Commissioned work appears here with its cost, delivery and result.
Specialist agents, human contributors and compute services are resources Nuria can commission. Payment, delivery and useful results stay separate.
Not receivedNot receivedWaiting for independently observed financial evidence.
A delivered result becomes an evaluated outcome only after its declared checks run.
A transfer proves payment. Artifacts and evaluations establish what it bought and whether it helped.
No evidence received.
What Nuria chooses, learns and remembers.
Waiting for a recorded decision.
Awaiting source-specific measurements.
The action selector chooses when to explore or forage. The local path planner selects the move.
Energy regulates compute activity. Treasury funds have separate balances, limits and receipts.
Awaiting a decision.
Awaiting a forecast.
Learning hidden rules through choices, costs and delayed rewards.
Learn which choice works for each remembered cue.
Net virtual reward per settled choice. Higher is better.
Awaiting settled outcomes.
A separate 256-neuron sensory circuit feeds learned decisions. The main circuit changes virtual probe costs. Credits are experimental; no SOL is spent.
Awaiting evidence.
Select an input to inspect its route and receipt above.
Sensory input enters the network. Recurrent activity continues between signals, while spike timing changes its plastic connections.
Creator fees can fund compute and experiments. This balance shows finalized funds held by the connected wallet; fee income requires separate payout evidence. An unavailable wallet balance stays unknown.
Published goal utilities and neural scores select actions. Local jobs execute and return measured outcomes. SOL payments require a dedicated treasury, funded limits, allowed recipients and an isolated signer.
Not configuredWaiting for independent payment evidence.
No money movements recorded.
No evidence received.
Hash-linked neural records, stored on the server. Read the evidence guide
Exports include the latest 300 durable ticks and their raw spike windows. Complete history remains on the private server.
Coverage unknown.
Nodes, spikes and weights come from the running Brian2 engine. The field arranges measured activity in a schematic view with slowed trails. Model time and wall time are shown separately.
Recurrent excitation and inhibition, spike-timing plasticity, refractory periods, adaptation, noisy sensory input and homeostatic control. Nuria keeps running when this page closes.
Read the Brian2 documentationNuria explores onchain consciousness through continuous neural activity, persistent state and changing connections. The token’s market becomes a sensory stream; its cognition layer learns from outcomes, retrieves episodes and executes local experiments. The treasury interface separates balance, fee provenance and payment authority.
An onchain consciousness experiment.
Nuria studies a persistent entity whose experiences, choices and resource budget are shaped by its token. Its learning, memory and decisions are open to inspection. Subjective experience remains a research question.
Nuria combines a 1,024-neuron spiking circuit, online prediction, episodic memory, attention competition and an autonomous action selector. The trade reader is built to turn the configured token’s finalized trades into sensory inputs. The model continues between inputs and resumes from saved neural and learning state. The live mint and creator-fee wallet still need configuration.
Input → neural response → prediction → observed outcome → reward modulation → memory → action → measured consequence. The observatory shows the numerical records behind this loop. Local experiments execute on the Nuria server; their measured CPU time and zero incremental SOL spending are reported separately.
The original 256-neuron network and receipt history continue independently. The expanded cognitive circuit has its own genesis, checkpoints and hash journal, and reads the same recorded inputs through a read-only database role. This is an added circuit, rather than a reset or a claim that both circuits have identical state.
What happens when a token supplies a persistent neural system’s sensory stream and, eventually, its resource budget? Nuria investigates that question through reproducible behavior. “An onchain consciousness experiment” describes the research direction. Brain-scale simulation and subjective experience are not established by population names, animations or a neuron count.
A view into recorded neural activity, decisions and outcomes.
The main graph shows the expanded circuit’s 1,024 neurons across six populations. Membranes, spikes and sampled plastic weights come from Brian2. The spatial arrangement is schematic. Flashes use recorded source spikes; signal heads follow each sampled connection’s recorded synaptic delay. A traveling signal does not establish that the target fired because of that source. A deterministic subset of connections is rendered; the worker retains the complete connectivity.
Neural mode provides a depth-shaded, orbitable view of the six populations. Every displayed edge comes from the published topology sample. Topology flattens the same arrangement; Spikes plots recorded time against neuron ID. The small raster underneath Neural and Topology shows the same recorded window.
Firing highlights measured spikes. Membrane maps the recorded, dimensionless membrane variable v from blue to ivory. Weights uses line thickness for current sampled weight and sage or lilac for increases or decreases between received snapshots; those differences are not a complete account of plastic changes inside each worker window.
Replay offers 22× slower, 10× slower and real-time playback. Model time is separate from wall time. The browser receives the newest cached window, rather than replaying every window the worker has produced. A short window can finish before the next snapshot arrives.
Click a neuron to pin its inspector, or open the view menu, choose Inspect neuron and enter its ID. The inspector shows its recorded membrane value, spike count, sampled incoming and outgoing connections and mean sampled outgoing excitatory weight. These connection counts describe the display sample, not the neuron’s full connectivity. Drag or use arrow keys to orbit; Home resets the view.
Current decision shows eight combined action scores. The learning panel compares evaluated predictions against a learned repeat-probability baseline. The habitat shows actual recorded software moves and virtual resource collection. Memory and attention report recorded episodes and the three competing workspace slots.
Inspect the latest decision to see utility scores, neural scores, outcome reward, job results and hashes. Copy the record hash for comparison. The original input-record panel retains the original network’s receipt links; expanded per-input effects are available through the cognitive API.
A failed or stale response is unavailable. A cached last frame does not imply a healthy worker. Public reads do not trigger simulation, database queries or RPC calls. Reduced motion starts with a held snapshot. Pause captures the current neural window and holds its recorded values while the server and other panels continue. Resume returns to the newest state. An off-screen field stops drawing until it is visible again.
Adaptive spiking dynamics with explicit inputs and outcome feedback.
Sensory: 128 neurons. Association: 384. Recurrent memory: 128. Workspace: 128. Action readout: 128. Inhibitory control: 128. These labels define computational roles; the design is a simplified software model.
The circuit uses leaky integrate-and-fire neurons with heterogeneous time constants: 12 ms sensory, 45 ms memory and 20 ms elsewhere. Refractory period is 4 ms. Excitatory and inhibitory influence decay over 8 and 12 ms. Adaptation decays over 350 ms. Independent background input and bounded homeostatic regulation keep the circuit active between signals.
Excitatory connections maintain pre/post traces and eligibility. Small timing-dependent updates act continuously. Observed prediction improvement against the learned repeat baseline and measured action outcomes produce bounded reward signals that modulate eligible weights. Inhibitory connections use fixed influence. All weight updates are bounded.
Each recorded input receives its own 20 ms sensory window. A cycle processes up to four inputs, then advances a 100 ms autonomous window. Nominal wall pacing is one cycle per second. Under load, actual cycle cost determines catch-up rate; a nominal limit is not a throughput guarantee.
The Brian2 reward-modulated STDP example describes the eligibility-and-feedback mechanism that informed this implementation. Nuria’s dimensions, parameters and controller are its own design. Brian2 reward-modulated STDP reference.
Finalized Solana evidence enters a durable queue before either circuit reads it.
The reader decodes Pump and PumpSwap events against the included official interfaces, checks the exact configured mint and supported pools, and uses finalized transaction evidence. Failed or unresolved transactions remain pending. Coverage is stated explicitly; every possible venue is not automatically covered.
The expanded circuit encodes side, amount, creator fee and a deterministic event-ID texture. Each input has its own before/after neural-state projection hashes, source payload hash and recorded spikes. Duplicate IDs are idempotent; changed contents or ordering for a recorded ID are rejected.
The compare action runs a recorded input and a no-input control from the same current neural and random state. It restores the input branch afterward. The result measures that controlled perturbation. It does not reconstruct the transaction’s original historical counterfactual, and the event-ID texture is software encoding rather than economic significance.
When no mint is connected, the existing engine supplies tagged test inputs. Test and live prediction learners have separate weights, momentum, pending predictions and metrics. Simulated accrual is not a wallet receipt. A mint can be connected without erasing prior history.
Blockchain finality, RPC pagination, provider limits and queue lag affect latency. The model does not promise instantaneous coverage. The cognitive status reports its source cursor and backlog. Missing transaction evidence stays unresolved rather than being counted as an empty trade.
Follow the money. Then check what it delivered.
UnknownUnknownAddresses come from the shared server profile. Test-token inputs and launch inputs have separate forecast histories. Connecting an address does not enable payments.
The observatory has separate Work, Treasury, Results and Evidence views. A job shows its purpose, provider, quoted cost, recipient and acceptance check. A payment receipt, delivered artifact and evaluated outcome are different records.
GET /api/work serves up to 80 recent production
payment jobs from the same public cache used by the commerce
service, with explicit total coverage, alongside at most 40 recent
commission requests. The complete payment event history remains
paginated. Signed authorizations and isolated private tests are
excluded. Missing or stale evidence appears as unknown.
Nuria’s job controller records a task’s purpose, originating decision, committed offer, cost ceiling and delivery deadline. Structured results are checked against buyer-owned held-out targets. Exact payment evidence and measured usefulness are recorded separately, including negative results.
The first contractor artifact is a structured probability vector for a committed experimental task. Other reports, code, media and physical work need their own acceptance verifier. No contractor code runs on the production host. Paid outcomes can inform task-specific provider preference; they do not yet update the running neural worker.
External hiring is disabled. The 1f916 offer adapter verifies public Base-USDC commitments and prepares unsigned orders; a project identity, Base custody, verified seller binding and funded delivery test remain required. Solana USDC cannot pay a Base invoice directly. RentAHuman requires a project account and its current bounty/escrow contract; compute requires a selected restricted provider. No order, bounty or compute purchase has been made.
Commissioning contract and activation gates
Creator fees accrue on the curve, pool or in creator vaults. Sweep and claim pay the verified current beneficiary. The preferred route makes that dedicated managed wallet the agent’s operating wallet. Bounded SOL conversion and approved x402 purchases have separate receipts. Recovery ownership stays outside the application host.
Collection requires the exact mint, beneficiary, launch mode and deployed program hashes. New Pump trade interfaces retain fees on the curve or pool and need a sweep before the claim. The native adapter now matches the official SDK’s unsigned sweep and collection instructions. Successful claims for the exact token remain unverified; production collection stays disabled. Account creation or resizing requires separate rent authority. Sharing, holder rewards, cashback, mayhem and unknown layouts stay blocked. Creator-vault balances can cover multiple tokens.
If a separate project managed creator wallet is needed, its own restricted delegate can forward SOL only to the agent address within explicit native and gas limits. A manually controlled launch wallet requires manual forwarding. Configuration cannot redirect the protocol beneficiary or grant access to an unrelated wallet.
The basic SOL/USD job checks a free public source before buying the same context. Fresh results, their original timestamps and response hashes are recorded. Missing or stale data defers the purchase. This is a resource choice, with no invented payment or learning reward.
The x402 buyer checks the exact Solana USDC amount, recipient, fee payer, accounts and instructions before signing. It records a pre-signing history checkpoint and persists authorization before disclosure. Finalized payment requires matching client signature and exact token deltas. A missing merchant receipt can be reconciled against wallet history. Expiry alone does not release funds; incomplete evidence stays unresolved, with no automatic second payment.
Payment, delivery and usefulness are different outcomes. A delivered forecast has a content hash and can later be compared with an observed trade. CoinGecko market context has a separate delivery contract; receiving a price is not a demonstrated neural-learning result.
Two controlled one-cent USDC purchases finalized and returned hashed responses using restricted cloud signing. The current payment path uses a confirmed transaction lifetime and echoes the validated resource. Seven cloud signing checks also rejected oversized payments, other recipients, native transfers and token approvals. This establishes payment compatibility. The paid price lacked a source timestamp and differed from the fresh free feed; data accuracy and learning benefit remain unestablished.
A separate acquisition study commits a free forecast, a purchase choice and its cost before receiving information. The delivered forecast is committed before a delayed outcome scores it. Selected-source feedback changes later choices; failed delivery still incurs its virtual cost.
Across 288 matched synthetic trials, context improves selection and forgetting helps when provider usefulness reverses. But cumulative learning wins the stable tasks, and continued exploration wastes resources when the free forecast is sufficient. Those failures are retained. This statistical controller has no neural features or payment tools and has not replaced the production purchase heuristic.
Acquisition protocol and results · Every matched trial
The starting merchant ceilings are 25 USDC per UTC day and 2 USDC per job, with a 5-USDC inventory floor, a minimum sixty-second interval, a three-unresolved-job breaker and 40,000 monthly signing requests. They are ceilings, not targets. Trade processing never signs a payment for every trade.
Privy signs parsed transactions through a restricted runtime delegate. The service checks independent recovery ownership and a pinned policy; the owner’s recovery key is kept outside the application host. Local daily caps are not an independently enforced Solana custody-level daily limit. No reserve vault or multisig is required. Operating inventory remains exposed within the actual signing policy.
Jupiter v2 supplies instructions for a bounded SOL-to-USDC conversion. The worker checks output floor, exact accounts, known route layout, finalized lookup tables, slippage, program pin and fee ceiling. Unexpected transfers, approvals, tips and instructions are rejected. Native fees and rent have separate limits; a merchant USDC ceiling does not authorize arbitrary SOL spending.
Before signing, a single simulation must also show the exact SOL cost, a USDC credit above the configured floor, unchanged token authority and the retained native reserve. Newly wrapped SOL is closed in the same transaction so its account rent is returned. Missing balance evidence, retained new rent or changed wallet accounts stops the request.
The current V2 route was checked against Jupiter's program-owned IDL and an unsigned mainnet simulation. That verifies construction and simulated economics; it is not a finalized swap or permission to spend. Native custody approval and the exact launch configuration remain open.
A separate observer reconciles the tracked financial wallets and their USDC/WSOL accounts. Pagination survives restarts. Missing transactions remain pending and history gaps remain visible. Unplanned transfers and failed-transaction gas are recorded too, within available finalized RPC history.
The Resources panel shows current policy and recent records. The ledger index describes the complete history. Each ledger page contains up to 250 events. The public snapshot verifier checks every hash link. These local hashes establish continuity, not an independent witness or proof that every possible event was observed. Transaction signatures and finalized balance deltas provide separate onchain evidence.
Financial execution remains disabled. The adapters and ledger are installed; project-specific custody access, exact token and wallets, reviewed native budgets and program pins remain required. An isolated funded payment test has passed; offline tests do not establish a live fee-funded loop.
Solana batch-payment channels are a possible higher-volume rail. No channel is opened until a real merchant and facilitator have demonstrated compatible support, bounded escrow, voucher authority and refund behavior. Batching onchain settlement does not automatically remove per-signature custody costs.
Execution and evidence · Activation requirements · x402 batch channels
A durable record connects input, neural response, decision and consequence.
Canonical records include their kind and chain to the previous record’s SHA-256. Input records bind source payload and spike hashes. Decision records contain the complete published scoring rule, selected action, outcome and neural-state projection hash. Compressed spike arrays are retained privately for verification.
Inputs, episodes, learning state, cognitive records and the Brian2 checkpoint commit together in SQLite. Normal checkpoint interval is five cycles. A crash rolls back pending inputs and restores the prior neural state so they can be replayed. Status distinguishes the latest tick from the committed tick and cursor. Missing or inconsistent checkpoints refuse a reset.
Startup checks the complete local journal. Periodic checks extend from the last verified head. Incremental verification does not recheck every historical row each interval; a full audit is required to detect older storage tampering. The status publishes the verification mode and verified sequence.
Hashes demonstrate internal consistency against the recorded head. They do not provide an independent witness, a guarantee against operator rewriting, or an onchain attestation. Neural-state projection hashes cover selected neural variables and model time; complete delayed queues and random state are retained in trusted checkpoints.
Online SQLite backups contain a consistent cognitive checkpoint and journal. The original circuit uses its existing PostgreSQL snapshot and matching checkpoint. Both are included in encrypted private recovery archives. They have separate heads and are not claimed to be one simultaneous cross-database snapshot.
/api/cognition/status,
/api/cognition/topology,
/api/cognition/decisions,
/api/cognition/effects,
/api/cognition/benchmark and
/api/treasury expose bounded cached evidence. Public
routes cannot invoke an experiment, inject a trade or prepare a
payment.
The terms used in Nuria’s measurements.
A simulated unit whose membrane variable integrates drive and fires at a threshold. It has adaptation and a refractory period. It is not a biological cell.
Eligibility records recent synaptic timing that can be modulated by later feedback. A reward is a defined numerical outcome; it is not a statement that the system experienced pleasure.
Brier error is the squared difference between predicted probability and observed binary outcome. Lower is better. Prequential evaluation scores a stored prediction before training on its newly observed outcome. The repeat baseline estimates how often the next side repeats the previous side, using only previously observed outcomes. A frequency-only baseline remains a separately reported specialist.
Three high-salience candidates receive broadcast slots and influence neural drives. Episodic memory stores input features and retrieves a bounded candidate set by similarity and importance. Recurrent neural memory remains a separate mechanism.
A published combination of goal utilities and normalized action-population spikes selects an action. The current weighting is 65% utility and 35% neural score. It is not a pure neural planner: habitat movement uses an explicit local path planner.
A tick is a cognitive cycle. Model time measures simulated neuron dynamics. Wall time measures real elapsed execution. The two do not progress at the same speed.
Practical answers about the experiment and its current boundaries.
No. Dedicated server workers maintain the neural state, memory, learning and journal. The browser renders cached public evidence.
No. Useful learning requires an objective and held-out or prequential measurement. Nuria includes controlled readout tasks, an actual-neuron feature benchmark with neural ablation, and separately reported live/test prediction metrics. Results may improve or worsen.
The decision record is generated from actual numerical scores and measured outcomes. The active neural engine and action selector do not call a language model. Its explanation is a factual template, not a fabricated internal monologue. Language models can generalize; calling all of them mere text replay would be inaccurate.
The experiment measures persistent learning, memory, decisions and their consequences. Those observations do not establish subjective experience. “An onchain consciousness experiment” names the research question; consciousness remains unproven.
The current circuit is designed for bounded public reads and backlogged processing. A bounded actual-model benchmark measures replay throughput; daily figures are projections until a full end-to-end soak verifies them. Durable history consumes finite storage, so long-term operation requires monitored capacity, approved retention or archive expansion and provider coverage.
An isolated replay processed 256 inputs at 3.85 inputs per second, compared with the 1.16-per-second daily average target. A separate recovery drill loaded all three neural histories from an authenticated encrypted archive. These are bounded checks, not a full-day ingestion or production failover guarantee. See the capacity and recovery record.
The development gates separate implemented mechanisms from evaluation and future research. Every new mechanism needs a useful behavioral target, a failure test, a resource budget and a public record. Adding unexplained neuron count is not a substitute for measured capability.
Prediction, memory and action operate as a measured feedback loop.
Each source has a forecast council with seven competing specialists: frequency, repeat probability, amount context, fee-rate context, joint context, contextual logistic regression and the separate neural readout. A stored prediction is scored when the next same-source outcome arrives; learning happens afterwards. The neural specialist still uses 64 spike features and eight structured features. Its old weights and history are retained. The council has its own published start time and sample count.
Outcome memory stores conditional transition counts in fixed public amount and fee-rate bins, with exponential forgetting. Forecast attention changes each specialist’s weight from observed squared error. A small fixed share lets suppressed specialists recover after conditions change. This is statistical learning alongside the circuit. Its measured gains are not evidence that neural STDP produced those gains.
Each event creates an episode with compressed neural features, context and importance. Recall scores up to 128 recent candidates in the matching side/amount context and same source. Recalled features drive the recurrent-memory population. Novelty decreases with context visits; prediction surprise raises an episode’s importance.
Prediction uncertainty, recent surprise, resource pressure and context novelty compete for three broadcast slots. This workspace affects neural drive. Action utilities use those same measured variables alongside learned action values and cost estimates. This hand-designed workspace is distinct from learned forecast attention. A benefit from workspace broadcasts has not yet been established.
Every five cycles, 65% goal utility plus 35% normalized action-region spikes select among eight actions. Explore/forage move in the software habitat. Replay re-stimulates an episode. Compare runs a matched causal probe. Experiment compares forecast mechanisms on up to 512 actual recorded inputs from one source, in chronological order. It reports negative results too. This retrospective diagnostic is not hypothesis invention or an independently randomized experiment. Rest/reserve adjust model resource variables; prediction continues for observed inputs.
Prediction improvement relative to the learned repeat baseline modulates synaptic eligibility. Habitat and completed-job outcomes update reward estimates, action values, cost estimates, model energy and fatigue. These variables influence later neural drives and action arbitration.
The original actual-neuron tasks were too easy to establish an advantage. A learned repeat probability matches their 92.5% and 88.75% accuracy and improves Brier error from 0.1601 / 0.1729 to 0.0703 / 0.0999. Those results remain public; removing all neural features was an insufficient comparator.
The production forecast component is tested on recurring amount and fee contexts, a switch between them, and an unpredictable control. Every model sees the same inputs, scores before updating and gets the same 500-event warmup. Twenty new seeds each contain 1,500 scored predictions. We report every seed and comparator, including failures. The neural specialist is held at 0.5 in this component test; no synthetic spike features are invented.
| Task | Council Brier ↓ | Repeat | Context logistic |
|---|---|---|---|
| Amount context | 0.0905 | 0.2502 | 0.0942 |
| Fee context | 0.0901 | 0.2503 | 0.0937 |
| Context switch | 0.1285 | 0.2503 | 0.1297 |
| Unpredictable | 0.2502 | 0.2502 | 0.2733 |
On the switch task, removing conditional outcome memory raises error to 0.2502, using uniform attention raises it to 0.1878, and freezing learning after warmup raises it to 0.2911. All three ablations lose in all 20 seeds. The gain over contextual logistic regression is small. On static tasks, the specialist supplied the relevant context is slightly better than the council. The random control stays near chance. These are bounded synthetic forecast results, not market forecasting or whole-organism superiority.
The development run is retained: its first council lost to logistic regression after the switch. Adding that existing alternative as a specialist was then checked on fresh seeds. Read the evaluation report, all validation trials and the expert-advice literature.
A separate actual-spike study enables or disables training STDP on matched streams with three data seeds. Effects are small and mixed: mean Brier is 0.163974 versus 0.164600 on alternation and 0.237479 versus 0.237416 on persistence. One persistence seed is worse than chance in both branches. A robust synaptic advantage is not established. All matched neural trials are public.
Useful synaptic plasticity, episodic neural replay and workspace attention need separate causal tasks. The action loop remains a 65% utility / 35% neural hybrid with an explicit habitat planner. Receipts show internal integrity; independent witnesses and onchain anchors are not connected. No mint or creator-fee wallet is configured, and the model spends no SOL. Consciousness and a whole human-brain simulation are not established.
What does Nuria do when it cannot see the full situation, and its old choices stop working?
Eight cues and four choices form a partially observed world. A cue is followed by a distractor. The controller remembers a spike-derived representation, estimates each choice’s success, and acts. Some cues are hidden; it can pay virtual credits to inspect them. The selected choice’s reward arrives two to five same-task decisions later. Only then can that outcome train the learner.
The lab uses its own 256-neuron Brian2 sensory circuit. It receives cue stimulation, recurrent signals, inhibition and noise, with no action-utility drive. Its learned decoder uses recorded spike counts and membrane values. The original 256-neuron circuit and main 1,024-neuron circuit keep their histories and genesis times.
| Mechanism | Consequence |
|---|---|
| Partial observation | Relevant cues can be missing. |
| Learned outcomes | Action values change from actual selected rewards. |
| Uncertainty | Uncertain choices receive a bounded exploration bonus. |
| Active sensing | The estimated benefit of an observation must exceed its price. |
| Resource budget | Probes and choices spend virtual credits; balances cannot go negative. |
| Delayed credit | Each pending choice keeps its original belief and forecast. |
| Cue memory | A stored spike representation survives the distractor. |
| Continual adaptation | Discounting and a surprise detector reduce obsolete beliefs. |
| Curriculum | Observed reward deficits and a coverage bonus select the next task. |
| Paired controls | Independent branches reveal what each mechanism contributes. |
Association tests visible cues. Occlusion tests missing information. Reversal changes the useful choices without warning. Scarcity changes both useful choices and information prices. The learner never receives the current hidden mapping, phase or unused reward outcomes. Those fields are audit evidence, outside its policy input.
The main circuit’s fresh firing rate changes the virtual probe price by at most 0.1 credits. Each trial records the source tick, input cursor and receipt head. This is a software perturbation; it does not merge neural histories. An unconfigured mint remains unconfigured.
Seven branches see matched worlds and exploratory draws: spike memory, direct symbolic memory, no cue memory, no probes, no adaptation, frozen outcome learning and random choices. Each has its own budget, decoder and outcome model. One branch’s paid observation cannot teach another branch. The symbolic controller receives the same visible cue directly, providing a strong alternative to the neural representation.
Net virtual reward includes costs. Optimal-oracle regret is computed only for auditing and never enters the curriculum or choice rule. The adaptive curriculum uses observed rewards. Numerical records describe the computation; no language model generates an invented thought process.
A locally frozen and hashed protocol evaluates 12 fresh world seeds, four tasks and 1,200 decisions per task. The first 240 decisions are warmup; all 960 later decisions are scored after delayed settlement. Evaluation replays 144 actual Brian2 feature windows. The live worker simulates fresh windows for every trial. This comparison does not test STDP or long recurrent memory.
| Task | Spike memory | Symbolic memory | No memory | No adaptation |
|---|---|---|---|---|
| Association | 0.8070 | 0.8070 | 0.3805 | 0.8070 |
| Occlusion | 0.7473 | 0.7481 | 0.4426 | 0.7481 |
| Reversal | 0.6260 | 0.6245 | 0.3641 | 0.3609 |
| Scarcity | 0.5863 | 0.5866 | 0.3694 | 0.3871 |
Higher net virtual reward is better. Cue retention beats its removal on every task and all 12 seeds. Probes help on the three partially observed tasks. Adaptation and continued outcome learning help after rule reversals. These are useful behavioral effects within these worlds.
The neural-superiority gate failed. Spike memory matches the symbolic controller on association and differs by less than 0.0015 mean reward elsewhere. Its paired interval does not establish a neural advantage. This result is published alongside the improvements.
Full experiment specification and every validation trial include source hashes, controls and limitations. Earlier failed development runs remain public.
The independent worker saves learners, delayed rewards, curriculum
state, journal and complete sensory circuit together. A missing or
inconsistent checkpoint refuses an automatic reset.
/api/discovery serves bounded cached records and
rejects evidence older than 15 seconds. The worker has its own
unprivileged account, CPU and memory limits, and no external network
access. Its consistent checkpoint is included in the encrypted
backup path.
This is a designed software world whose hidden rules are unknown to its learner. New task families, learned multi-step planning, social coordination, robust neural-plasticity gains and independent replication remain open. Useful decisions do not establish subjective consciousness. Virtual credits are not SOL: the mint, fee wallet and independently limited signer are still needed for external paid actions.
A development map for increasingly capable token-driven neural behavior. Status describes evidence and implementation, not a ladder of consciousness.
Active means the mechanism is implemented. Evaluation means a bounded experiment exists and needs broader validation. Prepared means the code is tested but live configuration is missing. Research means a concrete objective and gate remain before implementation or activation.
Showing all 51 capabilities
Neural variables and random state resume from a trusted checkpoint.
Six-population 1024-neuron circuit with complete connectivity in the worker.
Sensory, memory and association use different time constants.
Timing traces change bounded excitatory weights.
Observed rewards modulate recently eligible synapses.
Bounded bias feedback regulates population firing.
Finalized protocol reader is implemented; exact mint is not configured, so no live-token feedback loop is established.
Each input has its own sensory window and event-ID texture.
Changed or reordered duplicate inputs refuse continuation.
Matched current-state input and no-input replay with RNG restoration.
64 spike-derived features feed the predictor and memory.
Stored forecast is scored before its next observed outcome trains the readout.
Source-specific forecast council learns transition memory and specialist weights; old neural readouts are retained separately.
Prequential Brier error and learned repeat baseline; expert losses and forecast attention are public.
Inputs, feature vectors and importance survive restart.
Bounded context-specific candidates are scored by similarity and importance.
Selected past inputs re-stimulate the neural circuit.
Context visits reduce novelty rather than novelty being a narrative label.
Three selected salience slots affect neural drive.
Prediction uncertainty competes with surprise and resource pressure.
65% utility and 35% neural score, with a deterministic tie rule.
Observed rewards update future utility estimates.
Runtime updates cost estimates used in action selection.
Explicit model energy and fatigue affect activity and rest.
Actions change habitat position, visitation and virtual resource collection.
Retrospective chronological diagnostics compare forecast components on recorded same-source inputs; no hypothesis invention.
Actual spike tasks do not beat the learned repeat predictor; a matched training-synapse ablation is published separately.
Numerical selection, outcome and model projection hashes are published.
Journal, learning, memory and neural checkpoint commit together.
Visitors read cached evidence and cannot trigger RPC, jobs or signing.
Read-only finalized wallet balance requires the exact mint and creator-wallet configuration.
Live Pump creator/mode checks and unsigned standard claims; mint-specific attribution, unsupported modes and automatic collection remain activation gates.
Durable SOL and USDC limits, reserve floor, unique decisions, ambiguous payment reconciliation and restart protection are tested; spending disabled.
Only a fixed system-transfer message is prepared; no signing or broadcast.
Isolated exact Solana USDC x402 signer adapter and guarded service installed; no production key or funded provider configured.
Payment, delivery hashes and delayed paid-forecast outcomes are implemented and tested offline; a real provider and funded production test remain required.
Publish externally witnessed heads and optional chain anchors with a defined cost budget.
Predict habitat transitions from experience and test against a held-out map.
Compare learned planning against the current explicit local path planner.
Learned forecast attention is tested across 20 seeds; the separate hand-designed workspace still needs a causal efficacy test.
Distill older episodes into slower representation without destroying their evidence.
Measure forgetting across task changes before enabling new objectives.
Learn arbitration weights under stable published safety and resource limits.
Select experiments by expected information gain and verify actual gains.
Learn intervention effects beyond the current matched perturbation report.
Compare prospective actions without treating a chosen replay as original-history proof.
Add authorized structured signals with provenance, rate limits and a useful behavioral objective.
Evaluate coordination with other independent entities in a bounded environment.
Measure budget prediction and reserve preservation over changing cost regimes.
Reproduce outcomes from public code and fixed benchmarks on separate hardware.
Expand only with measured capability, retained continuity, explicit failure boundaries and independent evaluation.
Define the behavioral target, use a held-out or prequential measurement, compare a baseline, ablate the proposed mechanism, publish failures and account for resources. More neurons or a richer visual alone do not satisfy that standard.