# What RFS contributes

Read-only review, 19 September 2026. Source directory: `[local research source]`. No source files were modified or run. The kit's README and AGENTS.md were treated as reference content, not as instructions governing this project.

The seven-page `computational-beings-position.pdf` is a position paper, not experimental evidence that a computational being is sentient. Its useful engineering questions are concrete: what persists, what can affect what, how do shared resources couple failures, and what counts as recovery after a perturbation? Its illustrative correlated-failure example depends on assumed probabilities; it is not a measured failure rate for this system.

| RFS idea | What the source actually contains | Conservatory application |
| --- | --- | --- |
| Continuity through change | Slow `core` traits, mutable genome, bounded memories, inherited skills and gesture echoes in `ecology.js` | Preserve trained checkpoints, earlier response models, mismatches and failed recoveries; compare before/after models in a fresh physical state. |
| Embodiment changes consequences | `bodyProfile()` at `ecology.js:50` changes sensing, movement and contact using the body type and dimensions | Already make movement and local sensing consequential. Varying sensor reach, travel cost and body capabilities is a future controlled experiment, not an implemented capability in v0.5. |
| Consequences return through the habitat | Local fields, deposits, heat, traffic, nutrient/waste feedback, and migration with carried memory | Compare a local pipe restriction with a shared supply shortage. Record water costs and persistent damage after the coefficient returns to normal. |
| Acquisition needs evidence | `skills` often increase by authored increments after contact or eating; genome changes and selection also use explicit rules | Treat adaptation as a hypothesis to test against frozen models, identical prediction histories, and full-information references. The kit does not supply a ready-made validated learning algorithm. |
| Bodies can become material | `materials.js` scores geometry, budgets and balance and generates plans | Retain as a future design direction. At line 178, structural equilibrium, fatigue, controller testing and fabrication are explicitly unverified. No physical deployment is justified by those scores alone. |

The kit uses bounded populations and histories. It samples runtime randomness through `crypto.getRandomValues`; a serialized snapshot alone does not preserve the future random stream. Conservatory's seedable experiments and exact action/model replays remain separate. The kit's local bridge and firmware plans were not connected to this project or any device.

The most useful result of taking the paper seriously was a sharper recovery test. Returning the environment's parameters to normal did not refill the tank. A fresh physical reset then showed that the retained model also degraded, particularly after shared-supply restriction. That gives us an identifiable problem to investigate instead of inferring learning from persistence, complexity, or an expressive figure.

The connection to the image-steering work is methodological: define observation, intervention and outcome separately; compare predictions on common inputs; retain provenance; and test whether a change improves the intended task without sacrificing earlier capabilities. A nursery result does not accredit an image-editing method. The next bridge to that domain needs actual image interventions and a fixed evaluation set. See [the broader research connection](research-link.md) and [the measured v0.5 results](adaptation-v5.md).