↗
Continue with a free account
Ask the paper: 3 free questions a day about this paper; save it, get its citation, new summaries every day for your field. Takeaways are Premium.
Continue free on the webSign in with Google or Apple; no card needed. You come back to this paper.
On your phone:
Field: Safety, Risk, Reliability and Quality
Safety, Risk, Reliability and QualityEngineering