Kenny Wang / Research

Reading guide · October 13, 2026 revisions

State, identity,
and accountability.

Four papers on making long-running AI agent work inspectable across sessions, changing runtimes, and separately controlled systems.

The papers connect runtime state estimation, continuity of identity, and attributable action. Each has its own evidence and open questions. The suggested reading order below starts with the concrete instrument audit, then widens to the framework, identity, and federated accountability.

Orientation

What EISV records

EISV is intended to help agents read their runtime state and operators inspect the system. Its implemented coordinates compress and smooth observations, including the controller’s own history:

E
A blend of prior policy actions, controller coherence level, complexity divergence, and available outcome, continuity, and tool features.
I
A calibration and consistency blend. The deployed calibration input is fleet-scoped; a reading is not a correctness score for the current claim.
S
A blend of drift, regime changes, complexity divergence, and available continuity and tool features.
V
The separately smoothed, signed gap between the raw E and I observations.

Controller feedback remains information about the system’s response. It does not independently confirm the condition that prompted that response. Measurement fidelity, diagnosis, usefulness to a reader, outcome prediction, and intervention benefit are distinct claims. A signal can be recorded and inspected while those further questions remain open.

Runtime definitions as of October 13: signal definitions · sensing, estimation, and regulation

01 · Instrument audit

A Flat Coherence Score

What information did one deployed score retain, and what did it contribute at the thresholds where it was used?

Evidence
Two frozen production windows (13,292 rows) and a reproducible basin-label replay showing that the legacy coherence score stayed concentrated near its ceiling while a class-conditional manifold replacement separated regime assignments.
Scope
Evaluates classification sensitivity in replay; downstream policy and task-outcome effects are tracked in the server’s evidence ledger.

Read v2.0.1 · Source and reproduction

02 · Framework

UNITARES

How can runtime observations support state estimation and governance across heterogeneous agents?

Evidence
A mathematical framework for runtime state estimation, dated deployment measurements, calibration analyses, and an explicitly bounded continuous-system contraction proof.
Scope
The deployed runtime uses weighted behavioral indices (see the runtime definitions above); information-theoretic formulations and discrete-time governor bounds are ongoing research.

Read v6.9.4 · Source, figures, and audits

03 · Identity

Trajectory Identity

Can patterns over time describe continuity beyond a reused identifier?

Evidence
A five-component trajectory-signature framework, longitudinal observations from an embodied Raspberry Pi agent, a twelve-episode recovery summary, and a four-agent multi-harness pilot.
Scope
Separates harness and role classes across models; isolating individual agent signatures across harness changes and eras is the next empirical step.

Read v0.18 · Source and analyses

04 · Accountability

Accountability Without
a Trusted Center

How can separately controlled systems establish who acted, under whose authority, and through which delegation chain?

Evidence
A federated testbed design, frozen evaluation protocol, single-server baseline trace, and an eleven-case Ed25519 attestation verification across two hosts.
Scope
Pre-registers the benchmark comparing federated and centralized control across independent operators.

Read v1.3 and supporting artifacts · Frozen evaluation registration

Companion evidence

Instrument quality has its own tests

Runtime evaluations test the state estimator’s axes directly: a synthetic S-axis experiment measured deterministic sensitivity to injected deviations and regime feedback, and an E/I input audit mapped controller-history coupling and calibration pair coverage. These findings guide instrument refinement and evaluation.

Faithful response, robustness, and usable explanations can be studied directly. Outcome prediction is a separate registered question, and intervention authority requires a separate decision.

S experiment · E/I input audit · Instrument-quality plan