license	cc-by-4.0
language	en
size_categories	10K<n<100K
task_categories	other
pretty_name	UNITARES Verdict Counterfactual — Paper v6.8 Repro Kit
tags	agent-governance ai-safety time-series reproducibility

UNITARES Verdict Counterfactual — Paper v6.8 Reproducibility Kit

Reproducibility artifacts for §11.6 of UNITARES: Information-Theoretic Governance of Heterogeneous Agent Fleets (Wang, 2026).

• Paper: CIRWEL/unitares-paper-v6, tag paper-v6.8.1
• Paper DOI (concept): 10.5281/zenodo.19647159
• Data DOI (this release, v6.8.1-repro): 10.5281/zenodo.19705151
• GitHub: CIRWEL/unitares-repro-v6
• Source: UNITARES production governance database, 30-day rolling window of core.agent_state
• License: CC-BY-4.0 (same as paper)

What this is

The paper's §11.6 counterfactual asks: how many basin assignments would flip if we replaced the legacy fleet-wide tanh coherence with a class-conditional grounded coherence? On a 13,310-row production slice it reports a 28.9% flip rate, with directional bias into the low basin — empirical support for the homogenization-failure thesis.

This dataset publishes two snapshots of that counterfactual:

File	Window	Rows	Flip rate
verdict_counterfactual_v6_submission.csv	30 days ending 2026-04-18 21:00 MDT	13,292	28.8%
verdict_counterfactual_2026-04-23.csv	30 days ending 2026-04-23	16,879	44.3%

The submission snapshot is within 18 rows / 0.1pp of the figures in the published paper (exact rerun timing varies by seconds).

The 2026-04-23 snapshot shows a ~15pp increase in overall flip rate over 4 days of window shift, with the Phase 2 calibration constants held frozen at v6.8 submission values. We do not claim this is a steady-state drift signal — a 4-day gap with ~87% window overlap is one measurement, not a trend, and the magnitude is within the kind of shift class-conditional calibration drift could absorb in a single re-calibration pass.

Volume checks confirm the shift is not dominated by observer effects (daily row counts over the interval are ~800–1,500, no spikes). The per-class pattern is informative (Sentinel +25.4pp, Vigil +19.6pp, Lumen +15.8pp, default +15.3pp, Watcher −11.3pp). We publish both snapshots so others can reproduce §11.6 exactly and so the v7 empirical agenda has a concrete before/after pair to iterate from. See paper §11.6 and unitares-v7-outline.tex for the open questions this speaks to.

Columns

Each row is one agent-state observation, already pseudonymized to a class label — no agent UUIDs, session IDs, prompts, or KG content are in this export.

Column	Type	Meaning
class	str	Agent class: Lumen, Sentinel, Vigil, Watcher, or default (tag-derived for non-resident agents)
E	float ∈ [0, 1]	Energy (productive capacity)
I	float ∈ [0, 1]	Information integrity
S	float ∈ [0, 1]	Entropy (lower is better)
V	float ∈ [-1, 1]	Void / accumulated E–I imbalance
risk	float ∈ [0, 1]	Risk score at time of observation
c_legacy	float ∈ [0, 1]	Legacy fleet-wide tanh coherence
c_grounded	float ∈ [0, 1]	Class-conditional grounded coherence
basin_legacy	str	Basin assignment under legacy coherence: high, boundary, low
basin_grounded	str	Basin assignment under grounded coherence
flipped	int	1 if basin_legacy != basin_grounded, else 0

Basin thresholds and class-conditional calibration constants are in scripts/verdict_counterfactual.py.

How to reproduce

pip install pandas matplotlib
python analysis.py  # prints the Table 5 equivalent + drift comparison

To regenerate either CSV from a live UNITARES instance (requires access to the governance DB):

python scripts/verdict_counterfactual.py --window-days 30 \
  --end-date "2026-04-18 21:00:00-06:00" \
  --csv --output verdict_counterfactual_v6_submission.csv

Limitations

• Fleet composition is UNITARES-specific (five resident agents + ephemeral coding assistants + embodied Lumen). The absolute flip rates are not directly transferable to other fleets; the method is.
• Paper §12.5 (write-path hygiene) and §11.7 (identity system maturity) describe known caveats in the underlying trajectory data. Read both sections before citing these numbers in derivative work.
• Class-conditional calibration constants in the script are frozen at v6.8 submission. Regenerating the 2026-04-23 snapshot against the current production calibration (not frozen) would yield a different number — we deliberately use the frozen constants to isolate fleet drift from calibration drift.

Citation

Cite the paper for the method, cite the dataset for the data.

@article{wang2026unitares,
  author  = {Wang, Kenny},
  title   = {UNITARES: Information-Theoretic Governance of Heterogeneous Agent Fleets},
  year    = {2026},
  version = {v6.8.1},
  doi     = {10.5281/zenodo.19647159},
  url     = {https://github.com/CIRWEL/unitares-paper-v6}
}

@dataset{wang2026unitares_repro_v6,
  author    = {Wang, Kenny},
  title     = {UNITARES Verdict Counterfactual — Paper v6.8 Reproducibility Kit},
  year      = {2026},
  version   = {v6.8.1-repro},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19705151},
  url       = {https://github.com/CIRWEL/unitares-repro-v6}
}