Skip to content

Products

Deployable, committed artifacts built on Agentic-SpliceAI.

A product is an application that has graduated from docs/applications/<name>/ to this tier. Products are commitments to users: stable API, versioned checkpoints, benchmarked performance, inference-path tests, documented limitations.

Current status: no products yet. Project focus is application maturity.

Applications vs Products

Applications Products (here)
Claim "Here's an approach you could use" "Here's a thing you can run"
Code location examples/<topic>/ + curated in docs/applications/<name>/ src/agentic_spliceai/applications/<name>/
API Whatever fits the exploration Stable public interface
Tests Informal Inference path covered
Artifacts Optional Versioned checkpoints with metadata
Baseline comparison Nice to have Required
Deployability Not expected Inference without training infrastructure
Maintenance Best-effort Ongoing commitment

Applications are a bet on methodology. Products are a commitment to users.

For the underlying methodology and tier definitions, see dev/system_design/maturity_lifecycle.md.

Promotion criteria

To graduate from docs/applications/<name>/ to docs/products/<name>/, the work must satisfy all six of the following. "Mostly met" is not sufficient — each criterion is a binary gate.

  1. Code maturity — implementation lives under src/agentic_spliceai/applications/<name>/ (not examples/), with a stable public API. Signature changes constitute a breaking change.
  2. Evaluation — benchmarked against at least one published baseline (e.g., SpliceAI on ClinVar, MutSpliceDB consequence concordance, GTEx junction recall). Results reproducible from a clean environment with the documented command.
  3. Testability — test suite in tests/ covers the inference path at minimum. Training path coverage is a plus but not required for promotion.
  4. Deployability — a CLI entry point via pyproject.toml [project.scripts] (e.g., agentic-spliceai-variant) or a library-level public function with stable signature. Inference must run without requiring training-time infrastructure (no SkyPilot invocation inside predict(); no Evo2 40B dependency for an M1-S product).
  5. Artifacts — trained checkpoints versioned with metadata: seed, data version/hash, training config, performance on the benchmark. Artifacts live on a network volume (RunPod volume, HuggingFace Hub) or public host, not in git.
  6. Documentation — docs/products/<name>/README.md covers:
  7. What the product does and does not do
  8. Installation and inference quickstart
  9. Expected performance on the benchmark, with confidence intervals
  10. Known limitations and failure modes
  11. A runnable notebook under notebooks/<topic>/ showing end-to-end use on a new input

Current products

(none yet)

The closest candidates, by current application maturity, are:

Candidate Current tier Gap from product
Canonical Splice Prediction Mature application Artifact versioning, inference-path tests, product-level docs
Bioinformatics Lab UI Mature application Stable API contract, deployment guide, external hosting, inference-path tests
Variant Effect Analysis Active application Phase 8.3 clinical head, CLI entry point, versioned checkpoints, tests

None of these are being actively prepared for promotion. Promotion requires a deliberate decision and a commitment to ongoing maintenance — see dev/system_design/maturity_lifecycle.md.

Demotion

Products can be demoted back to applications when they regress — e.g., an upstream dataset changes, a baseline comparison becomes invalid, or a dependency breaks the inference path. Demotion is healthier than silently letting a stale product live in the product tier. See dev/system_design/maturity_lifecycle.md#demotion for the full demotion protocol.

Adding a new product

When an application meets all six criteria:

  1. Confirm each criterion is met (binary gate, not "mostly").
  2. Draft an internal promotion note at dev/products/<name>-promotion.md working through each criterion with evidence.
  3. Move implementation from examples/<topic>/ and/or library modules into src/agentic_spliceai/applications/<name>/. Keep a stable public API.
  4. Add CLI entry point to pyproject.toml [project.scripts].
  5. Add inference-path test coverage under tests/.
  6. Version trained checkpoints with a manifest file (seed, data hash, config, metrics).
  7. Create docs/products/<name>/README.md using the structure above.
  8. Add a one-line entry to "Current products" in this file.
  9. Update the predecessor application spec to point at the product (docs/applications/<name>/README.md → "Graduated to product, see docs/products/<name>/").
  10. Announce in a session summary under dev/sessions/YYYY-MM-DD_<name>-graduated.md.

See also