Two dates. What is published now, what follows, and what could move.
Published in this repository:
- SPEC.md — what is being built, the base model and the reasoning behind it, intended and out-of-scope use, and the constraints that are known before any results exist
- EVAL_PROTOCOL.md — how it will be measured, fixed before any training run
- COVERAGE.md — target crops and conditions
- DATA_STATEMENT.md — what the dataset will contain and how it is licensed
- LICENSING.md — the licence decision and why
- RESPONSIBLE_USE.md — known failure modes, prompt injection, deployment obligations
- Governance, contribution terms, conduct and security policy
Not published: model weights, training or inference code, the dataset, any evaluation result.
- Model weights, in safetensors
- Training, evaluation and inference code
- The schema — the documented interface contract, enough to build your own harness
- The training, validation and test dataset under CC-BY-4.0
- Full results against the protocol above, across all five axes
- Nomination filed with the Digital Public Goods Registry
DPG recognition requires an AI system's training data to be published under an open licence, alongside its code and weights. Two things take time: preparing a dataset that meets that bar, and building a corpus of agronomic reasoning trajectories, which does not exist openly and has to be constructed and reviewed.
Published in advance, so a slip is visible early.
| Risk | Effect |
|---|---|
| Consent for open publication takes longer than expected on some sources | Narrower dataset, or a later release |
| The trajectory corpus needs more agronomist review time than budgeted | Later release. This is the longest lead item. |
| Results miss the failure thresholds in the evaluation protocol | We narrow the published coverage, or delay. The criteria are fixed before the runs. |
| A licence cannot be verified at primary source for a material data source | That source is excluded, reducing coverage |
If the date moves, we will say so here and say why, early.
Not commitments, and deliberately not dated:
- Additional crops and conditions, prioritised by what people building on this actually ask for. These are added to the reference library by curation first — see the retrieval section of SPEC.md — with fine-tuning as the fallback route.
- Reference-library material for local hosts and conditions, and guidance on curating your own
- Fine-tuning recipes and adapters for local crops and languages
Agronomist-2, revisiting the base-model choice once newer architectures are proven on the hardware available to us
If you need a crop, a condition, a language or a deployment target that is missing, open a Discussion. That is how it gets prioritised.