Computational Detritivory: A Boundary Report

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Machine-learning development produces checkpoints that are discarded after early stopping, poor optimization, overfitting, or architectural replacement. This exploratory report records six experiments totaling 1,377 recipient-training runs on MNIST-family datasets to evaluate whether reusable first-layer structure can be recovered from abandoned checkpoints. Healthy donors improved Fashion-MNIST recipients, whereas deliberately failed donors provided no average benefit. Across several filter-selection experiments, no hand-designed label-aware, label-free, or target-blind screening rule provided a statistically significant advantage over random filter subsets. In a continuous destructive-optimization sweep across 36 donors, dead-filter fraction correlated with transfer harm, but adding dead fraction to a regression containing source-task accuracy improved explained variance by only 0.0040, indicating that deadness mainly tracked general donor quality. The current evidence supports the reuse of competent obsolete checkpoints, but neither autonomous fragment selection nor useful whole-layer salvage from severely degraded models has been demonstrated. All dataset comparisons remained exploratory and repeatedly queried the official Fashion-MNIST test set. A separate, independent exploratory pilot reproduced a progressive-damage retraining phenomenon in a small MNIST convolutional neural network. An additional fresh-sparse control matched on post-consumption training epochs revealed no consistent history-dependent reorganization advantage, with early performance differences reversing at the highest levels of weight consumption. This manuscript is presented as an exploratory experimental series rather than a settled scientific result or a peer-reviewed claim of novelty, priority, or superior performance. ChatGPT, Gemini, Grok, and Claude assisted with brainstorming, code scaffolding, analysis, and drafting, while the author selected, edited, and remains responsible for all methods, interpretations, and error checks. The account is built on execution logs and result summaries rather than independent rerun trials during drafting, and readers are encouraged to independently verify the experiments using the provided notebook and conversation logs.