# Research-preview staging checklist

## Complete for research-preview repository hygiene

- [x] Stage 1 and Stage 2 FP16 TensorFlow Lite weights packaged.
- [x] Machine-readable model metadata included.
- [x] Minimal end-to-end inference entry point included.
- [x] Model-artifact SHA-256 checksums recorded and locally verified.
- [x] Intended use, limitations, safety language, and development-split metrics documented.
- [x] Training and evaluation images excluded from the package; source trail recorded separately.
- [x] Stage 1 → symptom crop → Stage 2 execution exercised on a limited local field sample.
- [x] The limited field-sample result is labelled as an implementation smoke test, not accuracy or field-validation evidence.
- [x] Split audit recorded: 80,313 records, 68,786 source + group IDs, and 2,205 groups crossing train/validation/test.
- [x] Historical 0.992/0.993 results relabelled as legacy development metrics, not release or field results.
- [x] Reproducible split-audit tool and machine-readable audit report included in `eval/`.
- [x] Deterministic group-disjoint assignment plan generated: 57,914 train, 11,107 validation, and 11,292 test records, with zero cross-split groups and records.
- [x] Corrected split plan explicitly records that model retraining, corrected-split evaluation, and an independent Ghana field holdout are incomplete.
- [x] Fail-closed package validator added and run locally with the split-tool unit tests.
- [x] Editable and rendered artwork uses research-preview language.

## Required before approving or promoting a model release

- [ ] Resolve OSF northern-leaf-blight provenance and reuse license.
- [ ] Confirm MIT Indoor Scenes terms permit the intended model-weight release, or retrain without it.
- [ ] Review all source terms for compatibility with derived-weight distribution.
- [ ] Decide and publish an explicit model license after compatibility review.
- [ ] Materialize and independently verify the corrected group-disjoint manifest against the source metadata.
- [ ] Retrain both model stages using the corrected training assignment.
- [ ] Re-evaluate both retrained stages on the corrected validation and test assignments, then replace the legacy development metrics in the model card.
- [ ] Complete a documented, agronomist-reviewed field evaluation across representative cameras, farms, regions, seasons, cultivars, and crop stages.
- [ ] Publish the field protocol, tested conditions, aggregate results, sample hashes, and responsible sign-off with the release record.
- [ ] Have an agronomist independently label field samples and verify them through the complete Stage 1 → Stage 2 path.
- [ ] Benchmark latency and peak memory on every claimed target device.
- [ ] Obtain named agronomy, privacy/legal, security, product, and release-owner approval.
- [ ] Obtain explicit public-release and production-use approval only after the preceding gates pass; then update the repository and artwork status language.

## Current visibility and recommended posture

The existing Hugging Face repository is **publicly visible**. This update does
not change its visibility. Public access is not public-release approval,
production approval, field validation, or a license grant. The current evidence
supports this statement only: **Research-preview maize screening model evaluated
on documented development datasets and limited field samples.**

Keep the research-preview and no-license warnings prominent and do not promote
the package as an approved release. Whether to restrict public access while the
license review remains open is a separate legal and release-owner decision. The
package is not ready for a performance, field-validation, deployment, or
license-clearance announcement.

Audit artifacts: [`development-data-audit.json`](eval/development-data-audit.json)
records why the historical metrics cannot support public claims;
[`corrected-split-plan.json`](eval/corrected-split-plan.json) records the repaired
assignment and the still-false retraining, corrected-evaluation, and independent
field-holdout gates. A leakage-free assignment plan is not a new trained model or
an evaluation result.