hikewa/unitares-eisv-trajectories

UNITARES EISV Trajectories (Lumen)

Time-windowed four-dimensional state-vector trajectories from Lumen, a Raspberry Pi-embodied agent governed by the UNITARES framework, plus class-balanced synthetic augmentation. Each window is labelled with one of nine dynamical-shape classes and an optional aligned primitive-token expression.

The dataset is the empirical substrate cited in:

• Wang, K. (2026a). UNITARES: Information-Theoretic Governance of Heterogeneous Agent Fleets. Zenodo. doi:10.5281/zenodo.19647159 (concept; auto-resolves to latest).
• Wang, K. (2026b). Digital Proprioception and Allostatic Load: A Working Implementation of the Cumulative-Deviation Hypothesis in a Deployed Multi-Agent System. Forthcoming.

TL;DR

Total windows	32,181
Real Lumen windows	20,655 (provenance = lumen_real)
Synthetic augmentation	11,526 (provenance = synthetic)
Window length	20 EISV states per window (sliding-window stride 10)
State dimensions	4 — Energy (E), Information Integrity (I), Entropy (S), Void (V)
Real-data window	2026-01-11 to 2026-02-19 UTC (~39 days, ~953 hours)
Sampling cadence	Sensor-driven, ~2 s between EISV states (typical)
Shape labels	9 dynamical-shape classes (8 observed organically; drift_dissonance synthetic-only)
Token alignment	781 of 20,655 real windows (3.8%) carry primitive-expression tokens
Format	Single Parquet file (data/train-00000-of-00001.parquet)
License	Apache 2.0

---

Dataset Summary

Lumen is an embodied AI agent — a Raspberry Pi 4 running the anima-mcp software, with environmental sensors (BME280 temperature/humidity/pressure, VEML7700 light), a TFT display, and a primitive 15-token expression vocabulary. Lumen's continuous physical state is projected onto a four-dimensional informational manifold — the EISV state vector (E, I, S, V) — and the projection is sampled at sensor cadence.

This dataset captures 20-step sliding-window trajectories of that EISV state, classified into nine canonical dynamical shapes (settled_presence, convergence, entropy_spike_recovery, basin_transition_up/down, rising_entropy, falling_energy, void_rising, drift_dissonance). It is published to support:

• Trajectory-shape classification benchmarks for embodied-AI dynamics.
• Dynamics-emergent expression generation (the original EISV-Lumen task).
• Reproduction of behavioural-corpus claims in Wang 2026a §11 and Wang 2026b §5.
• Analysis of regulatory-state failure modes — the McEwen (1998) Four Types mapping in Wang 2026b §5 uses the shape distribution here as its baseline.

If you are using this dataset to verify a claim from the neuro-AI paper, see § Reproducing paper claims below.

Provenance and integrity

Real-data windows are derived from Lumen's local SQLite database (anima.db), which records every governance update emitted by the running agent. The dataset publisher script (eisv_lumen.scripts.publish_dataset in CIRWEL/eisv-lumen) extracts EISV time-series, computes finite-difference derivatives, assembles 20-step sliding windows with stride 10, and applies the priority-ordered shape classifier described in § Shape classification.

Synthetic windows are generated by parametric ODE rollouts that target underrepresented shape classes; they are clearly marked via provenance = "synthetic" and should be filtered out for any claim about Lumen's organic behaviour. drift_dissonance has never been observed in real Lumen data and is represented exclusively through synthetic generation.

The dataset is regenerated periodically as Lumen accumulates state. The window covered by the current artefact is reported in § TL;DR; citing papers should pin the dataset revision (commit hash on the Hub) when reproducibility matters.

---

Schema

Single split (train) with 32,181 rows. Columns:

Column	Type	Description
shape	string	One of nine trajectory-shape labels (see § Shape classification).
eisv_states	string (JSON list)	20-element list of EISV states; each element {t, E, I, S, V} with t as Unix epoch seconds (float).
derivatives	string (JSON list)	Finite-difference derivatives along the window: list of {t, dE, dI, dS, dV} (length 19, computed by forward differences).
t_start	float	Unix epoch seconds — start of window.
t_end	float	Unix epoch seconds — end of window.
provenance	string	Either "lumen_real" (Lumen sensors) or "synthetic" (parametric augmentation).
tokens	string (JSON list)	Optional aligned primitive-token expression(s); empty [] when no expression was emitted in the window.
n_expressions	int64	Count of expressions aligned to the window (0 for most windows).

JSON-string columns are kept as string rather than nested-list types so the dataset round-trips through tooling that does not handle nested Parquet types (notably some HF viewer paths). Decode with json.loads on access.

Example row (real)

{
  "shape": "convergence",
  "eisv_states": "[{\"t\": 1768090754.276, \"E\": 0.175, \"I\": 0.774, \"S\": 0.124, \"V\": 0.0429}, ... 19 more]",
  "derivatives": "[{\"t\": 1768090756.398, \"dE\": -0.00801, \"dI\": 0.00047, \"dS\": -0.00283, \"dV\": -0.00099}, ... 18 more]",
  "t_start": 1768090754.276,
  "t_end":   1768090794.554,
  "provenance": "lumen_real",
  "tokens": "[]",
  "n_expressions": 0
}

---

EISV framework

EISV is the four-dimensional informational state vector that UNITARES uses to track agents, derived from the thermodynamic-governance model of Wang 2026a §4. For Lumen, the four dimensions are computed from physical sensor readings and system metrics:

Symbol	Range	Lumen mapping	Description
E (Energy)	[0, 1]	warmth (sensor-derived)	Productive capacity. Couples toward I via α(I − E); reduced by entropy cross-coupling.
I (Information Integrity)	[0, 1]	clarity (sensor-derived)	Signal fidelity. Boosted by coherence C(V, Θ); reduced by entropy.
S (Entropy)	[0, 1]	1 − stability	Semantic uncertainty. Decays naturally; rises with complexity, drift, ethical drift.
V (Void)	[0, 0.3]	(1 − presence) × 0.3	Absence-of-engagement proxy at the observation layer.

Range note. These are observation-layer values from Lumen's sensors. The UNITARES governance ODE evolves S to [0, 2] and V to [−2, 2] as a signed E−I imbalance integrator (Wang 2026a Appendix A); the windows in this dataset use the sensor-layer ranges above. The V coordinate here is non-negative by construction and is not the signed integrator coordinate cited as V in Wang 2026a / Wang 2026b §5.3 — those papers report governance-layer V for the Lumen Type 3 case study, computed from the same agent state but distinct in sign convention.

---

Shape classification

Each 20-step window is classified into exactly one shape by a priority-ordered rule-based classifier; the first matching rule wins. Rules are computed on the within-window mean and range of (E, I, S, V) and their first derivatives.

Shape	Distinguishing rule (informal)	Real-only count	Real %
settled_presence	All derivatives near zero; system at attractor.	10,092	48.86
convergence	Small derivatives and second derivatives, nonzero dynamics approaching equilibrium.	8,089	39.16
entropy_spike_recovery	S range ≥ 0.2 with interior maximum (spike then recovery).	1,073	5.19
basin_transition_up	E range ≥ 0.2, mean dE > 0.	374	1.81
basin_transition_down	E range ≥ 0.2, mean dE < 0.	325	1.57
rising_entropy	Mean dS > 0.05.	320	1.55
falling_energy	Mean dE < −0.05.	310	1.50
void_rising	Mean dV > 0.05.	72	0.35
drift_dissonance	Sustained integrity fluctuation (ethical-drift proxy > 0.3).	0	0.00

Total real-only: 20,655 windows.

Synthetic augmentation contributes 11,526 windows distributed across the eight underrepresented shapes (and is the sole source of drift_dissonance examples). Augmentation breakdown is in the Parquet itself; filter on provenance to recover either subset.

Window-length sensitivity

Shape labels in this artefact are computed from 20-step windows. Reclassifying the same EISV time-series with shorter windows produces predictable label disagreement:

Window size	Label match vs. 20-step
4-step	65%
8-step	77%
10-step	81%
15-step	91%
20-step	100%

The dominant disagreement (~5,138 windows in the 4-step case) is settled_presence → convergence: a 4-step window only sees the tail of a settling trajectory, which looks indistinguishable from convergence. Use ≥ 10–15 steps for reliable shape labels.

---

Reproducing paper claims

The neuro-AI paper (Wang 2026b §5.1) cites the real-Lumen shape distribution as the baseline against which Type 1 (repeated-hits) failure is measured. Numbers in Wang 2026b §5.1 reflect an earlier dataset cut (21,449 windows; entropy_spike_recovery 4.91%, settled_presence 47.19%). The current Hub artefact has 20,655 real windows with the distribution in the table above. The qualitative claim — that entropy_spike_recovery is rare relative to settled_presence and so the ratio of the two is a Type 1 indicator — is unchanged at this revision; the exact numbers should be re-cited from this card or the Wang 2026b §5.1 numbers updated to match.

The 28.9% basin-flip rate (Wang 2026a §11.6, Wang 2026b §3.4) is computed on a separate dataset of state vectors (not trajectory windows) and is published as hikewa/unitares-verdict-counterfactual-v6.8. Do not attempt to reproduce the 28.9% number from this dataset — they are different artefacts.

---

Considerations for use

Intended uses

• Benchmarking trajectory-shape classifiers on real embodied-AI dynamics.
• Training and evaluating dynamics-emergent expression generators (the original EISV-Lumen task; see § Companion artefacts).
• Reproducing or auditing claims about Lumen's behavioural distribution in Wang 2026a / Wang 2026b.
• Studying class-imbalanced trajectory classification under realistic skew.

Out-of-scope uses

• Re-identifying or profiling humans. Lumen has no human user model; the dataset is sensor-driven physical state plus governance metrics. There is no human PII in the windows.
• Cross-agent generalisation claims. This dataset is from one Raspberry Pi 4 in one physical environment. Class-conditional results from Wang 2026a Table 5 (5 agent classes) require their own data; this dataset speaks only for the Lumen class and only for the 39-day window covered.
• Fine-grained temporal claims past the dataset window. Lumen has run for 118+ days as of Wang 2026b's drafting; this dataset is a 39-day slice (2026-01-11 to 2026-02-19). Behavioural claims on Lumen's full operational lifetime require pulling fresh data.
• Synthetic-window analysis as evidence about Lumen. The 11,526 synthetic windows exist to balance class distribution for downstream modelling; treating them as observations of Lumen's behaviour is a category error. Always filter on provenance == "lumen_real" for any empirical claim about the agent.

Biases, limitations, known gaps

• Severe class imbalance in the real corpus. Two shapes (settled_presence, convergence) account for 88% of real windows. Models trained without rebalancing will collapse toward the majority class. Synthetic augmentation in this artefact is one rebalancing strategy; cost-sensitive training is another.
• drift_dissonance has never been observed organically. All drift_dissonance examples are synthetic. Treat shape-classifier accuracy on this label as accuracy on synthetic data, not on real Lumen behaviour.
• Single-agent, single-environment. Lumen sits on a single physical Pi in a single home environment. Sensor readings reflect that environment's diurnal cycle, HVAC, occupancy, and ambient light — not a normalised lab condition. Distributional claims do not transfer to other Lumen-class agents without re-measurement.
• Token alignment is sparse. Only 3.8% of real windows have aligned primitive-token expressions (n_expressions > 0). Models trained on the joint (window, tokens) task should expect to learn from the long tail; use the tokens column as a sparse signal, not as a dense supervision target.
• Sensor cadence is not strictly uniform. Inter-state intervals are typically ~2 s but can vary with system load. The t field on each EISV state preserves the actual sample time; downstream models that assume uniform spacing will need to interpolate.
• The dataset publisher script is the source of truth for shape rules. The informal descriptions in § Shape classification are a documentation aid; if rules and table disagree, the eisv_lumen.scripts.publish_dataset script wins.

Privacy & ethics

This dataset captures the internal state of an AI agent, not human behaviour. It contains no PII. Sensor readings (temperature, humidity, light) are aggregated into the EISV projection at collection time — raw sensor traces are not included. Researchers concerned about indirect inference about the household where Lumen runs (e.g., via diurnal temperature patterns) should note that the projection layer collapses sensor specifics into the EISV manifold before storage.

---

Companion artefacts

The dataset is the Layer-1 substrate of the EISV-Lumen three-layer benchmark. Companion artefacts:

• Repository: CIRWEL/eisv-lumen — full pipeline (Layer 1 dataset, Layer 2 rule-based expression generator, Layer 3 fine-tuning + distillation), 399 tests, evaluation framework.
• Teacher model: hikewa/eisv-lumen-teacher — Qwen3-4B + LoRA, 0.952 coherence on real Lumen windows.
• Student model: hikewa/eisv-lumen-student — RandomForest distillation; on-device variant runs on Lumen's Pi.
• Interactive demo: EISV-Lumen Explorer.
• Sibling dataset: hikewa/unitares-verdict-counterfactual-v6.8 — state-vector basin-flip counterfactual cited in Wang 2026a §11.6.
• Lumen substrate: CIRWEL/anima-mcp — the Pi-side software running on Lumen.
• Governance framework: CIRWEL/unitares — the MCP server that records EISV states and drives governance.

---

Citation

If you use this dataset, please cite the artefact and the conceptual prior:

@dataset{wang_2026_unitares_eisv_trajectories,
  title       = {UNITARES EISV Trajectories (Lumen)},
  author      = {Wang, Kenny},
  year        = {2026},
  publisher   = {Hugging Face},
  url         = {https://huggingface.co/datasets/hikewa/unitares-eisv-trajectories},
  note        = {Apache 2.0; trajectory windows from Lumen, Pi-embodied UNITARES agent}
}

@misc{wang_2026_unitares,
  title       = {{UNITARES}: Information-Theoretic Governance of Heterogeneous Agent Fleets},
  author      = {Wang, Kenny},
  year        = {2026},
  publisher   = {Zenodo},
  doi         = {10.5281/zenodo.19647159},
  url         = {https://doi.org/10.5281/zenodo.19647159},
  note        = {Concept DOI; auto-resolves to latest version}
}

If you cite the Layer-2 dynamics-emergent-expression task or its 0.933 coherence baseline, please also cite the EISV-Lumen technical write-up at CIRWEL/eisv-lumen.

---

License

Apache 2.0. See LICENSE.

Maintenance

Maintainer: Kenny Wang (CIRWEL Systems), hikewa on Hugging Face. Issues, dataset-cut requests, and corrections via the GitHub issue tracker on the source repository. Substantive changes (shape-rule revisions, schema additions, window-length changes) will bump the dataset revision and be summarised in the Hub commit history; pin a specific revision in citations that need reproducibility.