Skip to main content

Sprint 2 delivery

Sprint 2 direction​

The main clarification for Sprint 2 was that RODENT is a research platform, not a project whose primary task is to invent a biologically accurate virtual rat. The demonstration agent remains useful for proving that configuration, timing, recording and replay work, but future research models should connect through stable interfaces.

Sprint 2 therefore focused on the platform around the model. A researcher should be able to create a project, choose or build an arena, configure conditions, collaborate with a team, operate a run, preserve results and later substitute a more appropriate model.

What we delivered​

Research project website​

The website was reorganised around a researcher workflow rather than exposing every Godot control at once. It provides account creation, sign-in, password reset, projects, experiments, project membership and a workspace for setup, running and review. New-project creation includes template selection, and the simulation can be focused without permanently hiding the surrounding controls.

The website and Godot interface use a shared dark visual language with consistent typography, cyan actions, yellow warnings, green success states and bordered panels. Controls are grouped by task so researchers do not need to understand the internal plugin structure.

Hosted identity, projects and access​

Supabase Auth and Postgres support the hosted account and project workflow. The project creator becomes its lead. A lead can add an existing account and assign a project role and edit scopes such as arena, light, odour, sound, treatment, simulation or agent settings. These scopes support multidisciplinary teams without pretending that every researcher requires unrestricted access.

The browser uses the Supabase project URL and publishable key. Database Row Level Security remains the authority for protected data; hiding a control in the interface is not treated as sufficient access control. A local Python and SQLite path is retained for isolated development.

Arena templates and custom design​

The integrated template catalogue contains six configurable starting arenas, including the MCSF-inspired layout. Researchers can choose a template when creating a project and can revisit template choices inside the workspace.

The 2D arena template builder provides a clearer authoring surface for regions, walls and doors. It includes grid-based placement, object properties, materials, colours, dimensions, agent start position, validation checks and JSON preview. A saved design uses the same experiment schema as the supplied templates so it can be opened in the 3D workspace.

Configurable experimental conditions​

Researchers can configure light, odour and sound sources, treatment or placebo conditions, dose and administration timing, seeds, fixed-step timing and relevant demonstration-agent settings. Doors can be opened or closed during setup and operated during a run when the researcher has permission. These changes are recorded as interventions.

The platform records the configuration and events needed to explain a run. A visual effect is not treated as scientific evidence of how a real animal responds. The eventual behavioural response comes from a selected model.

Internal headless control API​

The local Python client can launch Godot without rendering and use reset, step, observe, act, load and close commands. This demonstrates that simulated time and recording do not depend on a visible frame or mouse interaction.

The existing act command is intentionally limited. It proves the control boundary but is not yet the final observation, action and reward contract for machine-learning training. The next stage will expose readable local sensations and a versioned model adapter, described in the Research and development roadmap.

External environment API​

The Open-Meteo integration lets a researcher select a location, date and time, review retrieved environmental conditions and apply accepted suggestions to an experiment. It can suggest light and ambient sound settings and preserve temperature data. Odour is not invented from weather data.

Fetched values are normalised into RODENT's own data structure and presented for review before application. Accepted conditions are saved with the experiment so replay does not depend on a later network request. The generated mappings are setup heuristics, not calibrated claims about laboratory exposure.

Recording, replay and results​

Runs record step-based state, position, region, active stimuli, treatment events and researcher interventions. Researchers can download JSON for the complete structured record and CSV for tabular analysis. Replays can be saved and reviewed separately from a live run.

This separation is important for the next stage. Large groups of episodes can eventually run without rendering, after which selected episodes can be reconstructed in the 3D view from their recorded state and events.

Testing and delivery​

The automated workflow checks website JavaScript syntax, experiment presets, deterministic configuration, Python backend and headless-client behaviour, researcher-facing UI contracts and Godot headless behaviour. Python application coverage has a 60% minimum gate. JavaScript and GDScript are tested separately and are not included in that Python percentage.

The application is hosted on Vercel, the documentation is built with Docusaurus and published through Cloudflare Pages, and source control follows short feature branches and pull requests into main. A Gitea Actions run still depends on an available shared runner.

The redesign was informed by feedback from Adilah Abed, Devon Jarvis, Muhammed Akhalwaya and Irfaan Fulat. The common result was that the platform became easier to understand once its purpose, navigation and feature grouping were clearer. The full summary and resulting actions are recorded in Sprint 2 User Feedback.

Researcher workflow delivered in Sprint 2​

Create or join a project
          |
          v
Choose a supplied template or build an arena
          |
          v
Configure conditions and team permissions
          |
          v
Run visually or control locally without rendering
          |
          v
Record interventions and results
          |
          v
Replay selected runs or export data

What Sprint 2 does not claim​

  • The demonstration agent is not a validated model of real rodent behaviour.
  • Weather-derived settings are not laboratory calibrations.
  • A model drop zone does not make an arbitrary file safe or compatible.
  • A reproducible seed proves software repeatability, not biological validity.
  • A visual marker for risk or emotion is meaningful only if a documented model supplies and validates that variable.
  • Passing automated tests does not replace user testing, scientific validation or role-by-role security checks.

Handover to the next stage​

The next development stage should freeze a researcher-facing model contract. It must define observations, sensations, actions, interaction rules, termination, result records, model metadata and compatibility errors. Once that contract is stable, RODENT can support high-volume headless runs, including batches of 1,000 or more episodes, and render selected completed runs afterward.

This approach keeps training fast and makes the visualisation an auditable replay of recorded data rather than part of the model's decision loop.

AI Attribution: This page was drafted with OpenAI Codex assistance from the integrated Sprint 2 application and documentation. The team should verify feature status and attach its own sprint evidence.