Spoiled Deliveries as a Hidden Driver of Food Rescue Volunteer Loss

Volunteer food rescue platforms send ordinary people to pick up surplus food from grocers and restaurants and drive it to shelters. The literature on these platforms treats every rescue as the same kind of job: a pickup with a deadline. Plant physiology says that is wrong for fresh produce. How fast a crate of lettuce or ripe fruit falls apart depends heavily on what kind of produce it is and how hot it gets, and the relationship is not linear. This proposal tests a simple, falsifiable claim. Some volunteers quit because their trips were futile: they arrived to find wilted or rotting produce that the shelter did not want. The risk of such a trip can be predicted before dispatch from three things, namely the produce type, the hourly outdoor temperature, and how long the food waited. If the claim holds, platforms get a new lever. They can notify volunteers earlier for heat-exposed perishables, which saves both the food and the volunteer.

The claim combines the four directions in the brief, but they play different roles. Fruit perishability supplies the mechanism: a produce-specific "heat dose" that estimates accumulated spoilage. Volunteer retention is the outcome. Keyword co-occurrence analysis, the bibliometric technique of linking terms that appear together, is the measurement lens for turning thousands of free-text volunteer comments into auditable spoilage and frustration signals. Small-sample validation is the discipline the study imposes on itself. Expert-labelled comments and spoilage complaints are rare, so the study borrows the discovery-then-independent-validation design of biomarker studies rather than reporting a single fitted model.

What exists and what does not

Four papers on the same Pittsburgh platform, 412 Food Rescue, frame the gap. Improving Efficiency of Volunteer-Based Food Rescue Operations predicts whether a rescue will be claimed without a dispatcher stepping in, using a stacked Gaussian-process and random-forest model on 4,574 rescues with weather attached, and tunes notification timing. It never asks what happens to the food or the volunteer after the claim. Bandit Data-Driven Optimization for Crowdsourcing Food Rescue Platforms and Contextual Budget Bandit for Food Rescue Volunteer Engagement learn whom to notify under sparse data, fatigue and regional fairness. In both, perishability is at most an exogenous deadline. RescueLens uses a prompted GPT-4o mini to sort volunteer comments into seven issue types, including Inadequate Food, and ranks problem donors. It does not link those complaints to temperature, produce type, or later volunteer behaviour.

These four papers motivate the proposal, because they establish the data, the feedback stream and the retention problem. Online Fair Allocation of Perishable Resources constrains it. That paper shows fair allocation breaks down when items perish faster than recipients arrive, but it models decay as abstract geometric loss. Our heat-dose model makes that decay rate concrete and weather-dependent, so hot days become the regime where its condition is likely to fail.

The physiology papers supply the functional form rather than specific numbers. Respiration, ethylene signalling and oxidative damage all speed up with temperature and differ by commodity. That is shown in Effects of Modified Atmosphere Packaging on Postharvest Physiology and Quality of 'Meizao' Sweet Cherry, Transcriptome profiling reveals the roles of pigment mechanisms in postharvest broccoli yellowing, Selenium Treatment Regulated the Accumulation of Reactive Oxygen Species … in Postharvest Broccoli, and the ethylene-absorbent work of Chachin and Iwata. The file contains no rate constants, so the rates are estimated from the platform's own data. The physiology fixes only their structure and their ordering across produce classes.

Core Claim

For each rescue $r$ of commodity class $c$, the study defines an accumulated heat dose:

$$E_r = \sum_{t} \Delta t \; Q_{c}^{(T_t - T_0)/10},$$

where $t$ runs over hours from donation posting to delivery, $\Delta t$ is the hour length, $T_t$ is the outdoor temperature in hour $t$, $T_0$ is a reference temperature, and $Q_c$ is how much faster class $c$ deteriorates per 10°C rise. This is the standard "Q10" temperature-coefficient form used for respiration rates.

The claim has three parts. First, $E_r$ predicts food-quality complaints better than raw delay plus temperature. Second, a volunteer's exposure to such a complaint raises their hazard of quitting after accounting for how active they already are. Third, notification timing that responds to $E_r$ reduces the predicted number of futile trips.

The commodity classes are leafy greens, climacteric fruit, other produce, prepared food and shelf-stable goods. $Q_c$ is fit hierarchically, with an ordering constraint that leafy greens and climacteric fruit are faster than shelf-stable goods. Shelf-stable goods have $Q_c$ fixed near 1 and serve as a built-in negative control.

Research Program

The first stage is measurement. The logistics dataset and the 14,439-comment feedback corpus are joined by rescue. A key risk is that the logistics file, as described, lists timestamps and weather but not commodity type. The first deliverable is therefore a commodity label per rescue, taken from donation descriptions if they exist. Otherwise it is extracted by an LLM from the comment and the donor type. Both routes are audited against roughly 300 newly hand-labelled rescues.

From each comment the LLM extracts short normalized terms of three kinds: produce ("strawberries", "lettuce"), condition ("wilted", "moldy", "smelled"), and friction ("waited", "recipient refused", "threw out"). A keyword co-occurrence network is built over these terms.

Standard co-word studies such as Mapping longitudinal scientific progress … of the Alzheimer's disease neuroimaging initiative keep edges above an arbitrary frequency cutoff. The file itself flags that cutoff as the method's weak point. Here an edge is kept only if its count exceeds what a degree-preserving random rewiring of the network produces, at a false-discovery-controlled level. Betweenness centrality, which measures how often a term lies on shortest paths between clusters, is then used to find the bridge terms linking the spoilage cluster to the frustration and quitting cluster. Those bridge terms, for example whether "recipient refused" sits between "wilted" and "not worth it", define a validated binary outcome: a futile-rescue complaint.

The second stage is inference under small samples. The biomarker designs of APOC1 as a novel diagnostic biomarker for DN and Identification and validation of IgG N-glycosylation biomarkers of esophageal carcinoma separate discovery from an independent validation cohort. We copy that separation in time. Everything, including the network, the lexicon, the $Q_c$ estimates and variable selection, is fit on 2018–2022 data. The analysis plan is frozen, and the model is then evaluated once on 2023–2024 data. Variable selection uses LASSO-penalized regression with stability selection, meaning resampling to see which variables survive most refits, rather than a single fit.

Retention is modelled as time to a volunteer's last rescue, using a survival model. A volunteer's exposure to futile-complaint rescues enters as a time-varying covariate. Participation follows a power law, with a few super-volunteers doing most trips, so the model conditions on cumulative rescue count. It also compares volunteers who received similar numbers of trips but different heat doses.

The third stage is intervention. The study uses the counterfactual claim-time estimator of Improving Efficiency of Volunteer-Based Food Rescue Operations, which estimates how fast rescues would have been claimed under other notification settings from logged data. With it we simulate a rule that shortens the second-notification delay when predicted $E_r$ is high. The rule holds total notifications fixed so the gain cannot come from spamming, which the literature ties to app fatigue.

Baselines

The naive reference predicts futile-rescue complaints from the base rate and from elapsed time alone. It shows how much signal any structure adds.

The strongest existing competitor combines two published pieces. One is the feature set of the stacked claim-prediction model of Improving Efficiency of Volunteer-Based Food Rescue Operations: distance, timing and raw temperature and precipitation. That paper reports an AUC of 0.81 for claim prediction, which is a different target, so it is a capability reference rather than a direct comparison. The other is RescueLens's Inadequate Food label as the outcome. RescueLens reports 96% recall and 71% precision on issue classification against its 125-entry expert set, with Cohen's κ of 0.73 between annotators. That κ sets a ceiling on how cleanly any text-derived outcome can be measured.

The competitor gets the same joined data, the same temporal split, and the same hyperparameter search budget, with gradient-boosted trees and the stacked architecture both tried. The proposal beats it if the heat-dose model improves held-out AUC for futile-rescue complaints and the 2023–2024 bootstrap confidence interval excludes zero. No reported result was found on Lacuna for predicting food-quality complaints or volunteer churn on this platform, so these numbers are established here.

Two ablations isolate the key ingredients. The first replaces commodity-specific $Q_c$ with one shared coefficient, testing whether produce physiology matters or just heat-times-hours. The second swaps the significance-filtered co-occurrence outcome for the raw RescueLens category, testing whether the network lens changes conclusions. For retention, the reference is a survival model with complaint counts but no heat dose or text. The PROOF and contextual-budget bandits are not re-run as competitors, because they optimize a different objective. They are the natural policies into which a validated $E_r$ would later be inserted.

Why not just…?

Why not just use RescueLens's Inadequate Food label and skip the co-occurrence network? This is the most serious threat, and the study treats it as an ablation, not a strawman. The LLM labels are good at saying whether a comment is about food. They do not say whether the problem was spoilage, too little food, or the wrong food, and those have different causes. The network test is concrete. If the significance-filtered futile-rescue outcome and the raw category yield the same heat-dose effect size and the same churn hazard, the network adds nothing and should be dropped in favour of the LLM label. If the raw category dilutes the effect, the lens has earned its place.

Why not just add temperature and delay to the existing claim model? That model already sees weather, so a generic heat-by-time interaction may capture most of the signal. The shared-$Q_c$ ablation settles this directly. The shelf-stable negative control strengthens it. If heat dose "predicts" complaints about canned goods as strongly as about lettuce, the effect is confounding, such as hot days being busy days, and not spoilage.

Why not just use a topic model or embedding clusters on the comments? Topic models also find themes. However, they give no edge-level significance test and no bridge structure between spoilage and quitting. They also depend on a chosen topic count, which has the same arbitrariness as a frequency cutoff. A topic-model outcome is run as a secondary comparison. If it reproduces the same held-out effect, the conclusion holds and the choice of lens is a convenience.

Why not just model churn from rescue counts? Busy volunteers see more bad pickups and also churn less, so a naive complaint-count effect could even take the wrong sign. Conditioning on cumulative trips, and comparing matched volunteers who differ in heat dose, is what separates "bad trips drive people away" from "active people report more."

Decisive Checks

The idea succeeds or fails on four results in the frozen 2023–2024 validation period.

The first is whether commodity-specific heat dose beats both the shared-coefficient ablation and the strongest-competitor feature set on futile-complaint AUC.

The second is whether it fails, as it should, on two placebos. One placebo is shelf-stable rescues. The other is non-food complaints such as locked doors and bad directions, which heat should not cause.

The third is whether exposure to a futile rescue raises the hazard of quitting after conditioning on activity, with a stable direction across bootstrap resamples.

The fourth is whether the heat-aware notification rule, holding notification volume fixed, lowers the simulated count of high-dose deliveries under counterfactual claim times.

A positive first check with failed placebos would mean the heat dose is picking up busy hot days rather than spoilage. A positive first check with a null third check would mean spoilage matters for food waste but not for retention. That would still be a useful redirection of the retention literature toward other causes.