NeuraxonLive2M

Two months of continuous logs from Neuraxon GoL Live (nxon.online), a persistent multiplayer artificial-life server in which every agent (an NxEr) is driven by its own spiking, trinary-state neural network organised into a six-sphere CHC brain. Agents forage, move, mate, and die in a shared 360×360 world that runs 24/7 at ~20 ticks per second. The dataset contains two uninterrupted runs — 44.01 days and 10.07 days — totalling ~17 million simulation ticks, ~126,000 agent lifetimes, ~53,000 sampled neural architectures, and ~77,000 population-level metric samples, plus the final world snapshot from each run. This card describes the data only. It makes no claims about what the data shows.

Contents

NeuraxonLive2M/
├── 44DaysRun1.53/
│   ├── history/
│   │   ├── timeseries.jsonl     121 MB   63,176 records
│   │   ├── obituaries.jsonl      26 MB   86,455 records
│   │   ├── nas_trials.jsonl      21 MB   46,957 records
│   │   └── lineage.jsonl        7.0 MB   39,747 records
│   └── snapshot.json            6.0 MB   final world state
└── 10DaysRun1.54/
    ├── history/
    │   ├── timeseries.jsonl      31 MB   14,464 records
    │   ├── obituaries.jsonl      18 MB   39,642 records
    │   ├── lineage.jsonl        5.9 MB   33,750 records
    │   └── nas_trials.jsonl     5.2 MB    6,344 records
    └── snapshot.json            6.5 MB   final world state

Total ~247 MB. Every .jsonl file is newline-delimited JSON, one record per line, appended in chronological order. Every record carries a t_unix wall-clock timestamp; most also carry the simulation tick. ## The two runs | | 44DaysRun1.53 | 10DaysRun1.54 | |---|---|---| | duration | 44.01 days (1,056.2 h) | 10.07 days (241.7 h) | | ticks | 1 → 13,881,816 | 1 → 3,111,151 | | population at start / end | 32 → 513 | 32 → 593 | | agent lifetimes recorded | 86,455 | 39,642 | | architectures sampled | 46,957 | 6,344 | | metric samples | 63,176 | 14,464 | | snapshot tick / agents | 13,881,900 / 712 | 3,111,293 / 790 | Both runs begin from a cold boot with 32 founder agents and grow under the simulation's own dynamics. The 10-day run carries additional per-agent fields (see Schema differences below). > Note on folder names. The directories are named after the project release > the data was associated with when archived. The server version string written > by the running binary was GoL Server V 1.072 for the 44-day run and > GoL Server V 1.074 for the 10-day run. Users wanting the authoritative > version should treat those strings as canonical; the provenance.jsonl > stream that records them is not included in this dataset.

Streams

timeseries.jsonl — population-level samples (~1/min)

One record per sampling sweep, describing the whole living population. | field | description | |---|---| | tick, t_unix, uptime_s | simulation tick, wall clock, seconds since boot | | alive, managed, tracked | population counts | | rates | births_mating_per_min, deaths_per_min, food_eaten_per_min, matings_per_min, spawns_per_min | | fitness, g, offspring, lifespan_ticks | distribution blocks, each with min, p10, p50, mean, p90, max, sd | | traits | population means of inherited traits: thr_exc, thr_inh, learning_rate, refr, ahp | | trinary | M1_exc, M1_inh, M1_neutral, excitation_dominant, neurons_sampled, L1_to_paper | | g_structure | mean_r, pos_manifold, lambda_ratio, pc1, n — cross-agent ability-correlation structure | | m12 | the M-metric block, below | | perf | tick_ms, tps_est, phase_a_sense_ms, phase_b_brains_ms, phase_c_act_ms, brains_due_per_tick | | pool_mode | brain-execution mode string | The m12 block. Neural measurements aggregated across all sampled brains, named after the study's M-claims: | key group | fields | |---|---| | M1 trinary balance | M1_E, M1_I, M1_N, M1_deviation | | M2 inter-sphere gating | M2_mean_gate, M2_gate_xlink_std, M2_n_links | | M5 criticality | M5_branching, M5_branching_std, M5_subcritical_frac, M5_supercritical_frac | | M6 timescales | M6_acw_mean, M6_acw_heterogeneity | | M7 self-sustained activity | M7_zero_input_ratio, M7_zero_motor_amp, M7_driven_motor_amp, M7_n_zero, M7_n_driven | | M8 sphere specialisation | M8_sensory_act, M8_assoc_act, M8_motor_act, M8_sensory_vs_assoc | | M10 robustness / heredity | M10_dead_neuron_frac, M10_lesion_retention, M10_heritability_r, M10_heritability_pairs | | band tagging | _in_band (per-metric 0/1), _in_band_count, _band_total, _n_brains_sampled | Each metric has an associated target band; _in_band records whether the sampled value fell inside it. M3, M4 and M9 are not recorded online and are absent. ### obituaries.jsonl — one record per agent death id, name, is_male, managed, born_tick, death_tick, tick, lifespan_ticks, cause ("idle" or "starved"), lifetime outcome counters food_found, food_taken, explored, mates, n_offspring, fitness, g, and the agent's inherited thresholds thr_exc, thr_inh. ### lineage.jsonl — one record per birth id, name, is_male, managed, born_tick, tick, parents (a two-element array of parent ids), and the newborn's thr_exc, thr_inh, refr. Joining this to obituaries.jsonl on id yields parent→child outcome pairs. ### nas_trials.jsonl — one record per sampled architecture Agents spawned by the system (rather than born from mating) each carry a freshly sampled neural architecture; the record is written when that agent dies. trial, arch, born_tick, death_tick, lifespan, cause, and outcomes food_found, food_taken, explored, mates, n_offspring, fitness, g. The arch object holds the seven sampled parameters: firing_threshold_excitatory, firing_threshold_inhibitory, post_spike_mp_reset, intrinsic_timescale_default, learning_rate, connection_probability, sensorimotor_coupling. This stream is an architecture → outcome table: one row per architecture, with the lifetime that architecture achieved in the live world. ### snapshot.json — final world state A single JSON object: - tick, next_nxer_id, next_food_id, nxer_names, names_state - foods — 2,600 food items with positions - all_time — best value reached for each of food_found, food_taken, explored, time_lived, mates_performed, fitness, g - lifetime — cumulative counters: total_ticks, total_births_mating, total_deaths, total_matings, total_spawns, total_food_eaten, total_food_spawned, peak_alive, peak_managed, total_managed_registrations, uptime_seconds, started_at_unix - nxers — every agent alive at the end (712 and 790 respectively), each with id, name, pos, alive, food, is_male, is_managed, can_land, can_sea, parents, offspring_ids, a stats block (food_found, food_taken, explored, time_lived, mates_performed, fitness, g, branching, energy_efficiency), and a params object of 275 neural parameters — the agent's complete architecture specification. Synaptic weights are not included. params describes each agent's architecture, not its learned state; the serialised brains were excluded to keep the snapshots small. Agents can be reconstructed structurally, not restored to their exact learned condition.

Schema differences between the runs

The 10-day run was produced by a later build and carries extra fields. Anything listed here is absent from the 44-day run. timeseries.m12 gains 11 fields: - Mc_mean, Mc_median, Mc_best, Mc_frac_ge_0_75, Mc_frac_ge_0_90, Mc_n — the distribution of per-agent M-band compliance scores across the living population. - W_mean, W_mean_abs, W_std, W_pos_frac, W_n_sampled — sampled synaptic-weight statistics. obituaries.jsonl gains 7 fields: fitness_hi (an alternative, log-scaled fitness formulation logged alongside the original fitness), food_by_age (a map of agent age in ticks → cumulative food_found at that age, sampled at 1,250 / 2,500 / 5,000 / 10,000 / 20,000 / 40,000 / 80,000 / 160,000), and the per-agent compliance fields m_score, m_in_band, m_n_checked, m_deviation, m_samples. nas_trials.jsonl gains 9 fields: the same fitness_hi, food_by_age, m_score, m_score_last, m_in_band, m_n_checked, m_deviation, m_samples, plus m_last — that brain's own raw M-metric values at its final sampling. snapshot.json gains nas_trial_seq at top level, and each agent gains born_tick and food_by_age. lineage.jsonl is identical across both runs.

Loading

import json
 
def read(path):
    with open(path) as f:
        for line in f:
            line = line.strip()
            if line:
                yield json.loads(line)
 
# architecture -> outcome table
for r in read("44DaysRun1.53/history/nas_trials.jsonl"):
    arch, fitness, lifespan = r["arch"], r["fitness"], r["lifespan"]
 
# parent -> child outcome pairs
outcomes = {r["id"]: r for r in read("10DaysRun1.54/history/obituaries.jsonl")}
for b in read("10DaysRun1.54/history/lineage.jsonl"):
    p = b.get("parents") or []
    if len(p) == 2 and b["id"] in outcomes:
        child = outcomes[b["id"]]
        parents = [outcomes[i] for i in p if i in outcomes]

With datasets: python from datasets import load_dataset ds = load_dataset("NeuraxonLive2M", split="run_44day_nas_trials") timeseries.jsonl has deeply nested records; loaders that infer a flat schema may need pandas.json_normalize or an explicit feature spec. The 121 MB 44-day timeseries file is best streamed line by line rather than loaded whole.

License

Released under CC-BY-4.0-SA

Citation

@dataset{neuraxonlive2m
  title  = {NeuraxonLive2M: Two Months of Continuous Artificial-Life Logs
            from Neuraxon GoL Live},
  year={2026},
  publisher={Hugging Face},
  version={1.0.0},
  url    = {https://nxon.online}
}

Authors & Curators

• David Vivancos / Artificiology Research (https://artificiology.com) - Qubic Science
• Dr. Jose Sanchez / UNIR - Qubic Science Contact: For questions or issues, please open a GitHub issue at https://github.com/DavidVivancos/Neuraxon.