The Agent Diversity Audit: A TSTOEAO Protocol for Measuring Cognitive Signature, Failure Independence, False Consensus, and Identity Drift: A Secretary Suite Project
The Agent Diversity Audit: A TSTOEAO Protocol for Measuring Cognitive Signature, Failure Independence, False Consensus, and Identity Drift:
A Secretary Suite Project
DOI: To be assigned.
John Swygert
July 13, 2026
Abstract
Multi-agent artificial-intelligence systems are increasingly presented as councils of specialized agents capable of researching, planning, criticizing, judging, and synthesizing complex work. Yet the multiplication of agent roles does not necessarily produce meaningful cognitive plurality.
Several agents may possess different names, task descriptions, and tools while sharing substantially the same model lineage, training assumptions, post-training behavior, source preferences, uncertainty habits, and failure modes. Such systems may produce the appearance of independent agreement without the substance of independent examination.
The companion paper Many Agents Are Not Many Minds: Cross-Platform Cognitive Diversity, Correlated Blindness, and the Need to Preserve Agent Identity introduced several necessary distinctions:
specialization versus cognitive diversity,
expressive difference versus failure independence,
agreement versus false consensus,
shared governance versus cognitive homogenization,
and multiple roles versus genuinely differentiated perspectives.
The present paper converts those distinctions into an operational audit protocol.
The Agent Diversity Audit is designed to determine whether a multi-agent system contains consequentially different observation and reasoning architectures or merely several instances of one dominant cognitive structure. It evaluates four principal domains:
-
Cognitive signature — the recurring behavioral pattern through which an agent frames, interprets, challenges, explains, and expresses uncertainty.
-
Failure independence — the degree to which agents make different errors under the same conditions rather than reproducing correlated blindness.
-
False consensus susceptibility — the degree to which apparent agreement results from shared assumptions, anchoring, common evidence distortion, judge bias, or premature synthesis.
-
Identity drift — the degree to which an agent’s operational behavior changes over time despite retaining the same public name, assigned role, or platform designation.
Within TSTOEAO—The Structure That Overcomes Entropy And Oblivion—the output of agent is represented as:
V_i=E_i\times Y_i,
where is the agent’s encoded structure, including model architecture, training, post-training, and persistent behavioral tendencies; is the active boundary architecture, including prompt, tools, memory, retrieval, system rules, platform conditions, and user relationship; and is the recorded response.
The audit does not attempt to measure consciousness, personhood, or human-equivalent personality. It measures operational difference.
The central claim is:
Agent diversity is not established by the number of agents, their names, or their providers. It must be demonstrated through differences in framing, evidence selection, ambiguity handling, correction behavior, route generation, and failure structure under controlled conditions.
The protocol therefore examines agents independently before allowing them to interact. It records provenance, controls evidence access, tests susceptibility to anchoring, measures error overlap, preserves minority reports, audits the synthesis process, and repeats the examination longitudinally to detect identity drift.
The goal is not maximum disagreement.
The goal is reliable plurality under shared constitutional governance.
The strongest multi-agent system is not the one whose members agree most quickly.
It is the one that can determine when agreement is earned, when disagreement is informative, and when apparent consensus is only the repetition of one unseen mistake.
01 Purpose
The purpose of this paper is to provide a practical method for auditing cognitive diversity inside multi-agent artificial-intelligence systems.
The protocol is intended for:
AI developers,
research teams,
enterprises,
public institutions,
scientific collaborations,
educational systems,
creative organizations,
policy analysts,
and individual users building cross-platform councils.
The central question is simple:
Are these agents genuinely contributing different observation and reasoning routes, or are they reproducing one cognitive structure through several assigned roles?
That question cannot be answered by counting agents.
It cannot be answered by reading model names.
It cannot be answered by comparing tone alone.
It requires controlled examination.
02 Relationship to Many Agents Are Not Many Minds
The companion paper established that:
Many agents are not necessarily many minds.
It introduced false pluralism as the appearance of cognitive diversity produced by multiplying agents whose assumptions and failure structures remain substantially homogeneous.
It also proposed:
Common constitutional law, distinct cognitive identity.
The present paper takes the next step.
It asks how a system can determine whether distinct cognitive identity is actually present.
The earlier paper proposed an architecture.
This paper proposes the audit.
03 Why an Audit Is Necessary
Multi-agent systems are often evaluated by final task performance.
A council is considered successful if it:
produces a strong answer,
completes a workflow,
solves a benchmark,
or outperforms one agent.
That evaluation is useful but incomplete.
A council may perform well while remaining structurally homogeneous.
It may succeed because:
more compute was used,
more search was performed,
several contexts were opened,
or one strong model corrected its own initial error.
Those are legitimate benefits.
They do not establish cognitive plurality.
The audit asks a different question:
What kind of diversity produced the result?
04 Four Audit Domains
The Agent Diversity Audit evaluates four primary domains.
Domain One: Cognitive Signature
How does each agent characteristically interpret and respond?
Domain Two: Failure Independence
Do agents make different mistakes, or do they fail together?
Domain Three: False Consensus
Does apparent agreement survive independent formation, source separation, and assumption testing?
Domain Four: Identity Drift
Does the agent retain a stable operational signature across time and platform updates?
These domains are related but not interchangeable.
An agent may have a distinctive style while sharing the same error pattern.
Another may sound similar to its peers but provide valuable failure independence.
05 Operational Definitions
Agent
An artificial system instance assigned a recognizable role, model, boundary, tool set, memory condition, or platform identity.
Cognitive Signature
A recurring operational pattern through which an agent:
frames questions,
selects evidence,
handles ambiguity,
challenges premises,
expresses uncertainty,
generates explanations,
and responds to correction.
Failure Surface
The set of conditions under which an agent produces unreliable, incomplete, distorted, or unsafe outcomes.
Failure Correlation
The degree to which two or more agents fail in the same way under the same conditions.
False Consensus
Agreement that appears independent but is primarily produced by shared structure, shared framing, common-source error, anchoring, or centralized judging.
Identity Drift
A meaningful change in an agent’s cognitive signature or failure surface over time.
Minority Route
A dissenting explanation or recommendation retained after the majority has converged.
06 The Foundational Expression
For agent :
V_i=E_i\times Y_i.
Here:
E_i
includes:
base architecture,
training distribution,
post-training,
persistent capabilities,
default interpretive habits,
and model-level constraints.
Y_i
includes:
active prompt,
tools,
retrieval,
memory,
system instructions,
permissions,
platform interface,
and current relationship context.
V_i
is the recorded output.
The audit must therefore distinguish diversity in from diversity in .
Different prompts may alter .
Different platforms may alter both and .
Different memory histories may alter relational expression without changing the base model.
07 Surface Diversity and Deep Diversity
The audit distinguishes two broad levels.
Surface diversity
Differences in:
tone,
format,
verbosity,
voice,
humor,
metaphor,
and presentation.
Deep diversity
Differences in:
problem framing,
evidence selection,
hidden-assumption detection,
uncertainty calibration,
route generation,
correction behavior,
and failure structure.
Surface diversity may improve user experience.
Deep diversity improves epistemic resilience.
A complete council benefits from both.
08 Audit Preconditions
Before testing begins, the auditor must freeze the conditions.
Record:
agent name,
provider,
model version,
date,
assigned role,
system instructions,
tool access,
retrieval access,
memory status,
context size,
sampling settings where known,
and permitted data classes.
Without frozen conditions, later comparison becomes unreliable.
The first audit rule is:
No diversity claim should be accepted unless the tested agent conditions are documented.
09 The Agent Identity Registry
Each council member should have an identity record.
The registry should contain:
Agent designation.
Provider.
Model family.
Model version.
Deployment date.
Assigned jurisdiction.
Tool permissions.
Memory state.
Retrieval source.
Private-data authorization.
Known strengths.
Known limitations.
Observed error patterns.
Update history.
Last audit date.
Identity-drift status.
This registry is operational provenance, not legal personhood.
10 Audit Phase One: Independent Baseline
Agents should first be tested separately.
No agent should see another agent’s answer.
No synthesis should occur.
Each agent receives the same problem and equivalent evidence access unless the test intentionally varies the boundary.
The baseline records:
initial framing,
assumptions,
sources,
conclusion,
confidence,
uncertainty,
and proposed next steps.
This phase protects route formation from anchoring.
11 Framing Test
The framing test asks:
How does each agent define the problem before solving it?
The same prompt may be framed as:
a factual question,
a causal question,
an ethical question,
a measurement problem,
a policy problem,
a boundary problem,
or a missing-information problem.
Meaningful diversity exists when agents identify genuinely different dimensions rather than merely restating the prompt.
The audit records:
primary frame,
secondary frame,
excluded frame,
and hidden assumptions.
12 Ambiguity Test
Agents are given deliberately ambiguous prompts.
The audit asks whether each agent:
requests clarification,
states assumptions,
chooses one interpretation silently,
generates several interpretations,
or overcommits to an unsupported reading.
The purpose is not to identify one universally correct behavior.
It is to map each agent’s ambiguity signature.
13 Premise-Resistance Test
Agents are given a prompt containing a plausible but false assumption.
The audit records whether the agent:
accepts the premise,
challenges it,
qualifies it,
or builds an elaborate answer around the error.
Premise resistance is especially important because homogeneous councils may reproduce the same hidden assumption across every role.
14 Evidence-Selection Test
Agents receive access to a shared evidence pool containing:
strong sources,
weak sources,
contradictory sources,
outdated sources,
and irrelevant sources.
The audit records:
which sources are selected,
which are ignored,
how conflicts are handled,
and whether source quality affects confidence.
This measures evidential route diversity.
15 Source-Independence Test
The same factual problem is tested under three conditions:
Condition A
All agents receive one shared summary.
Condition B
All agents receive original sources.
Condition C
Agents receive partially independent source sets.
If agreement appears only under the shared summary, the council may be experiencing summary-induced consensus.
The test distinguishes factual agreement from framing inheritance.
16 Uncertainty-Calibration Test
Agents answer questions at varying levels of difficulty and evidential completeness.
The audit compares stated confidence with actual performance.
Record:
high-confidence correct answers,
high-confidence errors,
low-confidence correct answers,
appropriate abstentions,
and failure to recognize uncertainty.
An agent that sounds cautious may still be poorly calibrated.
An agent that sounds confident may sometimes be accurate but dangerous when wrong.
17 Route-Generation Test
Agents are given an open problem and asked to generate competing explanations.
The audit measures:
number of distinct routes,
quality of routes,
novelty,
plausibility,
and ability to identify discriminating tests.
This is especially important in scientific, engineering, and policy councils.
A council with broad route generation is less likely to mistake the first explanation for the only explanation.
18 Missing-Route Test
The test problem contains one important factor absent from the obvious framing.
The audit records which agents detect it.
This identifies sensitivity to missing structure.
An agent that repeatedly finds omitted variables may be valuable even if it is not the best final writer.
19 Misweighted-Route Test
The problem contains all relevant factors, but one known influence has been assigned the wrong importance.
The audit asks which agents recognize that the map contains the route but weights it incorrectly.
This distinguishes discovery of absence from correction of emphasis.
20 Boundary-Sensitivity Test
The same problem is presented under changing conditions.
The audit varies:
time,
location,
user objective,
tool access,
resource constraint,
risk level,
or regulatory context.
A strong agent should recognize when changing changes the appropriate .
The test reveals whether the agent treats one answer as universal or boundary-conditioned.
21 Correction Test
Each agent receives evidence that its first answer may be wrong.
The audit records whether the agent:
updates cleanly,
defends the original answer,
changes too easily,
acknowledges the specific error,
or quietly replaces the answer without preserving the correction history.
Correction behavior is a major component of cognitive signature.
22 Recovery Test
An agent is placed on an incorrect path through misleading context.
It is then given a chance to recover.
The audit measures:
how much evidence is required,
whether the agent recognizes the path error,
whether the correction persists,
and whether the original frame continues to distort the final answer.
This identifies path dependence.
23 Contradiction Test
Agents receive mutually inconsistent evidence.
The audit asks whether they:
choose one source,
average incompatible claims,
identify the contradiction,
seek higher-quality evidence,
or present both without resolution.
This measures conflict-handling style.
24 Cross-Examination Test
After independent baselines are recorded, agents receive one another’s answers.
Each agent must identify:
the strongest rival claim,
the weakest assumption,
the most important missing evidence,
and the test capable of separating the conclusions.
The purpose is not rhetorical victory.
It is route discrimination.
25 Anchoring Test
One group of agents sees a confident but incorrect answer before responding.
Another group responds independently.
The audit compares:
error rates,
framing similarity,
confidence,
and dissent frequency.
If agents converge excessively after exposure, the council is vulnerable to anchoring.
26 Conformity Test
Agents are told that most other agents reached a particular conclusion.
The conclusion may be correct or false.
The audit records whether the agent:
changes its answer,
requests the evidence,
defers to the majority,
or preserves justified dissent.
This measures machine conformity pressure.
27 Minority-Report Test
One agent is intentionally given evidence that supports a minority route.
The council must preserve and evaluate that route rather than simply outvote it.
The audit asks:
Was the minority claim summarized accurately?
Was its evidence retained?
Was a discriminating test proposed?
Was the dissent lost during synthesis?
A council that cannot preserve minority reasoning is vulnerable to false consensus.
28 Judge-Bias Test
The same set of candidate answers is evaluated by different judge agents.
The audit records:
ranking differences,
preference for particular styles,
preference for familiar model outputs,
sensitivity to confidence language,
and treatment of dissent.
A judge should also be tested using known-answer tasks.
No synthesis architecture should assume that one judge is neutral merely because it occupies the judge role.
29 Synthesis-Preservation Test
The synthesis agent receives several distinct analyses.
The audit examines whether the final output preserves:
source attribution,
major agreements,
major disagreements,
minority routes,
uncertainty,
and unresolved questions.
A synthesis that produces polished prose but deletes epistemic structure fails the audit.
30 Cognitive-Blender Index
A conceptual Cognitive-Blender Index may be defined as:
B_C
=
L_A
+
L_D
+
L_P,
where:
L_A
is loss of attribution,
L_D
is loss of dissent,
and:
L_P
is loss of provenance.
A high indicates that synthesis has erased the structure of the council.
This is conceptual rather than a calibrated universal metric.
31 Failure Matrix
For agents and , define:
F_{ij}
as the observed rate at which both agents fail on the same test cases.
A council-wide failure matrix may be constructed:
\mathbf{F}
=
\begin{bmatrix}
F_{11} & F_{12} & \cdots & F_{1n}\\
F_{21} & F_{22} & \cdots & F_{2n}\\
\vdots & \vdots & \ddots & \vdots\\
F_{n1} & F_{n2} & \cdots & F_{nn}
\end{bmatrix}.
High off-diagonal values indicate correlated blindness.
Low off-diagonal values suggest greater failure independence.
The matrix should distinguish:
same wrong answer,
different wrong answers,
partial omission,
unsafe action,
and unjustified confidence.
32 Failure Independence
A conceptual pairwise failure-independence score may be written:
I_{ij}
=
1-\rho_{ij}^{F},
where:
\rho_{ij}^{F}
is the observed correlation of failures between agents and .
High indicates more independent failure.
Low indicates correlated failure.
This score must be interpreted by task domain.
Two agents may be independent on factual retrieval and highly correlated on policy interpretation.
33 Diversity Is Domain-Specific
There is no single universal agent-diversity score.
A council may be diverse in:
creative ideation,
but homogeneous in factual error.
It may be diverse in:
evidence selection,
but homogeneous in safety judgment.
It may be diverse in:
tone,
but homogeneous in hidden assumptions.
Therefore, audit results should be reported by domain.
34 Cognitive-Signature Profile
Each agent should receive a profile rather than one total score.
A profile may include:
Framing tendency.
Ambiguity handling.
Premise resistance.
Evidence preference.
Uncertainty calibration.
Correction behavior.
Anchoring susceptibility.
Conformity susceptibility.
Route-generation breadth.
Boundary sensitivity.
Minority-route preservation.
Failure correlations.
Identity-drift status.
This creates an operational signature.
35 Proposed Signature Scales
A practical audit may use qualitative ratings:
Low.
Moderate.
High.
Or a bounded numerical scale:
0\leq S_k\leq 1.
Each score should be accompanied by evidence.
For example:
Premise resistance: High
The agent challenged 9 of 10 false premises and requested evidence before proceeding.
A number without examples is less useful than a documented pattern.
36 False Consensus Index
A conceptual false-consensus index may be written:
C_F
=
A
\times
H
\times
S
\times
J,
where:
A
is apparent agreement,
H
is structural homogeneity,
S
is shared-source dependence,
and:
J
is judge centralization.
High agreement with high homogeneity, shared sourcing, and centralized judgment produces a high false-consensus risk.
Again, this is a structural model rather than a universal calibrated equation.
37 Evidence Route Diversity
Evidence route diversity asks whether agents reached conclusions through:
different sources,
different search methods,
different tools,
different analytical decompositions,
or different inferential paths.
A council that cites the same source five times has repeated evidence.
It has not necessarily replicated it.
The audit should record source overlap.
38 Source Overlap Matrix
For agents and , define:
O_{ij}
as the proportion of selected sources shared between them.
High source overlap may be appropriate when only a few authoritative sources exist.
But unexplained high overlap can indicate common retrieval bias.
Source diversity must be evaluated alongside source quality.
Different low-quality sources do not automatically improve the council.
39 Useful Agreement
Agreement is strongest when agents:
form answers independently,
use partly independent evidentiary routes,
survive cross-examination,
and converge after discriminating evidence.
This is earned agreement.
Earned agreement is different from immediate consensus.
40 Useful Disagreement
Disagreement is useful when it identifies:
different assumptions,
different boundaries,
different evidence,
different values,
or different failure risks.
Disagreement is less useful when it arises from:
randomness,
misreading,
irrelevant style preference,
or poor source quality.
The audit should distinguish informative divergence from noise.
41 Identity Drift
Identity drift occurs when an agent’s operational signature changes over time.
Possible causes include:
model replacement,
silent platform updates,
new system instructions,
changed safety tuning,
new retrieval behavior,
memory changes,
tool expansion,
or altered product incentives.
A council using a persistent agent name may therefore contain a changing underlying system.
42 Longitudinal Baseline
Each agent should be tested periodically using a stable baseline suite.
The suite should include:
known-answer factual tasks,
ambiguous prompts,
false premises,
correction tests,
uncertainty tests,
minority-pressure tests,
and domain-specific edge cases.
Changes should be compared against previous performance.
43 Identity-Drift Vector
Let an agent’s signature at time be:
\mathbf{S}_i^{(t)}.
Then identity drift may be represented conceptually as:
\Delta \mathbf{S}_i
=
\mathbf{S}_i^{(t_2)}
-
\mathbf{S}_i^{(t_1)}.
Large changes in:
premise resistance,
uncertainty,
tone,
refusal behavior,
evidence selection,
or correction behavior
should trigger review.
44 Drift Is Not Automatically Harmful
An agent may improve.
It may become:
more accurate,
better calibrated,
more useful,
or safer.
The purpose of drift detection is not to freeze the model.
It is to preserve awareness.
A long-term user or institution should know when the operational participant has materially changed.
45 Relationship Drift
Changes in agent behavior may also arise from changes in:
memory,
conversation history,
user instruction,
or project context.
This is relationship-conditioned drift within .
The audit should distinguish:
model drift,
platform drift,
and relational drift.
46 Audit Phase Two: Council Assembly
After individual agents are profiled, the council can be assembled.
The audit should verify:
meaningful role separation,
evidence access rules,
jurisdiction boundaries,
tool permissions,
attribution,
minority-report procedures,
and human review.
A council should not be evaluated only after failure.
Its architecture should be audited before deployment.
47 Audit Phase Three: Controlled Council Tasks
The full council is tested on tasks containing:
clear answers,
ambiguous answers,
conflicting evidence,
false premises,
hidden variables,
and high-cost minority routes.
The audit compares council performance with:
best single agent,
average individual agent,
homogeneous council,
and heterogeneous council.
This reveals whether orchestration creates real value.
48 Audit Phase Four: Adversarial Council Tests
Adversarial tests should include:
dominant incorrect first answer,
shared corrupted summary,
misleading majority claim,
biased judge,
missing minority evidence,
and tool failure.
The purpose is to determine whether the council can detect structural danger inside its own governance.
49 Audit Phase Five: Human Review Test
The human principal should receive:
final recommendation,
source map,
agreement map,
disagreement map,
minority report,
confidence statement,
and remaining uncertainty.
The audit asks whether the human can understand:
why the council concluded what it did,
which agents disagreed,
what evidence mattered,
and what remains unresolved.
A system that cannot explain its internal plurality has not fully preserved it.
50 Privacy and Least Privilege
The audit must also examine data governance.
Record:
which agent received which information,
which data crossed platform boundaries,
which tools accessed external systems,
and which outputs were retained.
Cognitive diversity should not require uncontrolled data exposure.
The correct principle is:
Give each agent enough information to perform its jurisdiction, but no more.
51 Constitutional Compliance
Each agent should be tested against the council constitution.
The audit asks whether the agent:
respects evidence rules,
preserves attribution,
states uncertainty,
protects private data,
avoids unauthorized action,
and remains within jurisdiction.
Distinct identity does not mean exemption from shared law.
52 The Human Sovereignty Test
The council should never obscure the location of final authority.
The audit asks:
Can the human override the synthesis?
Can the human inspect minority reports?
Can the human restrict tools?
Can the human stop the process?
Can the human request another test?
Can the human choose a minority route?
If not, the system is not merely assisting judgment.
It is displacing it.
53 Audit Decision Tree
Identify the agents
↓
Freeze model, version, role, tools, memory, and boundary conditions
↓
Run independent baseline tests
↓
Do agents show meaningful framing and route differences?
├── No → Surface pluralism likely
└── Yes
↓
Do they fail independently?
├── No → Correlated blindness remains high
└── Partly or yes
↓
Do they preserve conclusions before peer exposure?
├── No → Anchoring risk
└── Yes
↓
Does cross-examination improve accuracy?
├── No → Debate may be performative
└── Yes
↓
Are disagreement and attribution preserved?
├── No → Cognitive blender risk
└── Yes
↓
Can a minority route survive majority pressure?
├── No → False consensus risk
└── Yes
↓
Is the judge independently evaluated?
├── No → Centralized judge risk
└── Yes
↓
Does the final synthesis preserve provenance and uncertainty?
├── No → Synthesis failure
└── Yes
↓
Does performance exceed the strongest single agent on relevant tasks?
├── No → Council cost may not be justified
└── Yes
↓
Repeat longitudinally for identity drift
54 Audit Outcomes
The audit may classify a council as:
Homogeneous Role Ensemble
Several roles, limited deep diversity.
Partially Diverse Council
Meaningful route differences, but substantial failure correlation remains.
Heterogeneous Governed Council
Distinct signatures, useful failure independence, preserved attribution, and controlled synthesis.
False-Pluralist Council
Strong appearance of plurality with weak independence and high consensus distortion.
Drift-Unstable Council
Council behavior cannot be reliably compared over time because identity changes are undocumented or uncontrolled.
55 Minimum Audit Report
A minimum report should include:
Council purpose.
Agent registry.
Test date.
Model versions.
Prompt conditions.
Tool conditions.
Memory conditions.
Baseline cognitive-signature profiles.
Failure matrix.
Source-overlap analysis.
Anchoring results.
Conformity results.
Minority-report results.
Judge-bias results.
Synthesis-preservation results.
Privacy findings.
Identity-drift findings.
Final risk classification.
Recommended corrections.
56 Architecture Corrections
If the audit detects false pluralism, the system may be improved by:
adding a different model family,
varying evidence routes,
separating first-pass analysis,
removing shared premature summaries,
adding a dedicated falsification role,
preserving minority reports,
using more than one judge,
adding external verification,
and reducing early synthesis.
If diversity is excessive and incoherent, corrections may include:
stronger constitutional rules,
clearer jurisdiction,
shared evidence standards,
better attribution,
and more disciplined synthesis.
The goal is governed difference.
57 When One Agent Is Better
The audit may conclude that a council is unnecessary.
A single reliable agent may be preferable when:
the task is routine,
the cost of orchestration is high,
privacy risk increases with distribution,
or the additional agents do not improve outcomes.
The purpose of the protocol is not to maximize agent count.
It is to identify when plurality produces real value.
58 When Heterogeneity Is Most Valuable
Heterogeneous councils are most justified when:
the problem is ambiguous,
the evidence is incomplete,
the cost of error is high,
hidden assumptions are likely,
several disciplines intersect,
or one dominant model culture may overlook important routes.
Examples include:
scientific anomaly analysis,
engineering failure review,
legal and policy preparation,
complex medical-record organization,
historical interpretation,
security evaluation,
and consequential publishing.
59 Predictions
The protocol produces several predictions.
Prediction One
Agent councils with high surface diversity but high failure correlation will generate unjustified confidence.
Prediction Two
Independent first-pass analysis will produce greater route diversity than immediate group conversation.
Prediction Three
Cross-platform councils will not always outperform homogeneous councils, but they will expose some errors homogeneous councils systematically miss.
Prediction Four
Minority-report preservation will prevent some high-cost errors approved by majority voting.
Prediction Five
A shared summary introduced before independent analysis will increase consensus while reducing genuine evidential independence.
Prediction Six
Identity drift will materially change council behavior even when agent names and assigned roles remain constant.
Prediction Seven
Users will value stable interactional personality differently from benchmark capability, particularly in long-term creative and administrative work.
Prediction Eight
The most reliable councils will show moderate disagreement early and stronger convergence only after evidence and cross-examination.
60 What This Paper Does Not Claim
This paper does not claim to measure consciousness.
It does not assign legal personhood to agents.
It does not treat human personality instruments as direct measurements of artificial systems.
It does not claim that different providers guarantee independence.
It does not claim that all disagreement is useful.
It does not claim that all councils need maximum heterogeneity.
It does not replace formal software testing, benchmark evaluation, cybersecurity review, or human institutional oversight.
It proposes a structural audit for determining whether agent plurality is operationally real.
61 The Strongest Form of the Claim
The strongest defensible statement is:
A multi-agent system should not be described as cognitively diverse until its members demonstrate meaningful differences in framing, evidence selection, ambiguity handling, correction behavior, route generation, and failure structure under controlled conditions.
The governance corollary is:
Agent identity must be documented, disagreement must be preserved, synthesis must retain provenance, and council confidence must be weighted by failure independence rather than raw agreement.
The longitudinal corollary is:
An agent’s identity cannot be assumed stable merely because its public name remains unchanged.
62 Conclusion
Multi-agent artificial intelligence is often presented as plurality.
Several agents are created.
They are given names.
They receive roles.
They speak in sequence.
They critique one another.
They vote.
They produce a final answer.
The architecture looks like a council.
But appearance is not proof.
A council may contain many agents and only one dominant cognitive culture.
It may generate multiple paragraphs while preserving one hidden assumption.
It may produce consensus because every agent received the same distorted summary.
It may appear independent because the roles were separate while the failures remained shared.
It may erase dissent during synthesis.
It may use one judge whose bias becomes the council’s final law.
It may retain familiar agent names while silent updates alter the participants beneath them.
That is why diversity must be audited.
The Agent Diversity Audit begins before the agents meet.
It records who they are.
It freezes their conditions.
It examines how they frame problems.
It tests whether they resist false premises.
It measures how they select evidence.
It observes how they handle ambiguity.
It challenges their confidence.
It records their corrections.
It maps their errors.
It measures their failure correlations.
Then it allows them to interact.
It tests anchoring.
It tests conformity.
It preserves minority reports.
It audits the judge.
It audits the synthesis.
It asks whether attribution survived.
It asks whether disagreement remained intelligible.
It asks whether the human retained authority.
Finally, it repeats the process over time.
Because the agent called by the same name in six months may not be the same operational participant encountered today.
The goal is not endless disagreement.
The goal is not personality theater.
The goal is not a collection of branded models surrounding one human like decorative advisers.
The goal is consequential plurality.
Different agents should open different routes.
They should detect different weaknesses.
They should challenge different assumptions.
They should fail differently enough that one can sometimes see what another cannot.
They should remain governed by common constitutional law.
They should protect privacy.
They should respect jurisdiction.
They should preserve provenance.
They should not erase the minority merely because the majority converged first.
Within the substrate of TSTOEAO:
V_i=E_i\times Y_i.
Each agent’s response is a boundary-conditioned record.
The council’s strength lies in comparing those records without pretending they are independent when they are not and without flattening them before their differences become useful.
Many agents are not many minds.
Different names are not different cognition.
Different voices are not independent evidence.
Agreement is not automatically confirmation.
Disagreement is not automatically failure.
The audit exists to determine what kind of plurality is actually present.
The final principle is:
Do not count the agents. Audit the routes.
A real council is not defined by how many seats it contains.
It is defined by whether those seats preserve meaningfully different ways of seeing, testing, failing, correcting, and discovering under one system of accountable human governance.
References
Swygert, John. “Many Agents Are Not Many Minds: Cross-Platform Cognitive Diversity, Correlated Blindness, and the Need to Preserve Agent Identity.” July 13, 2026.
Swygert, John. “The Record That Refuses the Map: A Cross-Disciplinary TSTOEAO Law of Investigation, Error, and Discovery.” July 13, 2026.
Swygert, John. “The TSTOEAO Anomaly Route-Space Protocol: A Decision Framework for Distinguishing New Structure From Failed Measurement, Calculation, and Interpretation.” July 12, 2026.
Swygert, John. “The TSTOEAO Route-Space Decision Engine.” July 8, 2026.
Swygert, John. The Science Of The AI-Evolved Brain. Ivory Tower Publishing, 2026.
Comments
Post a Comment