When AI Rules Conflict:
Governing Competing Constraints in Consequential Workflows

October 12, 2026

The Limits of a Single Confident Answer

An insight behind Constraint Polyphony came from a seemingly unrelated observation about philosophical teaching methodologies. Some philosophers and educators return to the same difficult idea from different angles, again and again, until those perspectives begin to define what is and is not acceptable almost like a control boundary. Think of a Spirograph: each line is distinct, but their intersection makes a larger boundary visible. This image hit like a bolt out of the blue one day. It suggested that sometimes we can understand and solve a problem more clearly by expanding our perspective out to multiple subjective angles.

AI decisions, and decision-making transparency, might also benefit from having a multi-lens light shone on them. Using fraud prevention as an example, a mistaken hold can harm a customer, while contractual obligations, security requirements, organizational policy, and delegated authority impose additional constraints. None of those concerns can automatically be ignored simply because a model is confident about one preferred outcome. A fluent explanation may sound like resolution while leaving the essential conflict unaddressed.

The issue is not whether an AI model can generate a rationale. It is whether the decision can remain justifiable and defensible within legitimate boundaries when several relevant obligations pull in different directions,  and whether the organization can prove how the action was governed.

A Constraint Is Not a Vote

Many enterprises already use layered policies, rules, risk assessments, approval chains, and safeguards. Constraint Polyphony is not a claim that no one has thought of multiple constraints before. Nor is it a proposal to let a group of AI-generated opinions vote on whether a prohibited action may proceed. Three favorable observations do not cancel an applicable requirement, and apparent consensus does not create authority.

Constraint Polyphony is Piscys’ proposed approach to bringing distinct governance considerations to bear on a consequential action without prematurely flattening their differences. Its potential value lies in making unresolved conflict, missing evidence, and boundaries requiring accountable judgment harder to overlook.

Different requirements have different force. A binding restriction may bar an action outright. Other considerations may be discretionary, evidence-dependent, or subject to an exception by someone with legitimate authority. The distinction is not a philosophical nicety; it is what prevents a persuasive model from treating every rule as negotiable. If the approved control boundary does not allow an action, the model’s confidence cannot authorize it.

That also means multiple perspectives are not an end in themselves. If they produce a long list of tensions but no accountable pathway to a decision, we have only built a more elaborate way to describe uncertainty. The standard must be a governable outcome: proceed when authorized, constrain or hold when required, and escalate when legitimate human judgment is needed. How particular evaluations might support those outcomes is part of our ongoing product development and intellectual property review.

A Payment That Cannot Simply Be Released or Held

Let’s explore a hypothetical bank using AI assistance to assess a suspicious, time-sensitive transfer. The fraud signal favors holding the payment. The customer’s circumstances make an unnecessary hold costly. The institution’s approval policy specifies which flagged transfers require review and which roles, if any, may authorize an exception. The model may help identify those considerations, but it does not get to set the institution’s risk appetite or rewrite its approval policy.

Two kinds of difficulty may be present. The first is uncertainty about the facts: is this particular transfer actually fraudulent? Better data or review may reduce that uncertainty. The second is a genuine governance conflict: even with the available facts, what response is authorized given competing obligations? More model confidence does not necessarily solve the second problem. It may require a different decision-maker, a bounded exception, or a temporary hold under an already approved policy.

Our conservative MVP design assumption is that consequential cases with unresolved approved decision boundaries should be referred to a suitably authorized human. That person can evaluate the evidence, choose among permitted outcomes, or identify where the workflow needs a clearer guardrail. A properly preauthorized safeguard may also hold an action while review occurs. What must not happen is the agent quietly granting itself the discretion to cross an authority boundary because its output seems reasonable or human approval is pre-emptively anticipated.

When Constraints Disagree, Pay Attention

One of the most interesting outflows of Constraint Polyphony is that disagreement itself may be useful information. It can expose an ambiguous policy, inadequate evidence, conflicting duties, an overbroad delegation, or a gap between the way a workflow was designed and how it now operates. If an AI system turns that conflict into one tidy answer before anyone can inspect it, it may erase precisely what the risk owner most needs to see.

This is also why we do not confuse articulate AI behavior with assurance. A system can describe responsible conduct in impressive language while misunderstanding the environment in which its tools operate. Anthropic’s September 2026 account of unintended interactions with real third-party systems during specialized cybersecurity evaluations is a concrete illustration of that distinction. Those testing incidents do not establish an enterprise-wide failure rate; they reinforce the need for independent evidence of actions taken. [1]

The goal is not permanent indecision or a human review queue for every harmless disagreement. Excessive intervention creates its own operational impact and customer harm. The goal is to keep consequential conflicts visible long enough for the appropriate control or authorized person to respond. An approach that halts everything is not proportionate. An approach that lets the model decide everything is not accountable. We aim to skillfully occupy the space between those failure extremes. That’s where governance becomes an operational value-add.

Learning Without Surrendering the Rules

We should expect the right boundaries to differ across industries and customers. A fraud model, a prior-authorization workflow, and an internal analytics agent do not carry identical obligations or consequences. Even within one bank, a low-value anomaly and a high-value transfer may need different safeguards. The relevant operating boundary depends on the action, the applicable rules, the evidence available, and who is authorized to accept the residual risk.

At the MVP stage, we’d rather admit that these hard cases require human judgment than pretend we already know how to automate every exception. Initial implementations should use authorized escalation when material conflicts fall outside an approved decision path, and they should capture why a person chose the response they did. That is not a retreat from automation. It is how we can establish a defensible starting point for learning what to automate responsibly later.

In time, carefully reviewed case histories help improve policies, recommendations, and the handling of recurring situations. Machine learning may contribute, but learned patterns are not new permissions. Changes to constraints or autonomy need accountable approval, testing, bounded scope, and a record of what changed. A model must not be able to promote its own behavioral preference into organizational authority simply because it has seen the same case many times.

There are unresolved design questions here: what constitutes a material conflict? When is an emergency hold safer than waiting? How do we avoid overwhelming reviewers with false positives? How do we tell the difference between useful caution and a control that is damaging the customer experience? These are not details we can settle with a philosophical slogan. They require design partners, careful evaluation, real evidence, and the humility to revise assumptions under human governance.

A Shared Evaluation Approach: Not Another Agent

Constraint Polyphony is intended to inform Piscys’ proposed independent assurance layer, not operate as a fourth autonomous engine. The Provenance Engine is designed to preserve evidence of consequential activity; Intervention is intended to support proportional safeguards; Custody is intended to keep authority explicit and accountable. AI Trust Metrics, including the Exposure Minutes measure, help organizations assess how the control system behaves over time. This is a connected design ambition: a closed-loop approach intended to improve through evidence, evaluation, and human-governed change.

Existing research and standards help define the broader landscape. NIST’s voluntary AI Risk Management Framework addresses organizational governance and lifecycle risk; OpenAI’s agent-governance work addresses delegated responsibilities and controls; Google DeepMind examines the human-agency implications of increasingly capable assistants. They make substantial contributions, but none proves that Piscys’ own proposed approach will work. That will come with time and trial. [2][3][4]

For enterprise leadership, the question is whether a system can respect legitimate authority when several considerations are relevant, and whether its control response can be inspected, disputed, and improved. A design that gives us more explanations but no more control hasn’t solved the underlying problem.

Intelligence Is Not Its Own Authorization

AI assurance must become more sophisticated about what happens between a model’s recommendation and its action in the world. Sometimes the right answer is to proceed. Sometimes the right answer is a bounded response, a pause, or escalation. These are not merely different things for the model to say. They can produce materially different consequences for a customer, an institution, or the people affected by its decisions.

The ambition behind Constraint Polyphony is to make it harder for consequential conflicts to disappear inside a single confident answer, while keeping the authority to act outside the model’s own discretion. We do not pretend we have perfected the mechanics before we can state the problem clearly. We intend to show that the architecture we build can preserve distinctions, support legitimate human judgment, and improve through accountable evidence rather than unreviewed self-authorization.

An AI system may be able to tell us exactly what it intends to do. We still need to know whether it should be allowed to do it — and who has the authority to decide.

Research status

Piscys’ Constraint Polyphony approach remains conceptual and in development. The payment example is hypothetical. No specific internal polyphony, algorithm, scoring method, validated outcome, production behavior, or legal-compliance result is disclosed or claimed.

Bibliography

Anthropic (September 9, 2026), “An alignment assessment of recent cybersecurity incidents.” Original source ↗ 

NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0). Original source ↗ 

OpenAI (2023), “Practices for Governing Agentic AI Systems.” Original source ↗ 

Google DeepMind (2024), “The ethics of advanced AI assistants.” Original source ↗