# OSUN Research Corpus (Cited + Design-Mapped)

## Selection criteria

The corpus is constrained to:

1. High-credibility venues (ICLR, Nature portfolio, International Journal of Forecasting, Transportation Research Part E, Journal of Cleaner Production, FDA primary guidance).
2. Direct relevance to OSUN product tasks (agricultural yield, fermentation sustainability, precision nutrition, demand/sourcing forecasting, restaurant automation, ghost kitchen dispatch, recipe compliance).
3. Methods that can be translated into reproducible benchmarks (time-series metrics, multi-horizon modeling, cost/latency tradeoffs).

## Core cited corpus

| ID | Source | Venue / year | Why it is in the corpus | Implementation mapping |
| --- | --- | --- | --- | --- |
| C1 | Makridakis et al., *The M4 Competition: 100,000 time series and 61 forecasting methods*. DOI: [10.1016/j.ijforecast.2019.04.014](https://doi.org/10.1016/j.ijforecast.2019.04.014) | International Journal of Forecasting, 2020 | Establishes large-scale forecasting benchmark framing and error metric discipline. | Benchmark protocol includes explicit baseline comparisons and sMAPE/MAE reporting. |
| C2 | Makridakis et al., *The M5 competition: Background, organization, and implementation*. DOI: [10.1016/j.ijforecast.2021.07.007](https://doi.org/10.1016/j.ijforecast.2021.07.007) | International Journal of Forecasting, 2022 | Hierarchical retail forecasting setup with exogenous drivers and intermittency. | SourceGrid + WasteZero tasks include intermittent behavior and exogenous disturbances. |
| C3 | Lim et al., *Temporal Fusion Transformers for interpretable multi-horizon time series forecasting*. DOI: [10.1016/j.ijforecast.2021.03.012](https://doi.org/10.1016/j.ijforecast.2021.03.012) | International Journal of Forecasting, 2021 | Multi-horizon forecasting under mixed covariates with interpretability constraints. | Multi-horizon tasks and model comparisons use explicit horizon-level scoring and readable result cards. |
| C4 | Nie et al., *A Time Series is Worth 64 Words: Long-term Forecasting with Transformers* ([OpenReview](https://openreview.net/forum?id=Jbdc0vTOcol)) | ICLR 2023 | Efficient long-horizon design with patching/channel-independence ideas. | Informs low-latency forecasting design and long-window test tasks. |
| C5 | Liu et al., *iTransformer: Inverted Transformers Are Effective for Time Series Forecasting* ([ICLR proceedings](https://proceedings.iclr.cc/paper_files/paper/2024/hash/2ea18fdc667e0ef2ad82b2b4d65147ad-Abstract-Conference.html)) | ICLR 2024 | Strong long-lookback performance with practical computation tradeoffs. | Cost-latency benchmark reports include median inference latency and cost proxies per 1k forecasts. |
| C6 | Tamayo-Vera et al., *Advanced machine learning for regional potato yield prediction*. DOI: [10.1038/s44264-025-00052-6](https://doi.org/10.1038/s44264-025-00052-6) | npj Sustainable Agriculture, 2025 | Regional crop-yield prediction under climate/soil/NDVI context. | FarmOS task generator includes seasonal + exogenous + regime behavior and model robustness scoring. |
| C7 | David et al., *The role of techno-economic and life cycle assessment in guiding precision fermentation towards sustainable food production*. DOI: [10.1016/j.tifs.2025.105488](https://doi.org/10.1016/j.tifs.2025.105488) | Trends in Food Science & Technology, 2026 | TEA/LCA framing for fermentation decisions and sustainability-aware tradeoffs. | FermentOS benchmark couples quality metrics with compute-cost proxies to support practical decision tradeoffs. |
| C8 | Wu et al., *A Scoping Review of Artificial Intelligence for Precision Nutrition*. DOI: [10.1016/j.advnut.2025.100398](https://doi.org/10.1016/j.advnut.2025.100398) | Advances in Nutrition, 2025 | Consolidates AI precision-nutrition practices and evaluation gaps. | NutriFlow scenario design emphasizes measurable, reproducible, feature-driven outputs. |
| C9 | Hess et al., *Real-time demand forecasting for an urban delivery platform*. DOI: [10.1016/j.tre.2020.102147](https://doi.org/10.1016/j.tre.2020.102147) | Transportation Research Part E, 2021 | Near-term operational demand forecasting under intermittent/double-seasonal signals. | WasteZero demand tests use short-horizon operational scoring with intermittent demand profiles. |
| C10 | Wu, *A Grover-Based Quantum Algorithm for Solving Perfect Mazes via Fitness-Guided Search*. DOI: [10.48550/arXiv.2507.21937](https://doi.org/10.48550/arXiv.2507.21937) | arXiv preprint, 2025 | Fitness-guided search strategy and bounded candidate refinement concept. | Inspires Lux SparseBeam bounded candidate search objective (error + compute penalty). |
| C11 | De Mello et al., *Machine learning methods for food demand forecasting in catering operations*. DOI: [10.1016/j.jclepro.2024.144726](https://doi.org/10.1016/j.jclepro.2024.144726) | Journal of Cleaner Production, 2024 | Waste-aware demand forecasting grounded in real catering operations. | WasteZero + GhostFlow benchmarks include short-cycle demand shocks and waste-sensitive signals. |
| C12 | Pabón et al., *AI-Driven Last Mile Logistics in the Food Delivery Sector* ([arXiv](https://arxiv.org/abs/2408.07417)) | arXiv preprint, 2024 | Route-level and dispatch-level optimization structure for food delivery contexts. | GhostFlow dispatch pressure benchmark and latency/cost-aware routing signals. |
| C13 | Zhang et al., *A robust dynamic scheduling method for mixed-model two-sided robotic assembly lines under uncertainty*. DOI: [10.1016/j.rcim.2024.102734](https://doi.org/10.1016/j.rcim.2024.102734) | Robotics and Computer-Integrated Manufacturing, 2024 | Dynamic scheduling under uncertainty and constrained robotic resources. | AutoKitchen benchmark includes uncertain demand with resource-constrained throughput optimization. |
| C14 | U.S. Food and Drug Administration, *Food Additive Status List* ([FDA source](https://www.fda.gov/food/food-additives-petitions/food-additive-status-list)) | FDA primary source | Regulatory status registry for food additive compliance checks. | RecipeProof compliance fields are explicitly constrained to approved-status planning workflows. |
| C15 | Ahn et al., *Flavor network and the principles of food pairing*. DOI: [10.1038/srep00196](https://doi.org/10.1038/srep00196) | Scientific Reports, 2011 | Quantitative flavor-network basis for ingredient pairing and recipe exploration. | RecipeProof scenario design includes gastronomy-driven pairing constraints with compliance overlays. |
| C16 | U.S. Department of Agriculture, *Local Food Directories: National Farmers Market Directory* ([USDA source](https://www.ams.usda.gov/local-food-directories/farmersmarkets)) | USDA primary source | Canonical location-indexed local supplier discovery inputs. | LocalSource module uses radius-aware supplier matching and procurement routing. |
| C17 | U.S. Food and Drug Administration, *Inventory of Effective Food Contact Substance (FCS) Notifications* ([FDA source](https://www.fda.gov/food/packaging-food-contact-substances-fcs/inventory-effective-food-contact-substance-fcs-notifications)) | FDA primary source | Primary source for food-contact packaging chemistry approvals. | Packaging rows in LocalSource and RecipeProof compliance notes stay linked to approved-contact workflows. |

## Evidence policy for product claims

- Product claims are generated only from measured benchmark artifacts in:
  - `web/data/benchmark_summary.json`
  - `reports/osun_outperform_report.md`
- Required validated data inputs and coverage contracts are tracked in:
  - `benchmarks/validated_data_registry.json`
  - `docs/research/benchmark_data_requirements.md`
  - `reports/benchmark_data_validation.json`
- Claims are limited to observed metric deltas (no unverifiable frontier claims).
- C10 and C12 are explicitly treated as non-peer-reviewed inspiration, not as standalone production evidence.