RELEASE_CHECKLIST.md

Research-preview staging checklist

Complete for research-preview repository hygiene

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