# The Hive: autonomous simulation validation

Experimental simulation, without scientific validation or replication of a biological brain. These results come from local tests, not observations of a real colony.

## Trained network

RecurrentPPO uses 32 LSTM units followed by a 32-unit action layer. Foragers share frozen weights but retain individual recurrent states and flower memories. Hive workers and queens follow explicit rules.

Training uses 1,800 imitation updates followed by 65,536 reinforcement learning transitions. Rewards favor nectar actually delivered and penalize time and bee loss. Perceptual flower selection and navigation remain rule-based; the network selects six actions rather than wing movements.

Validation covers 20 seeds unseen during training, with identical conditions for three strategies. Episodes last up to 256 five-second decisions. Initial states may place a bee on a flower or give it a load; these figures are not biological daily yields.

| Strategy | Mean nectar delivered per episode |
|---|---:|
| Trained network | 0.7144 g |
| Simple strategy | 0.7540 g |
| Random actions | 0.0280 g |

Model: `ec5a8f9c215fafc6`. Publication threshold: more than 1.5 times the random baseline and at least 60% of the simple strategy. Result: passed. This does not demonstrate generalization to all environments.

## Full-world scenarios

Each scenario starts with 1,000 workers and covers six simulated hours. Favorable conditions start with 8 kg of food and full flowers. Shortage and recovery start with 0.5 g, empty flowers and rain. Recovery starts after one hour. These are not causal comparisons with identical initial conditions across prosperity and shortage.

| Scenario | Surviving workers | Nectar delivered | Resource balance error |
|---|---:|---:|---:|
| favorable | 1000 | 832.207 g | 9.71e-10 g |
| scarcity | 0 | 0.000 g | 0.00e+00 g |
| recovery | 1000 | 177.449 g | 2.18e-11 g |

Eleven automated tests cover limited-nectar competition, resource accounting, identical save/resume including randomness and memory, swarming without duplication, independent starvation between hives, flower recovery, newborn IDs, and rejection of checkpoints linked to another model. Additional checks cover small-colony laying and construction costs, the emergence delay, no laying without food or a queen, and private neural states.

Swarming uses experimental population and food rules with queen-cell maturation. It transfers 40% of workers, their loads and 40% of stored food to a new hive. Travel is calculated physically. Nest selection is not yet a neural scout consensus.

## Local performance

Measured over 100 simulated seconds after warm-up, without reducing population:

| Workers | Foragers | Mean time per simulated second | JSON snapshot construction |
|---|---:|---:|---:|
| 10000 | 2417 | 3.28 ms | 19.69 ms |
| 40000 | 9944 | 11.00 ms | 71.46 ms |

The headless browser measured 109.2 ms per frame with 10,000 workers, 120.6 ms per frame with 40,000 workers. This does not guarantee 60 fps on your device. Rendering adjusts resolution and limits nearby 3D detail; every individual remains represented.

## Usage and limitations

- Open http://127.0.0.1:8767/ruche-vivante.html while the local server runs. One real second equals one simulated second; neural decisions occur every five seconds.
- “Inspect a brain” selects a forager. Each cell shows an actual activation. Click to inspect its hidden state and memory. Probabilities describe the latest decision, not a causal explanation of each unit.
- Live simulation is independent of replay and accelerated experiments. History retains snapshots every 30 seconds for 24 hours. Brain activations are not recorded in replays; inspect them live.
- Saves include individuals, memories, random state and model hash. After downtime, the engine catches up elapsed seconds and reports this in the interface. No permanent remote server is deployed.
- Distances, flow rates and life rates are experimental parameters. Food is measured in grams of food equivalent, without biochemical separation of nectar, honey and pollen.
- The archived BEEHAVE version remains in French on port 8766 and does not feed the new world. No blockchain volume is connected.

## Understanding decisions and starting small

The bee inspector follows four steps: sensed inputs, private memory, chosen action, and physical constraints on its execution. Six probes change energy, load or rain one input at a time, holding the exact pre-decision recurrent state fixed. They recalculate the network without changing the world. A policy probability is not a success rate, and these comparisons are not predictions of colony survival.

The network diagram shows the three strongest incoming candidate-gate recurrent weights per unit, not every connection in the full model. Orange and blue indicate sign. Gate ordering follows [PyTorch LSTM documentation](https://docs.pytorch.org/docs/2.14/generated/torch.nn.LSTM.html). Each worker has private hidden and cell states; the frozen weights are shared. Young workers have dormant foraging states until the model's foraging age of 18 days. Nursing and queen reproduction remain rule-based, not learned neural behavior.

“Start a tiny colony” creates a separate 28-day accelerated experiment with one queen, 32 mixed-age workers and 2 g of stored food. Flower nectar is separate, finite environmental food. Eggs need 21 simulated days to emerge. Small-colony reproduction requires a living queen, at least eight workers, food above a reserve threshold and available brood-care capacity. Laying and construction consume food. Outcomes are not guaranteed and these rules have not been biologically calibrated.

Experiments can be watched, paused and resumed. Pausing saves the exact state. Hourly simulated checkpoints support recovery after server interruption; restored unfinished experiments remain paused. Founding experiments retain hourly replay frames for their whole run. Neural inspection is available in the current experimental state, but not in historical replay frames.

Tests with the actual network reproduced its recorded decision from the saved prior memory and confirmed that six probes leave memory unchanged. Browser checks covered explanations, real connection data, 32-worker creation, isolation of the 10,000-worker live world, and mobile layout. A separate restart check confirmed exact paused-world restoration and subsequent progress.

A separate favorable-condition run (seed 823) completed 28.0 days: 32 initial workers, 116 survivors, 96 births and 12 losses. Resource balance error: -8.502e-09 g. This is one scenario, not evidence of reliable growth under natural randomness.


## Physical action constraints, 14 September 2026

The world now executes the highest-scoring physically possible action from the frozen network's unchanged probabilities. Unloading and resting require the hive entrance; collection requires a flower within reach, free carrying capacity and rain below 85%. Returning is unavailable at home; visiting is unavailable once within collection range. Exploration remains available. Migration still overrides foraging explicitly. This is an execution rule, not additional learning or a biological reflex model.

The earlier strategy comparison above predates this constraint layer and does not evaluate the combined controller. Three integration runs with the actual frozen network, 32 founders and favorable conditions lasted 1,200 simulated seconds each. Seeds 481516, 823 and 1364264692 delivered 3.84, 3.76 and 5.36 grams respectively. The constraints changed 102/1920, 99/1920 and 135/2400 choices. All 32 workers survived each short run; resource conservation error stayed below 0.0000001 gram. These are short regression checks, not a new generalization benchmark.

The inspector distinguishes raw network preference, executed action and measured movement/collection/delivery. Wings follow displayed movement. Accelerated overviews can still skip trips between snapshots; they are not a continuous recording. Existing colonies retain their state, with the constraint activation time recorded as `controller_since`.


## Queen succession audit, 14 September 2026

Implemented biological sequence: queen loss stops new worker egg-laying, not all adult activity. Existing worker brood can still emerge while workers provide care. One existing egg or young larva may be redirected to queen development, never duplicated as a worker. No suitable young brood means no spontaneous replacement. Queen emergence, virgin maturation, weather-dependent mating opportunity, mating loss and a delay before laying are distinct states. Failed succession leaves an aging colony; no automatic rescue is applied. No brood emerges after all adults have died.

During swarming the old queen leaves with the selected workers. The parent colony retains an actual developing queen cell and cannot instantly acquire a fertile queen. Provisions are debited from the parent and carried within individual load capacity, then deposited at the destination, including by migrating nurses. New colonies receive no invented food. Egg production now has an explicit resource cost in large colonies as well as small founding colonies.

Reference sequence: [University of Georgia, queenlessness and laying workers](https://bees.caes.uga.edu/beekeeping-resources/honey-bee-disorders/honey-bee-disorders-non-infectious-diseases-and-pests.html) describes emergency queen rearing from young larvae and eventual collapse when replacement fails. [Penn State, queen development](https://pollinators.psu.edu/assets/uploads/documents/An-Introduction-to-Queen-Honey-Bee-Development.pdf) describes the virgin maturation interval before mating. The simulation uses 16 days from egg to queen emergence, 6 days before mating opportunities and 2 days after a successful flight before laying. These fixed timing approximations replace biological variation.

Explicit experimental assumptions: one candidate queen at a time, external drones assumed available, 10% loss per mating attempt, rain cutoff 35%, 21-day virgin mating window, minimum 8 workers and 0.25 g stores to initiate rearing, simplified queen nursing cost. These probabilities and thresholds are not scientifically calibrated. Drone production by laying workers, pollen/protein nutrition, thermal regulation, seasonality and detailed diseases are still absent. The old fixed queen age limit also remains an approximation; this is not a scientifically faithful whole-colony model.

Regression coverage: 24 tests passed, including delayed succession, mating failure, unsuitable brood, queenless foraging without laying, no resurrection from an empty colony, swarm queen/resource conservation, saved-state reproducibility, physical actions and inference concurrency. Some succession tests jump to developmental boundaries; they are not multi-week end-to-end biological validation.

Existing saves retain their workers, resources and memories. `biology_since` records when the new rules begin. Legacy swarm timers are canceled because those timers did not reserve a real queen candidate. Historical results above predate these rules.


## Pollen and available nursing care, 14 September 2026

The earlier absence of pollen accounting is now addressed. Flowers hold finite pollen separate from nectar. Actual nectar collection can also collect pollen, limited by flower stock and a separate 12 mg carrying capacity. Pollen is delivered at the hive, including by swarm members; death loses the carried amount. There is an independent conservation ledger. The frozen neural policy receives total carried mass capped at its existing load-input range. Pollen collection itself is an explicit accompanying rule, not a learned pollen-specialist strategy. A flower without collectable nectar cannot currently attract a dedicated pollen trip.

Larvae, rather than eggs or capped pupae, consume pollen through a simplified nursing budget. Sustained shortfalls can kill larvae despite abundant calorie stores. Queen-cell rearing also requires pollen. Care availability now counts workers aged at least three days doing interior work or resting/unloading at home, excluding migrants and absent foragers. Limited nursing capacity constrains laying; prolonged insufficient care causes larval loss. No additional workers are invented. Forager-to-nurse role plasticity is still simplified to availability at the hive; it is not a trained behavior.

Reference: [Penn State, queen development](https://extension.psu.edu/an-introduction-to-queen-honey-bee-development) describes pollen and nectar use by nurses producing royal jelly. [UGA, honey bees and beekeeping](https://extension.uga.edu/publications/detail.html?number=B1045&title=honey-bees-and-beekeeping) describes age-related task progression. The amounts, regeneration, 3-larvae-per-carer capacity and six-hour shortage accumulator are experimental parameters, not measurements fitted to these publications. Seasonal flight, winter physiology, thermal regulation and detailed diseases remain incomplete.

Validation: 32 regression tests passed, including finite shared pollen stock, pollen-only unloading, loss accounting, nursing absence, larval shortage/recovery, queen nutrition, swarming and checkpoint migration. Three one-hour runs with the actual neural network (32 workers, favorable conditions, seeds 481516/823/1364264692) delivered 3.624/3.504/4.896 g pollen and laid one egg each; both resource ledgers balanced within 0.0000001 g. These short integration checks do not validate long-term natural yields.

Legacy saves have no recorded pollen inventory. Migration starts that newly tracked resource at zero and records `nutrition_since`, without inventing past stores. Flowers then regenerate pollen through the explicit world process. Worker identities, nectar reserves and existing brood are retained; older reported results predate this change. New worlds initialize their declared pollen reserves explicitly in the ledger.


## Synthetic temperate day/night and thermal budget

The calendar starts at model day 135, 10:00 solar time. It is a reproducible fictional temperate habitat, not local weather. Seasonal daylight and smooth temperature determine flight restrictions and floral renewal. Night, temperatures below 13 C and heavy rain explicitly send foragers home, allow unloading and then rest. This is a visible engine override of the unchanged network, not a claim of learned nocturnal behavior. Queen mating requires daylight and at least 16 C; swarm departure waits for flight conditions while a viable queen cell remains.

Workers in the nest provide a simplified heat capacity. Heating consumes existing calorie stores. Sustained insufficient warmth causes brood losses; available food alone cannot compensate for too few workers present. Autumn/winter-born workers receive an explicitly longer modeled lifespan; existing workers are not rejuvenated. Winter laying is suppressed in this simplified seasonal model.

References: [UGA, daylight and temperature constraints on foraging](https://bees.caes.uga.edu/bees-beekeeping-pollination/pollination-protecting-pollinators-from-pesticides.html); [Penn State, seasonal colony management](https://extension.psu.edu/honey-bee-management-throughout-the-seasons); [PLOS One, overwintering thermoregulation](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0266219). The analytic climate, heat-capacity equation, resource costs, thresholds and lifespan ranges remain experimental approximations. This is not a validated thermodynamic, seasonal or disease model.

37 regression tests passed, including nighttime return/unloading/rest, resource-funded heating, cold brood losses, seasonal lifespan assignment and restart reproducibility. Engine rules, neural explanations and sky lighting use the same weather snapshot. `climate_since` records activation for legacy worlds.

Open audit items under the expanded request: continuous truthful accelerated trajectories, entrance/obstacle navigation, full foraging-cycle observation, decentralized source recruitment, long-duration combined validation. Smoothing sparse endpoint snapshots does not reproduce the paths traveled between them and must not be presented as doing so. Collective signaling should be modeled as communication, not asserted to be a shared consciousness.


## Recorded physical observations

Watching an accelerated experiment now requests a bounded recording of consecutive authoritative states at one-second intervals and plays it at x3. All worker identities, positions, actions and colony state in those frames come from the engine. Independent recordings are separated by an explicit time cut; no connecting flight is fabricated. The experiment keeps progressing independently. The recording covers up to 120 seconds, with a shorter window for large populations to bound memory at roughly 600,000 worker rows. No simulated individuals are removed.

The displayed colony counts and weather belong to the recorded time. The inspector displays recorded physical action/energy/load. Historical neural activations are unavailable and are explicitly not substituted with current activations. This is sampled observation, not an uninterrupted view of the entire accelerated lifetime. Paused/finished experiments need an existing recording or must resume to capture new movements. Live and old sparse historical views retain their separate behavior.

40 tests passed, including identical world evolution with/without recording, every-second frame timestamps, exact identity/action preservation, bounded capture and partial capture. These tests do not yet verify entrance routing or biological navigation; those remain open.


## Entrance and conservative obstacle navigation

The engine now routes departures through an aligned inner/outer entrance corridor and returns through the front opening to an internal unloading point. The hive-home offset is now (0.03, -0.7, 0.9), with an 0.08-unit arrival radius; earlier saves used an external point and a broad 0.3-unit radius. A low external corridor passes below the cocoon, and a visibility graph avoids a conservative cylinder enclosing the main trunk. Flower endpoints stay above the rendered terrain. These are explicit navigation rules and simplified collision volumes, not learned visual navigation or a simulation of every leaf, comb cell, root or branch.

Navigation routes persist across one-second steps and checkpoints. Existing positions are not teleported on upgrade; targets are adjusted and bees travel physically. Detours use the existing movement/energy budget and are included in recorded observations. Internal nursing locomotion remains unsimulated.

44 regression tests passed. New tests cover return from behind the hive without crossing the shell, entrance-aligned departure, trunk-avoiding segments and a full delivered trip with no per-second movement above the speed limit. Three one-hour runs with the frozen neural network delivered nectar/pollen successfully; both ledgers balanced within 0.0000001 g. These remain short integration checks, not biological calibration.

Performance on this host after the change: 10,000 workers, 34.7 ms per physical tick; 40,000 workers, 155.2 ms per tick, measured over 100 simulated seconds after warmup. All individuals were retained. Snapshot serialization measured 22.8/78.3 ms respectively. These numbers cover the engine and serialization, not browser rendering performance.


## Local source communication and a newly trained policy

A forager may advertise a source only after collecting from it and actually delivering at least 8 mg attributable to that source. A 30-second local signal can be heard by nearby foragers at the same entrance. Recruitment is probabilistic (50% per decision opportunity), weighted by reported yield and remaining signal time. Information expires when the dance ends, the sender leaves/dies or changes hive. A recipient stores the source and the sender's identity; recurrent neural states are never copied. The engine computes a small energy-costing entrance movement while sharing; recorded frames include this state. This is an explicit, simplified dance-inspired rule, not learned symbolic communication, full solar-angle dance decoding, or evidence of collective consciousness.

[Arizona State University, Bee Communication](https://askabiologist.asu.edu/honey-bee-communication) describes returning foragers communicating a source's direction and approximate distance. The model's timing, distance, yield threshold and recruitment probabilities are experimental. It does not model every sensory cue, odor, vibration or receiver interaction.

The previous trained network failed to deliver in a tested world when its perception was restricted to 2.5 units. It was not silently replaced in existing colonies. A separate candidate was trained with local perception and no remote stock knowledge: 1,800 imitation updates followed by 65,536 RecurrentPPO transitions. Across 20 held-out seeds, mean delivered nectar per episode was 0.64935 g for the candidate, 0.65591 g for the simple strategy and 0.031 g for random choices. The published comparison thresholds passed. These are controlled learning-environment scores, not natural honey yields.

Three one-hour complete-world tests with the new network delivered nectar and pollen, retained all 32 founders and balanced both ledgers within 0.0000001 g. Reports were produced, but these initial cohorts had no recruits in those short runs. A separate controlled full-engine fixture placed six newly mature foragers at the entrance during a genuine neural forager's report; bee 3 learned flower 20 from bee 29. The fixture tests actual recruitment, not a guaranteed benefit in every colony.

New experiments use the validated local-perception policy. Existing saved colonies retain their original policy and perception settings; the UI identifies legacy perception. Saved worlds are restored against a version-keyed policy registry. Missing weights cause an error, never substitution. Inspection probes and connection diagrams use the inspected world's own policy version. This preserves frozen weights for each colony lifetime.

50 regression tests passed, including local-contact requirements, hive separation, expiry, no unearned reports, stale remote knowledge, movement cost, independent memories and deterministic checkpoint recovery. Broader long-duration combined-model validation remains open.

## Queen fertility and laying bounds

Expiry of the modeled mating window now leaves an alive, unmated queen. She cannot lay worker eggs; the colony continues to consume resources and age. The UI retains her 3D presence and explicitly labels the fertility problem. Drone egg-laying, laying workers and supersedure are not modeled. Actual mating-flight loss remains a separate mortality event. The existing 21-day mating-window threshold remains an experimental approximation, not a biological death deadline. [FAO describes unmated queens eventually laying unfertilized eggs](https://www.fao.org/4/t0104e/t0104e05.htm).

Foundation colonies now cap worker egg-laying at 83 eggs per hour (1,992 per day), in addition to food, nursing, population and season constraints. This experimental ceiling is within the commonly reported 1,500–2,000 daily range, not an absolute biological maximum. [FAO queen production reference](https://www.fao.org/4/x0083e/X0083E03.htm).

52 regression tests passed, including a full 24-hour laying budget with 10,000 workers and persistence of an alive, infertile queen after the mating deadline. The running 28-day combined integration test was started before these two corrections; its results must not be cited as validating them.

## Comb construction conservation

Large colonies no longer gain comb directly from their population count. Both population modes pay for incremental comb growth from their own food reserves, with workers present at the nest and outside the modeled winter. A swarm starts with zero comb at its destination; the parent retains its existing comb. Migrating bees must arrive before construction begins. An empty nest must build before worker egg-laying resumes. Initial colonies still start with their explicitly configured starter comb; existing saved comb is retained.

The cocoon is a stylized shelter. Internal comb instances may now be zero rather than forcing one visible cell. Built units, construction speed and the 10 g food-equivalent/unit coefficient are experimental rather than a measured wax mass conversion. [FAO describes the food expense of wax production](https://www.fao.org/4/t0104e/t0104e05.htm). No biological yield calibration is claimed.

56 regression tests passed, covering paid growth in both modes, no construction without food or local workers, swarm comb separation, and construction before laying. The already-running 28-day integration test predates this correction as well.

## Current complete-engine performance measurement

Policy `278c39db4138ffef`, local perception, current social/navigation/construction rules: 100 simulated seconds after five warm-up seconds took 4.836 s for 10,000 workers and 20.843 s for 40,000 workers. Mean tick costs were 48.36 ms and 208.43 ms. Snapshots took 23.41 ms and 114.48 ms and contained all 10,000/40,000 living IDs. Nectar accounting passed. These short measurements support real-time computation on this machine, not arbitrarily fast accelerated simulation or sustained worst-case performance. A separate long integration process was running concurrently. Source hashes accompany the benchmark artifact.

Headless Chromium at 1200x900 was substantially slower graphically: median frames were initially 231.9 ms and 272.5 ms. Adaptive rendering now lowers resolution, shadow cost, foliage density and nearby model detail after sustained slow frames. Distant bee representations retain every identity and position; selected bees are prioritized for detailed display. Repeating the same two fixtures yielded medians of 111.2 ms and 137.0 ms, with all IDs still represented (32 detailed models plus the remaining point representations). No browser script errors occurred. This is an improvement, not proof of smooth interactive frame rates or the user's actual browser performance. The adaptation does not change simulation counts or physics.

## Full neural foraging trajectory check

An isolated favorable world with 32 founders and policy `278c39db4138ffef` was stepped for 3,600 consecutive physical seconds. Eight initial foragers made 2,715 harvesting steps and 269 deposits. Every deposit occurred at the internal home location and equaled the individual's previously transported nectar load. Each individual's cumulative delivery was bounded by its own actual collection. Harvesting only occurred within flower reach with the collection action. Across all workers, the largest step was 0.18000000000000055 world units, within floating-point tolerance of the 0.18 speed limit.

The trace recorded 543 crossings of the hive's front plane at the entrance corridor; no checked segment intersected the conservative main-trunk volume. A complete example trajectory for bee 11 and source hashes were saved with the validation result. This validates the actual policy and physical loop for this seed and weather scenario, beyond the heuristic navigation regression tests. It does not prove arbitrary terrain collision handling or biological speed/distance calibration. Recorded playback uses consecutive physical positions; sparse legacy history remains a separate, incomplete record.

## Restart and persisted movement playback

A copy of the long-run checkpoint at day 18 was resumed with the current engine for two physical hours. Saving and restoring halfway produced identical snapshots and private neural states for the remaining hour; worker and both food ledgers passed. The running original integration job was untouched.

Movement recordings now persist atomically as compressed JSON. New experiments automatically retain a bounded final clip even without a viewer. Persisted clips are immutable: they never append newer positions across an unrecorded gap. Old completed experiments with no clip are explicitly directed to recorded history; no past motion is fabricated. After a real server restart, all 121 states of experiment `a1aba61d-1afe-4903-9fc5-8630f6b8eb30` matched the original clip exactly. Playback was also observed in the user's existing in-app browser at x3 after the restart. 57 regression tests passed, plus the focused persisted-clip gap test after the final immutability adjustment.

The user's open tab had retained an older JavaScript interface. It was reloaded to activate the corrected movement display; saved worlds were preserved. Playback wording no longer implies that a completed experiment is still running.

## Age and task plasticity

Inspection of the original long run at day 24 found 25 surviving workers, all classified as foragers, substantial food and pollen, and repeated cold-brood losses. The previous unconditional 18-day transition left no permanent hive workers. Real honey bees can delay task transitions and some foragers can revert to nursing under altered colony demographics. [Experimental reversion study](https://www.sciencedirect.com/science/article/pii/S000334721200509X), [social reversal study](https://pmc.ncbi.nlm.nih.gov/articles/PMC2409152/).

The engine now retains younger carers when brood is present and allows unloaded foragers already at the entrance to begin reverting. No distant bee is teleported or relabeled as an immediately available nurse. Reverting workers contribute full nursing capacity only after two model days, retain their own neural states and age, and can return to foraging after brood demand disappears and their commitment expires. The two-day adaptation, daily commitment, three-brood/carer target and 65% allocation cap are explicit experimental simplifications, not fitted biological parameters or learned policy actions. Individual nursing locomotion and physiological gland changes remain unmodeled.

61 regression tests passed, including local arrival/load requirements, delayed capacity, retention/release and checkpoint identity. A new fresh favorable 28-day physical run was started separately with source hashes to assess the whole cycle under these rules. Its outcome is pending; no births or reserve refills are injected.

The earlier normal-weather 28-day run completed: 32 founders, one birth, 11 worker deaths, 22 surviving workers and 57 developing brood. It recorded 181 source-recruitment events (not 181 distinct bees). Worker accounting passed every day; final nectar and pollen residuals were about -1.25e-7 g and -2.67e-9 g. The poor renewal is evidence of a declining colony under the previous rigid task allocation, not successful population growth. This run predates the latest queen, construction and labor corrections; the fresh run is required before attributing a long-term outcome to them.

## Inspection after task changes

The inspector now separates current care rules from retained foraging memory. Returning carers no longer display an old flight choice as their current controller, and current input/action probability panels are cleared while their private recurrent state remains inspectable. Dead bees are labeled as having stored values, with no ongoing activity. The server only computes counterfactual policy probes for living active foragers. A focused explanation regression covers a former forager with a prior decision and a dead bee. The corrected care panel was verified on a real interior worker in the user's existing browser.

## Current-engine adverse scenarios

Four independent 32-founder scenarios used the actual local-perception policy, the current labor rules, seed 20260918 and six hours of consecutive physical stepping after scenario setup. Nectar, pollen and individual accounting passed in each case.

| Scenario | Workers surviving | Result |
| --- | ---: | --- |
| Persistent shortage | 0 | All 32 workers died; no births or food deliveries; colony extinct. |
| Flower recovery after one hour | 32 | Foraging resumed and five worker brood were laid. No adult births yet. |
| Queen lost before brood exists | 32 | Workers kept delivering food; no new brood, queen cell or fertile replacement. |
| Queen lost after four physical hours of laying | 32 | One of four existing brood was reserved as a queen candidate; three worker brood remained. No instantaneous fertile queen. |

The last scenario therefore covers ten physical hours overall. Stored result checkpoints and source hashes identify the exact engine under test. These short scenarios demonstrate distinct causal outcomes, not natural field survival probabilities or calibrated nectar yields. Long-term replacement-queen success is separately constrained by development, weather and mating rules; it is not established by this six-hour test.

## Measurement in the actual in-app browser

The renderer exposes a small read-only DOM diagnostic sample every roughly 1.5 seconds while the document is visible. In the user's existing tab at `2026-09-13T23:33:07.851Z`, the live scene represented 8,176 bees with a mean frame interval of 8.34 ms (about 120 fps), quality level 0, with no adaptive reduction. This differs substantially from the headless measurements above. It establishes fluid rendering for this actual viewport and live population, not a universal 40,000-bee frame-rate guarantee. No population or world state was modified to obtain this sample.

## Completed current-model lifecycle validation

The fresh favorable run (seed 20260917, policy `278c39db4138ffef`) completed 28 days by consecutive one-second steps. It began with 32 workers, one queen and the declared foundation resources. There were no injected births, age jumps or reserve refills. It ended with 119 living workers, 96 births, nine worker deaths and 309 developing brood. Day 22 had 24 births and day 23 had 48 cumulatively; the original eggs matured on their scheduled timeline. The queen remained fertile.

Worker accounting and both resource ledgers passed every day. Final residuals were about 7.94e-9 g nectar-equivalent and -1.47e-9 g pollen. Runtime was 1,232.56 seconds on this machine. Source-hash comparison confirmed that simulation Python modules were unchanged during the run; only server inspection handling changed. The final 61-test suite and current/dormant/dead explanation checks also passed. This is evidence of a successful lifecycle in favorable conditions, not a comparison isolating one parameter or a guaranteed growth trajectory in random weather.

The completed checkpoint was imported as separate experiment `e3197742-a90a-4e00-9897-ccea24ff49e9`, labeled **Growth example · 28 days**. Two additional consecutive physical minutes were recorded for immediate playback (2419200–2419320 s). The import preserved existing experiments and included a backup of their index. That final clip is an actual continuation, not a reconstruction of the preceding 28 days. Sparse history for this imported example contains the two actual endpoint snapshots only.

Final in-app checks confirmed 119 workers, 309 brood, the growth counters and recorded movement playback. **Inspect latest state** provides direct access to the selected experiment's present neural state, separately from replay; bee 1 showed policy `278c39db4138ffef` and 32 actual recurrent units. Returning to playback closes the current-brain panel, and opening recorded history from that mode stays on the same day-28 experiment. Latest-state inspection displays snapshots without fabricated connecting flights. The final generated JavaScript passed its syntax check. [Daily lifecycle results](growth-validation.json) contain the completed run and source hashes.