# Hive simulation audit

Status: implemented and verified within the experimental model described below. See [detailed evidence](neural-validation.md) for sources, version-specific results and limitations.

| Requirement | Current evidence | Status |
| --- | --- | --- |
| Physical foraging and return through the entrance | Actual neural policy, 3,600 consecutive seconds, eight foragers: 269 exact deposits at the internal home point; no movement above the modeled speed limit; entrance and main-trunk checks pass. | Verified for the tested scenario; simplified collision volumes. |
| Queen loss has biological consequences | Workers continue collecting; no new worker eggs without a fertile queen. Young existing brood can become one queen candidate, with emergence and weather-dependent mating delays. No instantaneous replacement. | Rules and controlled scenarios verified. |
| Food, pollen and workers are conserved | Separate ledgers, load-limited harvesting/deposit, paid construction, finite pollen and brood care. Scarcity produces extinction; recovery produces survival. Swarms transfer individuals and provisions without duplicating them. | Regression and integration checks pass. |
| Colony develops through an actual lifecycle | A fresh favorable 28-day run produced 96 births and nine deaths: 32 founders became 119 surviving workers, with 309 developing brood. Care commitments and delayed task adaptation maintained nursing. | Full physical lifecycle verified for this seed and favorable scenario; growth is not guaranteed. |
| Favorable and unfavorable conditions matter | Synthetic day/night, temperature, rain and flowering constrain flights and renewal; food shortage and insufficient care/heat can kill brood or adults. RNG and states persist. | Controlled scenarios and deterministic restores pass; rates are experimental. |
| Collective behavior has an explicit mechanism | Food actually delivered can trigger a local source report. Nearby bees may learn a source while retaining separate neural memories. | Contact, locality, identity and no-free-report tests pass. No shared consciousness claim. |
| Neural explanations describe the actual controller | Frozen RecurrentPPO weights, private states, feasible action selection and explicit weather/care overrides. Dormant or dead bees show stored memory rather than active choices. | Policy comparison, inspection regression and actual browser checks pass. |
| Accelerated viewing preserves real movements | Bounded one-second recordings played at x3; explicit cuts between clips. Saved clips survive server restart exactly. Sparse older history is identified separately. | API and in-app playback verified. |
| Existing worlds survive updates | Checkpoints retain IDs, stocks, RNG, memories and model version. An 18-day checkpoint resumed deterministically under the updated engine. | Verified; no silent weight substitution. |
| Large populations remain represented | 10,000 and 40,000 worker engine benchmarks retain every live ID. Actual in-app view: 8,176 bees at approximately 120 fps. Adaptive graphics reduce detail, not population. | Measured; 40,000-bee interactive frame rate is hardware-dependent. |

This is an experimental stylized simulation, not a reproduced biological brain or a scientifically validated colony forecast. Distances, food-equivalent quantities and several rates remain uncalibrated. Drone populations, laying workers, detailed disease, neural nursing and collisions against every branch/comb cell are outside the current model. The cocoon is an artistic shelter.

The completed colony is available in the app as **Growth example · 28 days**. Its movement clip contains two further real simulated minutes after day 28, played at x3. It is a separate completed experiment; existing colonies retain their saved identities and neural model versions.

Use **Watch experiment** for those recorded movements and **Inspect latest state** for its current neural state. Returning to playback closes the current-brain panel. **Recorded history** keeps the selected experiment rather than switching to the live world. These transitions and the correct 32-unit policy were verified in the in-app browser. [Daily lifecycle results and source hashes](growth-validation.json) are included.