Skip to main content

Sprint 3 and final integration

Direction​

After Sprint 2 established the researcher portal, the remaining work moved from placeholder integration points to a usable model workflow. The final application keeps the built-in rat for demonstrations, but researchers can now train or supply their own controller and use it without rendering every training episode.

Delivered work​

Researcher model workflow​

  • rodent-sdk runs the same fixed-step simulation from Python and Google Colab.
  • A Gymnasium-style environment exposes named observations, continuous or discrete actions, researcher-defined reward and termination, episode logging and repeatable episode reconstruction.
  • The supplied notebook demonstrates training, before/after comparison, saving and reuse in another experiment.
  • Project models have immutable versions, checksums, training metadata and project-scoped access.
  • Compatible networks are exported to ONNX and can drive the rat live in the website.

Stimulus propagation and observation​

  • Light uses direct paths and reflections.
  • Sound and odour weaken through walls and doors.
  • Odour diffuses over simulated time.
  • The SDK and model driver receive local stimulus samples rather than only source settings.
  • Heat maps help a researcher inspect what the model can sense before training.

Research review​

  • Training logs plot learning curves and compare selected episodes.
  • Any recorded episode can be rebuilt with its seed and actions.
  • Selected runs can be saved into the website's Review area.
  • Replays identify whether the built-in rat or a particular project model version produced the run.
  • Rat-eye, follow, orbit and free-fly cameras provide different inspection views.

Reliability and regression checks​

  • CI builds and installs bundled and slim SDK wheels outside the repository.
  • The pack download is checked with SHA-256.
  • Training tests verify the Gymnasium contract, multiple simulator copies, episode reconstruction and that a small example policy genuinely improves on its declared task.
  • Godot tests verify model-driven movement, browser/Python agreement, stimulus sampling and propagation.
  • Fingerprints guard the built-in rat's six preset trajectories against accidental change.

Contributions visible in the final merge​

The final main history records Huzaifa's model-training, notebook, SDK, episode-log, project-model and analysis work, and Ahmed's stimulus-propagation work plus the live website ONNX model runner and rat-eye camera. The merge history should be used as contribution evidence rather than assigning work from memory.

Remaining limits​

  • The built-in rat remains a demonstration baseline.
  • The supplied PPO example is a worked software example, not a validated biological model.
  • Recurrent ONNX models and custom additional observations do not run live in the browser yet.
  • Very long recordings can exceed hosted replay limits.
  • Database migrations and storage policies must be applied and tested in the live Supabase project; committing SQL is not deployment evidence.
  • Final accessibility, performance and user tasks still need evidence from the deployed release.

Methodology reflection​

The project began mostly bottom-up: deterministic stepping, plugins, stimuli and recording were built before the complete researcher journey was known. Client clarification caused a deliberate shift to top-down workflow design around projects, Setup, Run, Review and external models. That change was useful, but it also created documentation drift and several integration rounds.

The partial microkernel remained appropriate. The kernel owns deterministic simulation time, randomness and ordered plugins, while authentication, project management, training notebooks and database access stay outside it. Treating every feature as a kernel plugin would have coupled unrelated product concerns to the simulation loop.

AI Attribution: This delivery record was prepared with OpenAI Codex assistance from the final merge history and current application source. The team should attach links to its actual pull requests, Actions runs and deployment evidence.