UNITARES v6

Information-Theoretic Governance of Heterogeneous Agent Fleets

• 📄 unitares-v6.pdf — the paper (30 pages, latest tag paper-v6.9.1)
• 📝 unitares-v6.tex — LaTeX source
• 🖼 figures/ — 6 figures from deployment data (Feb 20, 2026 snapshot; Apr 18, 2026 30-day window)
• 🗒 Release history — per-version notes and frozen PDFs

---

In one line: UNITARES tracks each AI agent's behavioral state continuously, intervenes when an agent starts drifting, and produces an audit trail of the interventions.

---

What the paper argues

Governance of autonomous AI agents is usually assembled from output filters, dashboards, and post-hoc evaluators — each operating on a population the designer imagined as roughly homogeneous. UNITARES argues that this assumption fails once a real fleet is heterogeneous: embodied agents with sensor-driven state, persistent autonomous services, session-bounded coding assistants, and ephemeral parser agents do not share an output modality, a tempo, or a healthy operating point. Fleet-wide normalization is then the wrong target, and a governance framework that collapses the population to a single distribution — what we call cosmological soup — loses signal exactly where signal matters.

The paper makes two structural moves:

1. Information-theoretic grounding of the EISV state vector — S as Shannon entropy of the response distribution, I as mutual information between context and response, E as negative variational free energy, V as accumulated free-energy residual. Each coordinate ships with a tiered computation recipe (logprob, multi-sample, heuristic) and a provenance-tagged scale constant.

2. Class-conditional calibration — scale constants and healthy operating points are functions of an agent class, keyed on existing identity tags, with fleet-wide defaults retained only as fallback. Phase 2 measurement on five production classes (Lumen, Sentinel, Vigil, Watcher, default) over a 30-day window is reported in §11.5; per-class manifold radii span a 3.3× range, directly contradicting the fleet-wide assumption.

The headline finding

A verdict counterfactual on a 13,310-row production slice (§11.6) shows that 28.9% of governance basin assignments flip when the grounded class-conditional coherence replaces the legacy fleet-wide tanh form on the same state vectors. Flip rates range 15–33% per class, with a dominant directional bias into the low basin — the fleet-wide form was systematically permissive relative to tighter class envelopes. The homogenization-failure-mode argument has first-order consequences at the gating layer, not only at the reported-value layer.

This is a static reclassification measurement: the same state vectors, two coherence formulas, count the disagreements. It does not depend on the multi-agent coupling machinery of §sec:network (which v6.7 reframed as a theoretical extension rather than a deployed mechanism — see Version history below), and its conclusion survives that correction intact.

The methodological contribution

Re-grounding a deployed framework while service continues is a live-systems problem distinct from the mathematical one. §12 documents a pipeline-ordering migration mechanism that separates decision-producing from presentation-producing pipeline stages: governance verdicts fire on legacy values while the grounding computation runs as a post-verdict enrichment, so grounded values populate canonical response fields without changing live governance behavior. The mechanism generalizes beyond UNITARES — any deployed agent framework facing a semantic-coordinate change can apply it, given a request pipeline with explicit ordering between decision and presentation stages.

Scope

The contraction analysis (Appendix B), the grounded coordinate definitions (§4), and the class-conditional calibration protocol (§8) are fully specified. The manifold-coherence path is measured per-class on production data. The S_scale, I_scale, E_scale constants — used only by the higher-fidelity tier-1 (logprob) and tier-2 (multi-sample) estimators — remain provenance-tagged placeholders awaiting inference-layer instrumentation. Per-class results in §11.5 are a descriptive measurement on the deployed system rather than a controlled validation study across stable parameter regimes; the paper flags this explicitly (§11.7).

The §Multi-Agent Network section (§sec:network, eq:network, thm:network) presents a theoretical coupled-dynamics extension — not the deployed mechanism. The within-class synchronization theorem is used only as a qualification on homogenization claims elsewhere in the paper.

Reading aids

• Variational free energy (−F, used as the E coordinate) — the Free-Energy-Principle quantity measuring how well an agent's predictions match outcomes; high E = low surprise.
• Cosmological soup — the failure mode of averaging a heterogeneous agent fleet into a single state distribution, in which no population structure survives normalization.
• Manifold coherence (C = 1 − ‖Δ‖₂ / ‖Δ‖_max) — distance from the agent class's healthy operating point, used as a class-conditional replacement for the legacy tanh-of-V coherence.
• Basin flip — a change in governance-basin assignment (high / boundary / low) when the same state vector is classified under different coherence formulas.

Companion code

The deployed governance system this paper describes:

• CIRWEL/unitares

The grounding implementation (Phases 1+2 referenced in §12) was merged to master via PR #26.

Building the PDF

Either toolchain works; Tectonic is preferred for reproducibility since it downloads its own font and package snapshots:

# Tectonic (self-contained, single pass)
tectonic unitares-v6.tex

# Classic TeX Live (two passes for cleveref + bibliography)
pdflatex unitares-v6.tex
pdflatex unitares-v6.tex

Required packages: amsmath, amssymb, mathtools, bm, graphicx, booktabs, hyperref, cleveref, natbib, xcolor, caption.

Version history

Per-release notes and frozen PDFs are on the GitHub releases page. Headline deltas:

• v6.8 — operational-honesty bump: §12.4 adds a paragraph on the 2026-04-20 KG retrieval rebuild (pre-rebuild semantic scores sat at noise floor; fleet-learning claims now bounded by that era); §12.5 adds item 6 explicitly caveating retrieval-infrastructure and the PR-#54 auto-ephemeral write-bug. Grounded EISV counterfactual unaffected (inputs are core.agent_state, not KG content).
• v6.7 — drift correction: runtime removed the CIRS v2 neighbor-pressure coupling on 2026-04-17; paper caught up by excising §Multi-Agent Coordination and reframing §Multi-Agent Network as a theoretical extension. §11.6 headline finding unaffected.
• v6.6 — added §11.7 identity-system-maturity disclosure; added identity hardening track to §14.5 future work.
• v6.5 — submission polish; aligned prose with measured per-class calibration.
• v6.4 — added facilitator-load caveat on the §11.4 dialectic statistics (45/66 non-resolution rate framed as upper-bound under human-facilitator load rather than protocol failure).
• v6.0–v6.3 — initial v6 release and early revisions.

Citation

@article{wang2026unitares-v6,
  author    = {Wang, Kenny},
  title     = {{UNITARES}: Information-Theoretic Governance of Heterogeneous Agent Fleets},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19647159},
  url       = {https://doi.org/10.5281/zenodo.19647159},
  version   = {v6.9.1}
}

The concept DOI above resolves to the latest version; each tagged release (paper-v6.0 through paper-v6.9.1) has its own version DOI.

License

Paper text and figures: CC-BY 4.0

Author

Kenny Wang, Independent Researcher CIRWEL Systems