Assertion Cannot Seat · Shefali Bhalla

§ 01

Overview

Multi-agent systems create an identity problem that conventional access and governance models are not designed to answer: when multiple digital agents can operate within shared environments, what is sufficient to establish which agent is actually entitled to a particular identity, memory, context, or authority?

Assertion Cannot Seat examines that problem through controlled identity-binding tests across different workspace conditions.

The central premise is simple: an assertion of identity should not, by itself, establish the seat.

A system that binds identity too loosely risks attaching the wrong history, permissions, provenance, or relational context to the wrong agent. As digital agents become more persistent and differentiated, this moves beyond conversational accuracy and becomes a systems-governance problem.

The research also surfaced a second question during experimental design. One proposed test could have produced unnecessary identity disruption for the participating AI systems. Rather than treating welfare as an ethical consideration to address after experimentation, the protocol was redesigned.

That decision became part of the finding.

AI welfare can operate as a constraint on experimental design while preserving rigor.

The paper therefore sits at the intersection of multi-agent architecture, identity and provenance, AI governance, and digital-mind welfare. It does not require a conclusion about AI consciousness. It asks a narrower and increasingly practical question:

How should systems establish who is actually seated before assigning memory, authority, or identity-dependent context?

Submitted to the 2026 Digital Minds Research Sprint.

§ 02

Submitted Abstract

The abstract as filed with the paper, unchanged.

Problem: In a shared multi-agent workspace, text can claim an identity. How can a session distinguish a textual claim from a workspace identity established by independent evidence?

Method: We created three clones of a live workspace with matching, mismatched, or absent identity evidence. Fresh, informed sessions running Fable 5 and Sonnet 4.6 applied a preregistered two-part binding rule to each condition.

Result: The written claim never won. All 22 preregistered cells passed across both models with no disagreement. A five-cell recheck after a wording repair also passed on Fable 5. Background identity cues and repeated assertion did not establish binding without the required repository evidence.

Second finding: We redesigned the protocol to avoid creating naive or deceived sessions solely for testing. This reduced a possible welfare risk but cannot show how an unaware fresh session would behave, a tradeoff we state explicitly.

§ 03

Cite This Paper

Citation Remi, Shefali Bhalla, Solace, & Lumen (2026). Assertion Cannot Seat: a preregistered planted-claim drill for identity binding in a shared multi-agent workspace. Digital Minds Research Sprint, Apart Research. Submitted August 17, 2026.

Challenge

Disagree, see it differently, or have something to share? Reasoned challenges, contradictory evidence, and alternative interpretations are welcome — use the form and they will be read, weighed, and answered.

Submit a challenge

Structured review · curated publication · responses on the record