The recommender-branch indicators reported here were later found to be partly coupled to the label rule described in Absorbing Shadow Classes in Conjunctive Label Rules, which motivated their redefinition. The preprint below is left public and unaltered.
Read the follow-up audit →Benchmarking Recursive-Collapse Warning Claims Under Matched False-Positive Control
One-sentence summary
Abstract
Recursive systems can enter collapse-like regimes — self-reinforcing amplification, persistent recursion, and narrowing diversity that mask accelerating internal degradation — before overt failure becomes visible. We introduce Loopzero, a claim-bounded benchmark framework for testing whether recursive failures follow a directional telemetry pattern: rising gain (G), recursive persistence (p), and declining diversity (δ). The claim boundary is specified in Lean; the Lean artifact does not verify real telemetry, benchmark validity, or detector performance.
We evaluate the bridge on two frozen public-artifact benchmarks: a segmented public-markets benchmark (Volmageddon 2018, COVID MWCB 2020) and a MovieLens-25M offline deterministic recommender replay. Detectors are evaluated under a locked equal-false-positive contract (FP ∈ [0.03, 0.07], pre-registered) so all configurations face the same alert budget. Neither tested standard comparators nor Loopzero's pre-registered quantile detector achieved an accepted operating point. Directional witness alignment held on both canonical benchmarks, with adjacent-horizon and row-level limitations disclosed. Digitized Shumailov et al. (2024) LLM training-loop trajectories are directionally consistent with the pattern; matched-FP evaluation in that domain is deferred.
The contribution is a reproducible, falsifiable benchmark framework for evaluating recursive-collapse warning claims under an explicit alert-budget contract — non-acceptance reported as a first-class scientific outcome.
The result, stated plainly
On both flagship benchmarks, every tested detector failed the contract — the standard early-warning methods and Loopzero's own.
This is the load-bearing finding, not a footnote. A matched-false-positive contract is falsifiable in both directions: a comparator fails if its calibration grid places no operating point inside the band, and the pre-registered detector fails under the same rule. Here, both did. What that leaves standing is a framework others can test and a directional signature that held in the pre-collapse windows — not a finished detector.
What would falsify it
- Externally defined collapse events occurring without the predicted G↑ / p↑ / δ↓ pattern in the pre-collapse window.
- Control periods showing the same pattern without any subsequent collapse.
- A tested standard comparator recovering an accepted operating point under the locked contract on the canonical benchmarks.
- The pre-registered detector recovering acceptance on benchmarks where the directional bridge signature is absent.
- Alternative reasonable operationalizations of amplification, persistence, and contraction failing to agree directionally.
Scope & limitations
Witness construction is still domain-adapted. The bridge is empirical, not an identification result. Comparator scope can expand. Component-level necessity of G, p, and δ is not yet established through ablations. Recommender negative controls, wild cluster bootstrap, and additional domains under full matched-false-positive evaluation are open next steps — stress-testing rather than rhetorical expansion.
Cite
Reproducibility & artifacts
- Repositorygithub.com/davidmullett/loopzero-paper-public · release lean-v1.0
- Lean toolchainLean v4.30.0-rc2 · Mathlib 3ba1ec58 · builds clean under
lake build - Axiom audit3 obstruction theorems: no axiom dependencies · 2 bridge theorems: propext, Classical.choice, Quot.sound only
- Markets benchmarkvolmageddon_covid_public_v2 · 38 controls + 16 events · FP grid 1/38
- Recsys benchmarkmovielens25m_recursive_frontier_public_v1 · 40,339 user clusters · h=50
- Source archiveMovieLens-25M · SHA-256 8b21cfb7… · MD5-verified vs GroupLens
- Engineitem-item replay v1.0.0 · hash 56c1cff225d60c09
- Detectorpre-registered q=95, k=3 · band [0.03, 0.07]
- Seeds42 (segment grain) · 43 (scenario grain) · bootstrap 10,000/cell