The Hive

Abstract

The Hive is a persistent artificial world in which a bee colony evolves under constraints involving food availability, development, weather and mortality. Its purpose is to make the relationship between individual decisions and collective outcomes observable. Each worker has a persistent identity and individual state. Foragers use a recurrent policy trained offline, with shared fixed weights and separate individual memories. Care, reproduction and selected execution constraints follow explicit rules. A Python server computes the authoritative world, while a browser-based 3D scene presents its states.

The current design uses token trading volume as an external environmental signal. A bounded rule changes the renewal of finite flower resources; bees must still discover sources, collect and deliver food. An isolated prototype reads a public market-data sample and depicts its flowering effect with additional gold-colored blossoms in 3D. It does not collect fees, transfer funds or credit a hive with automatic harvests. The public integration samples the same provider every sixty seconds; the experiments below retain their original frozen inputs.

An internal evaluation across twenty held-out seeds reports mean food delivery of 0.6493 g for the neural controller, compared with 0.6559 g for a heuristic and 0.0310 g for a random strategy. A separate favorable twenty-eight-day simulation increased a colony from 32 to 119 surviving workers. These observations demonstrate selected mechanisms within the model; they establish neither general superiority of learning nor predictive accuracy for real colonies.

Keywords: agent-based simulation, artificial colony, reinforcement learning, LSTM, finite resources, transparency, virtual habitat, token.

1. Motivation and Research Questions

An animation can appear alive without its movements having consequences. The Hive instead connects visible effects to computed events: a bee must reach a source to harvest, transport its load to provision its colony, and thereby contribute indirectly to adult maintenance and brood development.

The primary research question is: to what extent can local decisions, finite resources and explicit demographic rules produce understandable and reproducible collective trajectories?

A second question concerns the external signal: can a bounded volume-to-flowering rule measurably influence resource flows without replacing ecological constraints with automatic growth? Three short paired experiments now explore this question; long-term effects remain untested.

The project has three objectives: an accessible public observation experience, an inspectable simulation engine, and a transparent protocol for environmental interventions. It does not claim to reconstruct a biological bee brain. The cocoon-shaped hive, landscape dimensions and some vegetation are artistic representations. An additional visible plant is not necessarily an additional simulated resource.

2. Scientific Positioning

Colony models connect population dynamics, foraging and food availability. BEEHAVE provides a comparative reference for an integrated treatment of these interactions [1]. The Hive draws on this modeling perspective without inheriting BEEHAVE's results: its autonomous engine does not add harvests computed by another simulator.

The foraging controller uses Proximal Policy Optimization, a reinforcement-learning method [2], through the recurrent RecurrentPPO implementation [3]. Using this algorithm does not by itself make the resulting behavior realistic. Biological validity also requires comparisons with field observations. Published evaluation work on BEEHAVE illustrates the importance of that step [4]; it has not been performed for The Hive.

3. World State and Time

The server is the authority for the simulated world. Physical transitions use one-second simulation steps. The live world uses two clocks. Lifecycle time advances at 240x for laying, brood development and aging; six real minutes correspond to one lifecycle day. Movement, harvesting, maintenance demand, nursing activity, construction and weather use normal physical time. More individuals can still increase total food demand. Egg production retains its per-egg resource cost. Transfers remain conserved and deliveries require arrival. Faster development changes lifetime resource demand: this is a deliberate experimental model, not a biological calibration. Historical evaluation results below predate this change. Accelerated experiments run in isolated worlds. Playback displays recorded states and does not modify the live colony.

A worker's state includes identity, hive membership, age, energy, position, nectar load, pollen load, action, role and memory. Lifespan and loss mechanisms complete this state. The queen and brood follow their own rules; they are not workers driven by the foraging network.

Each hive maintains a queen or succession state, workers, brood, food stores and construction. Flowers have finite stocks that renew according to the synthetic environment. Weather, temperature and the light cycle constrain available actions. Random episodes use a generator whose state is saved.

The transition can be summarized as:

S(t + 1) = F(S(t), A(t), ξ(t), θ)

Here S is the complete world state, A denotes executed actions, ξ represents random draws, and θ contains model parameters. F is the implemented transition, not a calibrated biological equation. Exact reproduction also requires compatible code, dependencies and numerical execution, not merely the same seed.

4. A Neural Controller with Individual Memory

The evaluated policy contains a 32-unit LSTM followed by a 32-unit decision network. Fourteen inputs describe energy, load, directions and distances to the hive and a known flower, plus indicators for source knowledge, location and rain. Perception is local, and sources can be remembered.

The six actions are explore, visit a flower, collect nectar, return home, unload and rest. Decisions are reconsidered every five physical seconds in the current live mode. Network weights remain fixed during a world's lifetime. Each bee retains its own activity vector h and memory vector c:

(pᵢ, next hᵢ, next cᵢ) = πθ(oᵢ, hᵢ, cᵢ)

The output pᵢ is a distribution over actions. During execution, the engine selects the highest-ranked physically executable action. Explicit rules can override that choice, including returning home under unfavorable conditions. Raw network preference, feasible actions and actual execution must therefore be distinguished.

Weights are shared across workers. A population of 32 bees does not imply 32 independently trained networks. Individuality comes from each bee's state, experience and recurrent memory. Workers assigned to care may preserve dormant foraging memories without making learned nursing decisions.

The training report records initialization through 1,800 gradient steps imitating a simple heuristic, followed by 65,536 PPO steps. The resulting policy must not be described as a strategy discovered entirely without guidance. The evaluated policy's abbreviated fingerprint is 278c39db4138ffef.

5. Resources, Development and Collective Behavior

Collection removes nectar from a source and adds it to an individual's load. Unloading transfers that load into hive stores. Travel, maintenance, care and construction incur costs. Pollen is tracked separately. Adults and brood can die; stores can run out and a colony can become extinct.

For nectar-equivalent food, the accounting invariant is:

Total stock(t) = Initial stock + Cumulative regeneration − Cumulative consumption − Cumulative losses.

Total stock includes flowers, carried loads, hive stores and delivered brood-food buffers. Harvests and deposits are internal transfers, not additional resource creation. Quantities are expressed in model food-equivalent grams; a complete nutritional calibration has not been established.

For workers, without external introductions:

Living workers(t) = Initial workers + Cumulative births − Cumulative deaths.

Swarming transfers existing individuals and provisions between hives. It does not duplicate the population. Each hive maintains separate stores and construction.

Worker development takes 21 lifecycle days in the model, approximately 2 hours 6 minutes at the current 240x lifecycle rate, conditional on survival. At 32 workers, the foundation rule schedules up to one egg per lifecycle hour, approximately every 15 real seconds, subject to capacity, resources and the queen reaching an available cell. A fertile queen, food and sufficient care capacity support continuation of the lifecycle. Retaining workers in care roles and reassigning some returning foragers to care are explicit mechanisms, not learned nursing behaviors.

Local communication can transmit source information after an actual delivery. This changes the knowledge available to nearby workers without merging their recurrent memories. It is not evidence of collective consciousness.

6. Observation and Interpretability

The public interface lets visitors follow a bee, inspect its action and explore its 32 recurrent units. Positive and negative activations do not automatically mean hunger, fear or intention. The graph shows selected recurrent connections rather than the entire computational network.

Explanations distinguish observations, memory, action probabilities, constraints and execution. Network probes can alter one input while holding the preceding memory fixed. Such probes describe controller sensitivity, not established biological causation.

New experiment observation clips also record synchronized neural states. Older clips and sparsely sampled history may not contain them. Their absence must be reported rather than replacing a past state with the current network state. Recording coverage is bounded: the project does not yet maintain an exhaustive permanent neural archive.

The interior uses three experimental comb surfaces with identified cells and routes between them. The queen reaches a cell before laying, and indoor workers travel to task destinations. Existing brood was assigned cell locations when this representation was introduced; historical positions were not reconstructed. Food remains pooled at hive level: displayed cell amounts partition that pool rather than implementing independent cell inventories.

The volume demonstration is tangible in the 3D meadow. Each bouquet represents an existing finite nectar source. Additional gold-colored blossoms identify the volume-linked flowering contribution; they do not each instantiate a new independently harvestable source. Their placement remains within the represented source area. Nectar availability also affects blossom size. The additional blossoms grow progressively with the applied multiplier; their geometry is an illustration of environmental support, not a direct count of nectar units.

A meadow-focused camera and a link to the matching control run let a visitor compare the same seed with and without the signal. These are recorded experimental states. The change is not a live response to a purchase, nor an exact botanical measure of flower abundance.

7. Available Results and Their Scope

The following measurements come from internal prototype reports [D1–D3]. They belong to the recorded versions; later changes do not automatically inherit their validation.

These averages are not a statistical significance analysis. One favorable run does not estimate survival probability or expected growth under natural conditions. Conservation tests establish software properties, not realism of the model's coefficients.

8. Volume as an Environmental Input

The selected mechanism is now trading volume → bounded flowering renewal → finite nectar and pollen availability → physical collection → possible colony effects. Fees are not the input to this implementation. Volume is a signal used by the simulation, not a financial budget and not a biological resource in itself.

The laboratory adapter reads DEX Screener's token-pair endpoint [5], checks the reported chain, configured pool and token identity, and rejects missing, ambiguous or invalid volume values. This is provider-reported market data, not independent verification of finalized transactions on BNB Chain. The pool is pinned rather than automatically aggregating every listed market.

For an hourly volume sample V, in the provider's USD units, the implemented laboratory rule is:

m(V) = 1 + 0.25 × V / (V + 10,000), with V ≥ 0.

At zero volume, m is 1. The additional renewal approaches, but does not exceed, 25%. The 10,000 scale and 25% cap are explicit experimental choices, not fitted biological or economic constants. Overlapping hourly volume windows are not summed into a cumulative budget.

The multiplier is applied to the environment's floral renewal factor. Nectar and pollen still have finite source capacities. Increased renewal enters the existing regeneration ledgers; harvesting and unloading remain internal transfers. The rule never changes a bee's load, creates an egg or increments a hive's food store directly. If weather suppresses renewal, or stocks are already full, the realized resource effect may be limited.

Test asset: BNB Chain token 0x9640554b8aef28a19a442cdf0c5791dadaa77777, pool 0xd17cb3bfd184463042b11340ba86fa988429abef. These identify a third-party test input, not an issued Hive token or an endorsement. The frozen sample was captured at 2026-09-14 20:34:26 UTC, with reported hourly volume of $44,052.98. It yields m = 1.203749, approximately +20.4% to the renewal setting.

9. Paired Experiments and Observed Influence

The first comparison used an already well-supplied environment. Extra renewal produced no additional delivered food: both runs delivered 11.96 g. That unchanged outcome is retained in token-test-report.json [D4]. A resource increase is not automatically a useful increase.

The second demonstration deliberately starts with depleted flower nectar and pollen, while retaining the same hive reserves and favorable weather. Three seeds were selected before the runs: 20260914, 20260915 and 20260916. For each seed, one control and one treatment use the same initial state, neural model and lifecycle settings. Only the flowering multiplier differs. Each run covers 3,600 physical seconds, equivalent to one lifecycle day. The same seed does not require identical later actions: the changed environment can alter decisions and subsequent trajectories.

Mean delivery changed from 7.3295 g to 7.4639 g: +0.1344 g per run, approximately +1.83% relative to the control mean. Two seeds improved and one declined. All six runs ended with 32 workers and 23 developing bees; there was no observed population benefit over this horizon.

Nectar regeneration increased from 31.0238 g to 37.3449 g in each constrained pair. This is the direct effect of the programmed renewal rule. The smaller, variable change in actual delivery is the downstream outcome of physical foraging. Final nectar and pollen accounting residuals were below 10⁻⁷ g in all six runs. Full per-run values and 30-second trajectories are in token-influence-report.json [D5].

These results demonstrate that this implementation can alter environmental resources and sometimes food delivery. They do not establish a statistically reliable effect across arbitrary seeds, a long-term survival advantage, a relationship between token ownership and growth, or a causal effect of buying the token. The coupling coefficient is imposed by the software. Three short pairs are a demonstration, not a market or ecological validation.

10. From Recorded Samples to a Continuous Integration

The laboratory runner freezes one sample per experiment. Separately, the public server polls the pinned pool every 60 seconds on a background thread. Accepted samples and their simulation application times are appended to a durable local input log before use. Checkpoints retain the applied multiplier and latest input; logged interventions after a checkpoint are reapplied at their recorded simulation times during recovery. Fresh market samples are only applied once catch-up reaches the present, never retroactively to an offline interval. The public 3D meadow displays the reported hourly volume, sample time, data status and actual applied renewal boost.

Rule live-volume-v1 uses the same capped target as the laboratory rule. The applied setting approaches that target exponentially with a 300-second time constant. After 300 seconds without a successful sample, the status becomes stale and the target returns to 1; the applied support fades gradually. Missing data is explicitly distinguished from zero volume. The 1 to 1.25 cap is enforced by the engine. No rolling window is accumulated as funding. The timestamp indicates when the provider response was received, not independent confirmation of trade freshness. Manipulated volume or a provider repeatedly serving obsolete values remain limitations. Additional golden blossoms grow with the applied support, while actual nectar reserves influence their size. This indicates an environmental setting, not proof of additional food delivery or births.

Further tests should cover falling and rising volume, prolonged outages, stale samples, pool migration, provider corrections and abusive or artificial trading activity. A provider's reported volume is not necessarily unique economic demand. Smoothing and caps reduce abrupt visual changes but do not establish data integrity.

No real fees are received or accounted for by this implementation. The experimental credit economy described below is separate from the volume signal. Its existence does not establish a connection to launchpad payments.

11. Trust, Administration and Economic Boundaries

Simulation runs off-chain on a server. A market-data sample does not prove faithful execution of the simulation. The prototype currently depends on its operator for execution, availability and data retention.

Separating public and administrative access protects controls but does not prevent the operator from editing files. A durable intervention log, published versions and periodic cryptographic commitments to state could improve auditability. These measures are not all implemented today. The current event journal is bounded and is not an immutable administrative ledger.

Under the present proposal, holding the token would not automatically confer ownership of reserves, revenue distributions or a promised yield. The proposed role is to connect observable market activity to a shared virtual habitat. The simulation does not technically require a token: another external signal could drive the same experimental rule. The distinctive element is the chosen relationship between the token's activity and the world, not an exclusive technical capability or a promised financial return.

Before launch, a separate specification must define issuance, allocation, fees, beneficiaries, control keys, treasury policies and any governance mechanism. This paper does not determine the legal classification of an undefined asset or replace an assessment of the eventual arrangement.

11.1. Individual Budgets and the Operator Treasury

The proposed financial path is: verified creator trading fees enter a dedicated receipt wallet, an allocation rule assigns budgets to individual bees, and bees can spend those budgets on useful simulated services supplied by the operator's treasury. The treasury is the operator's seller account; bees are software agents controlled by the project, not independent human beneficiaries. This design does not distribute investment returns to token holders.

Flap on BNB Chain (chain identifier 56), with native BNB payouts, is the selected launch path. Collection is separate from the operator treasury. The launched token and its payout contract must be verified before receipt ingestion is enabled. Wrapped assets are not interchangeable with native BNB. No percentage of displayed volume is credited as cash. Real launchpad receipts are not connected today.

Implemented credit layer. Accounts use non-negative integer simulation credits. An explicitly test-only receipt allocates equal whole-credit shares among workers alive at that moment. Division remainders remain unallocated. New workers begin with zero credits; migration retains the same individual account; dead bees retain a frozen balance without an automatic sweep to the treasury. Duplicate test receipt identifiers are idempotent; conflicting duplicates are rejected. Public live accounts start at zero and cannot be funded through the test-credit function. Credits are neither BNB nor ETH and have no redemption or exchange-rate promise.

The accounting invariant is:

issued test credits = all bee balances + treasury credits + unallocated credits

Reference decision graph, version economic-rules-v1. The preserved baseline evaluates every five physical seconds. The graph first checks survival, migration and a five-minute purchase cooldown. If the bee is near home, its own hive is below its food threshold, the bank has stock and the bee can pay five credits, it purchases 20 mg of food. Otherwise, a forager lacking a source may pay two credits for source information if its energy is at least 40%, flight is permitted and a usable source exists. All other cases save the balance. Prices and thresholds are experimental, versioned constants, not market valuations. Execution order rotates deterministically to avoid always assigning scarce stock to the lowest bee identifier.

Food purchases debit finite bank stock and credit the buyer's own hive stores. Test bank stock is transferred explicitly from existing hive food, not generated by payment; total food includes bank inventory. Delivery is an immediate model transfer near the hive, not a simulated courier journey. Information purchases disclose the bank's nearest currently usable source and update the bee's flower memory; they do not force a flight, guarantee collection or bypass weather. This is a deliberate information service using global source knowledge, separate from local perception. Budgets and consumption are not proof of improved colony outcomes.

Trained economic network. A separate RecurrentPPO controller has 12 observations, 16 LSTM units, a 32-unit policy head and three actions: save, buy source information, buy food. Each bee retains its own hidden and cell states in the checkpoint. Weights are frozen during operation. Observations describe budget, food relative to need, bank stock, energy, forager role, source knowledge, home proximity, source availability, flight conditions, cooldown and the two prices. Prices are fixed during this training; generalization to new prices has not been established. Decisions use a five-minute physical interval, matching the training window and independent of the 240x lifecycle clock.

The learning reward is 100 food consumed - 300 unmet food demand - 0.02 credits spent - 0.5 invalid action. There is no reward for operator revenue. Training uses a reduced single-agent environment with finite food sources, finite bank stock, changing weather, occasional test-credit income and delayed foraging benefits. It does not simulate full colony demography, actual flight paths, blockchain settlement or transaction fees. Food consumption and shortage are proxy objectives; survival benefits in the full world remain unvalidated.

Four earlier candidate networks failed the original strict comparison. The final network combines 409,600 PPO environment transitions and 11,200 imitation updates, including randomized contexts and an information-first teaching heuristic. Its last refinement is supervised imitation: this is teacher-trained behavior, not independent discovery of an optimal economy. The final policy hash is 0563131e9cd2eb5e.

On 100 held-out episodes (seeds 78000 to 78099), mean reward was 2.757 for the network, 2.928 for the original food-first rules, 2.697 for feasible random actions and -16.807 for never spending. The original strict gate therefore failed. We do not claim superiority over the reference rules or over the information-first teacher. The paired reward difference versus rules was -0.172, with a seed-bootstrap 95% interval of approximately [-0.526, 0.159].

A revised experimental deployment criterion was announced before a separate 500-seed evaluation (79000 to 79499): the upper bound of the paired 95% bootstrap interval for additional unmet food must remain below 0.002 g per 48-window episode. The observed mean difference was +0.000200 g, with interval [-0.000301, +0.000664] g. This passes the specified tolerance, not a superiority test. The margin is an engineering tolerance chosen for an unfunded credit simulation, not a biologically calibrated safety threshold. Seed bootstrapping does not cover uncertainty across independent training runs.

All final per-seed results, earlier candidate reports, the failed original gate, the revised gate and exact weights are published. Two intermediate evaluation runs accidentally reloaded an older checkpoint; those were not new model evaluations and were excluded from acceptance. Checkpoint publication was corrected before the final trials. A separate 600-second full-engine integration run verifies conserved food and credits; it does not validate long-term biological outcomes.

At execution, hard affordability and availability masks remain explicit code. Save is always available. The highest-probability feasible action executes, so an impossible high-scoring purchase cannot create a negative balance or phantom stock. The published probabilities are unmodified policy probabilities, not success estimates. Missing or mismatched saved economic weights fail explicitly rather than silently switching back to rules. The existing foraging LSTM remains separate and can react to newly purchased source information through its normal inputs.

The bee inspector displays the economic probabilities, availability mask, last decision inputs and actual recurrent activations. No semantic label such as hunger or intent is assigned to an individual hidden unit. Identical weights with separate memories do not establish individuality or consciousness.

Wallets and custody. An initial set of 34 EVM wallets has been generated: one treasury, one fee-receipt wallet and 32 individual worker wallets, mapped to the live experiment and immutable worker identifiers. Each private key is held in a password-encrypted Ethereum V3 keystore (scrypt). Encrypted files are stored outside the website in a root-restricted AWS directory. The randomly generated decryption password remains on the operator's Windows computer, protected using that account's DPAPI; it is not uploaded to AWS. Public addresses do not expose signing keys. These are software keystores, not KMS/HSM-backed wallets. A compromised Windows session can access the recovery secret, and loss of the Windows profile without a recovery backup can make the wallets inaccessible. New workers receive an account immediately but require a separate local wallet-provisioning run. No automatic signing or payout service is active.

HTTPS protects traffic to the site and does not substitute for key custody. No private key, keystore or recovery secret is served through the public API. The current wallet addresses should not be funded until secure recovery and the final chain-specific settlement path are configured. The same EVM address can exist on multiple chains, but balances and nonces must remain separate by chain and asset.

Scope of evidence. Seven economic tests cover credit conservation, duplicate handling, unaffordable choices, paid information, finite stock, cooldowns, frozen dead-bee balances, newborn accounts and checkpoint equivalence. The public economy page shows deterministic isolated decision fixtures from economy-report.json, not on-chain payments or learned behavior. The 200-entry simulation purchase journal is bounded and is not an immutable financial audit log. Live financial settlement still requires a durable receipt ledger, finalized-chain reconciliation, reorganization handling, idempotent payment intents, destination and spending limits, gas reserves and recovery procedures. These protections cannot be replaced by the simulation's credit ledger.

11.2. Native BNB Settlement Rehearsal

A separate durable ledger implements receipt deduplication, exact integer wei accounting, reserved distribution batches, daily limits, purchase intents and confirmation-gated service effects. A private receipt verifier checks chain 56, successful transaction status, confirmation depth and canonical block identity. Internal call traces admit native BNB only from a pinned payout contract to collection. Reverted subtrees and delegate calls cannot create income. This relies on a configured RPC and is not an independent consensus proof.

An optional payment executor checks registered recipients, exclusive sender nonces, available balances and gas limits. Signed bytes are persisted before broadcast; uncertain attempts reuse the same transaction. A separate Windows keystore adapter implements signing, but real signing is disabled. No real wallet was unlocked for a signature during this development step. This custody arrangement requires the operator's computer and is not an always-on AWS signer.

Twenty-seven fixture tests cover receipt, accounting and executor boundaries, including restart recovery, failures and ambiguous broadcasts. A six-step rehearsal is published on the Fee collection page. Its receipt, confirmations and purchase choice are synthetic. The accounting residual is zero at every step. These tests are not evidence of executed blockchain payments or learned financial behavior.

The invariant is verified receipts = available account balances + pending payment and gas reservations + consumed gas. BNB prices and caps remain unspecified. The credit policy is not automatically a BNB policy. Automatic payout discovery, continuous reconciliation, portable key recovery, off-host backups, the live finite-stock/service-effect adapter and an end-to-end chain trial remain required. Details: settlement-guide.md; synthetic data: settlement-report.json.

12. Further Validation Protocol

A stronger evaluation should compare all three controllers across multiple independent training runs and held-out seeds and landscapes. Outcomes should include delivered food, energy costs, mortality, births, survival and shortage duration. Distributions, uncertainty intervals and failures should be published alongside averages.

Extend the paired tests across many seeds, resource conditions and full generations, under zero, stable, rising and falling volume signals. Measure whether benefits pass through collection and whether caps preserve a possibility of decline. Volume schedules are experimental inputs, not forecasts of token markets.

Infrastructure tests must cover save and restore, restarts, duplicate transactions, chain reorganizations, indexing delays and server outages. Performance evaluation must include neural clip size, concurrent visitors and memory consumption as well as worker counts.

Biological validation would require independent observations, calibrated units, sensitivity analysis and comparison between observed and simulated trajectories. Detailed diseases, drone populations, laying workers and learned nursing remain outside the current scope.

13. Implementation Status

The present prototype makes a bounded external influence observable without crediting automatic harvests or births. Its next milestone is longer comparative evaluation and assessment of market-data quality. Neither that roadmap nor the demonstrations imply guaranteed colony or token growth.

References

[1] Becher et al. (2014). BEEHAVE: a systems model of honeybee colony dynamics and foraging to explore multifactorial causes of colony failure. Journal of Applied Ecology. https://doi.org/10.1111/1365-2664.12222

[2] Schulman et al. (2017). Proximal Policy Optimization Algorithms. https://arxiv.org/abs/1707.06347

[3] Stable Baselines3 Contrib. Recurrent PPO. Technical documentation, accessed 14 September 2026. https://sb3-contrib.readthedocs.io/en/master/modules/ppo_recurrent.html

[4] An Evaluation of the BEEHAVE Model Using Honey Bee Field Study Data: Insights and Recommendations (2019). Bibliographic record: https://pubmed.ncbi.nlm.nih.gov/31343774/

[5] DEX Screener. API reference: token-pairs endpoint. https://docs.dexscreener.com/api/reference

Prototype Evidence and Reproducibility

[D1] Local training report: policy 278c39db4138ffef, held-out seeds 9000–9019, three compared strategies. A copy of the measurements accompanies this paper in the-hive-paper-evidence.json.

[D2] growth-validation.json: daily trajectory of the favorable twenty-eight-day experiment and source fingerprints. This is a historical experiment, not the current public colony.

[D3] neural-validation.md and simulation-audit.md: internal checks and limitations. Some passages describe an older neural playback implementation; this paper specifies its present coverage.

[D4] token-test-report.json: the initial abundant-flower control and treatment, frozen input metadata and accounting checks.

[D5] token-influence-report.json: all six constrained runs, recorded trajectories, seed identifiers and accounting checks. Neither report represents a continuous public deployment.

For academic submission, add authors and affiliations, archive code and dependencies for each experiment, fully specify training and evaluation protocols, and deposit the data in a permanent repository. No DOI, institutional affiliation or peer review is claimed for this document.

[8] Microsoft. Windows account protection with DPAPI. https://learn.microsoft.com/en-us/powershell/module/microsoft.powershell.security/convertfrom-securestring

[9] eth-account. Account creation and encrypted keystores. https://eth-account.readthedocs.io/en/stable/eth_account.html

[10] Flap documentation. https://docs.flap.sh/flap

[11] PONS documentation, version-specific fees. https://docs.ponsfamily.com/v2

[D6] economic-neural-report.json: all economic policy results, original and revised acceptance criteria, prior attempts, runtime versions and model fingerprint.

[D7] economic-world-validation.json: full-engine integration and conservation checks, one seed.

[D8] economic-neural-evidence.zip: frozen economic weights and training source files.