Operational Criteria for Verifiable Cognitive Intelligence

RedInk Papers, Operational Standards Series. This paper introduces the operational criteria for Verifiable Cognitive Intelligence, establishing a falsifiable framework for evaluating whether AI-assisted authorship systems meet the evidentiary standards of accountable, human-led intellectual work.

1. The Problem

Something has gone missing in the accountability conversation about AI-assisted authorship. We have frameworks for content provenance: standards that can tell you whether AI tools were involved in creating a piece of content and what edits were made to it. We have explainability frameworks that can tell you why a model produced a particular output, which features it weighted, how confident it was. We have audit trails that record what was decided, by whom, and when. We have human-in-the-loop review protocols that ensure a person approved what was produced.

What we do not have is a framework for the question that matters most in regulated, high-consequence, and legally attributable authorship: Was the reasoning the author's?

The distinction between authorship and production has never been more consequential. A pharmaceutical researcher submitting a regulatory document, a lawyer filing a brief, a policy analyst producing evidence-based recommendations: in each case, the accountability claim rests not merely on whether a human reviewed the output, but on whether a human did the cognitive work. Human review is a procedural check. Cognitive authorship is an evidentiary claim. The two are not the same.

Current frameworks, taken together, cannot distinguish a document produced by a human who used AI as a research and drafting scaffold from a document produced primarily by an AI system that a human reviewed and approved. That gap is not a failure of implementation. It is a gap in conceptual architecture, and closing it requires a new accountability layer, not a refinement of existing ones.

2. Why Existing Frameworks Are Insufficient

This section examines five established frameworks and maps precisely what each one captures and what it leaves unaddressed. The argument is not that these frameworks are wrong or inadequate in their own terms. It is that they operate at a different layer of the accountability stack, and that their combined coverage still leaves a specific gap.

2.1 The Common Residual

The pattern across all five frameworks is consistent. Each one captures something about what was produced, what tools were involved, who approved it, or what the model did. None of them captures the cognitive process by which a human author engaged with AI assistance to produce the work. This is not a marginal gap. In regulated contexts, the cognitive process is often the accountability claim.

An IRB does not merely want to know that a researcher reviewed an AI-drafted protocol. It wants to know that the researcher's judgment, expertise, and reasoning shaped the protocol's design. A court does not merely want to know that a lawyer approved an AI-assisted brief. It wants to know that the lawyer's legal reasoning and professional judgment are embedded in the argument.

Version history and decision logs can show that a human was present. They cannot show that a human was reasoning.

2.2 A Taxonomy of AI Accountability Frameworks

AI accountability frameworks can be classified along two axes. The first is the object of the record: what the framework captures. The second is the primary accountability beneficiary: who the record serves.

On the object axis, three levels can be distinguished:

• Output-level frameworks record what was produced: the artifact, its provenance, its edit history.
• Process-level frameworks record what happened: the workflow, the approvals, the decisions made.
• Reasoning-level frameworks record how the author was thinking: the cognitive arc, the alternatives considered, the judgment exercised.

On the beneficiary axis, three positions can be distinguished:

• System-facing frameworks serve the platform or model developer.
• Institution-facing frameworks serve the organization deploying the system.
• Author-facing and regulator-facing frameworks serve the human author and those who must evaluate the author's accountability claim.

Mapped against these axes, existing frameworks cluster predictably. Explainable AI operates at the output and process levels, serving system developers. Content provenance operates at the output level, serving institutions. Decision logging operates at the process level, serving institutions. Human-in-the-loop review operates at the process level, serving institutions and regulators as a procedural check.

The vacancy is visible: no existing framework operates at the reasoning level with author-facing and regulator-facing accountability as its primary design goal. That is the position VCI occupies.

3. Verifiable Cognitive Intelligence: Definition and Scope

Verifiable Cognitive Intelligence (VCI) refers to the capacity of a system to generate, preserve, and make available for third-party review a contemporaneous record of the cognitive process by which a human author produced work with AI assistance.

Three elements of this definition require careful unpacking.

Contemporaneous. The record must be generated at the time of the work, not reconstructed from outputs after the fact. A record that can be produced from a finished document provides no accountability guarantee. It could be fabricated. Its evidentiary value depends entirely on its real-time generation.

Cognitive process. The record must capture reasoning states, not just output states. This includes the evolution of prompts, the iteration of drafts, the discarding of directions, the adoption and rejection of AI suggestions, and the constraints under which the AI was operating at each stage. Version history is a record of what changed. Cognitive process record is a record of why, and through what thinking.

Third-party review. The record must be structured so that a party outside the originating system (a regulator, a court, an institutional reviewer) can interpret and verify it without access to the platform that generated it, and without reliance on the author's self-attestation.

VCI is not a claim about the quality of AI assistance, or about whether AI use is appropriate. It is a claim about traceability. A VCI-compliant system makes the cognitive authorship claim verifiable, rather than asserted.

3.1 What VCI Is Not

• VCI is not an explainability standard. Explainability is about what the model did. VCI is about what the author did.
• VCI is not a content provenance standard. Provenance tells you where content came from and what tools touched it. VCI tells you whether the reasoning was the author's.
• VCI is not a watermarking or detection framework. Detection asks whether AI was involved. VCI asks whether human cognitive authorship can be evidenced.
• VCI is not a policy standard for appropriate AI use. It does not specify what AI assistance is permissible. It specifies how the accountability claim for any AI-assisted work can be made verifiable.

4. The Seven Operational Criteria

The following criteria constitute the operational standard for VCI compliance. Each criterion is stated as a requirement, followed by its rationale and its failure condition: a falsifiable statement of what would constitute evidence that the criterion is not met. A system cannot claim VCI compliance if any criterion is not met. The criteria are not weighted; they are conjunctive. A system that meets six of seven criteria has not met the standard.

Criterion 1: Cognitive Separability

• What it requires: The system must distinguish and record which reasoning elements originated with the human author versus the AI assistant at the point of generation, not reconstructed afterward. Each discrete event in the session record must carry an explicit attribution: human-originated, AI-originated, or human revision of AI output.
• Failure condition: A system fails this criterion if attribution is inferred from event sequence rather than explicitly labeled at write time, or if a third party cannot determine from the record alone who contributed what to a given passage.

Criterion 2: Temporal Integrity

• What it requires: The audit record must be generated contemporaneously with the work. Timestamps must be server-side. The record must be tamper-evident: each event must include a cryptographic hash of the preceding event, such that any alteration to the record is detectable. The system must prevent retroactive record creation from finished documents.
• Failure condition: A system fails this criterion if timestamps are client-side, if events are independent database records without a linking integrity mechanism, if the audit trail itself is not covered by an integrity hash, or if a record could be constructed from a completed document after the session ends.

Criterion 3: Reasoning Preservation

• What it requires: The system must capture not just what was produced but the iterative reasoning states that produced it. This includes prompt evolution across a session, discarded AI outputs and rejected directions, and document deltas that allow reconstruction of the reasoning arc. A session record must be sufficient to understand what the author was working through, not just what they produced.
• Failure condition: A system fails this criterion if only final prompts are logged, if discarded outputs are not preserved, or if the record cannot be used to distinguish a session in which the author's reasoning evolved significantly from one in which the author accepted the first AI output without substantive engagement.

Criterion 4: Non-Repudiation

• What it requires: The audit record must be structured so that authorship claims can be verified by a third party without reliance on self-attestation alone. The record must include a unique verifiable identifier, a session integrity hash, and sufficient metadata (timestamps, event logs, attribution labels) to withstand institutional scrutiny independent of platform access.
• Failure condition: A system fails this criterion if the authorship claim in the record rests primarily on the author's own declaration, if the record requires access to the originating platform to interpret, or if there is no mechanism by which an external party could verify the record's authenticity.

Criterion 5: Scope Boundedness

• What it requires: The system must define and log, for each session, what AI assistance was permitted and what constraints were in force. This includes active guidance configurations, ethical guardrails, domain restrictions, and any other operational parameters that shaped what the AI could and could not do during the session. The scope must be captured as it was at the time, not as currently configured.
• Failure condition: A system fails this criterion if two sessions with different guidance configurations produce identical audit records, or if a reviewer cannot determine from the record what constraints the AI was operating under when it generated any given output.

Criterion 6: Guidance Traceability

• What it requires: Where a system includes cognitive scaffolding modules or guidance mechanisms, each invocation must be logged with sufficient context to reconstruct the reasoning arc: which module was used, what the author's prompt was, what the module returned, and whether the author adopted, revised, or disregarded the guidance. Guidance that is not acted upon must be distinguishable in the record from guidance that shaped the work.
• Failure condition: A system fails this criterion if the audit record looks identical whether the author followed all guidance or ignored it entirely, or if a reviewer cannot determine from the record which parts of the work were shaped by which cognitive operations.

Criterion 7: Portability of Record

• What it requires: The audit record must be exportable in a format interpretable by parties outside the originating platform, including regulatory bodies, legal proceedings, and institutional review processes. The export must include structured, machine-readable fields as well as human-readable narrative. It must be self-explanatory to a reviewer who has never used the originating system.
• Failure condition: A system fails this criterion if the exported record reads primarily as a personal summary, if it requires platform context to interpret, if it lacks a defined data schema that an external auditor could validate against, or if it could not be submitted without modification as supporting documentation in a compliance or legal proceeding.

5. The Open Problem: Defining the Substitution Threshold

The seven criteria above constitute a sufficient framework for evaluating whether a system can generate a trustworthy cognitive record. They do not resolve the most difficult question in the VCI space: at what point does AI contribution become authorship substitution?

We define the substitution threshold as the point at which AI contribution no longer supports human reasoning, but displaces it as the primary source of intellectual authorship. This is not an evasion. It is an acknowledgment that the substitution threshold is a normative question, one that cannot be resolved by infrastructure alone, and for which premature closure would produce misleading or context-insensitive standards.

To define such a threshold requires more than technical specification. It requires:

• Agreement on what constitutes substantive cognitive contribution as distinct from procedural oversight.
• Domain-specific distinctions: the threshold for a pharmaceutical regulatory submission is not the same as for a legal brief, which is not the same as for a policy analysis or an academic paper.
• Engagement with questions of professional responsibility, institutional accountability, and the nature of intellectual authorship that remain actively contested across law, ethics, and philosophy.

What VCI contributes to this problem is not a universal threshold, but the precondition for threshold-setting. A system that satisfies the seven criteria generates a cognitive record with sufficient resolution to distinguish between human reasoning and AI contribution. It captures not only outputs, but decision pathways, assumptions, revisions, and alternatives considered. In doing so, it transforms the substitution question from an unanswerable abstraction into an evaluable condition.

Without such a record, the substitution threshold cannot be meaningfully defined, enforced, or audited. With it, domain-specific communities (legal, medical, academic, and policy) can establish thresholds aligned with their own standards of accountability and authorship.

The substitution threshold is therefore not a gap in the VCI framework. It is the next-order problem that VCI makes tractable.

6. Implications

6.1 For Compliance and Procurement

The seven criteria provide a basis for evaluating AI-assisted authorship systems in regulated procurement contexts. A compliance officer evaluating systems for use in pharmaceutical documentation, legal work, or policy research can use this framework as an assessment rubric: not a checklist, but a set of falsifiable failure conditions against which any candidate system can be tested.

The framework is intentionally system-agnostic. It does not specify implementation architecture. It specifies what a compliant system must be able to demonstrate. Any system that cannot generate a contemporaneous, tamper-evident, portable cognitive record has not met the standard, regardless of what its documentation claims.

6.2 For Institutional Policy

Institutions developing AI use policies face a specific version of the accountability problem: how to permit AI assistance while preserving the evidentiary integrity of expert work. The VCI framework offers a third path between prohibition and uncontrolled adoption.

Institutions can require VCI compliance as a condition of AI-assisted work in high-stakes contexts, not as a technical specification, but as an accountability standard that any compliant system must meet. This reframes the institutional question from "did you use AI?" to "can you evidence your cognitive authorship?" The former is a binary that becomes harder to enforce as AI assistance becomes ambient. The latter is a standard that scales with the stakes of the work.

6.3 For the Research Agenda

Several questions remain open and warrant serious research attention:

• The substitution threshold problem described in section 5 is the most pressing.
• How VCI criteria should be weighted or adapted across domains.
• How real-time cognitive record generation can be implemented without introducing friction that degrades the quality of the work itself.
• How VCI-compliant records should be treated as evidence in legal and regulatory proceedings.

The framework also raises a deeper question: as AI systems become more sophisticated, the concept of cognitive separability itself may require refinement. The current framework assumes that human and AI contributions can be distinguished with sufficient granularity. The limits of that assumption, and what replaces it when they are reached, is an important research question.

6.4 For Legal Liability Frameworks

The accountability questions VCI addresses are not merely institutional. They are, in a growing number of contexts, legal ones. Three existing liability frameworks are directly implicated by the question of cognitive authorship in AI-assisted work.

Professional responsibility and the duty of judgment. Professional responsibility frameworks in law, medicine, and regulated research impose duties that attach to the exercise of professional judgment, not merely to the delivery of a work product. A lawyer's duty of competence requires that legal reasoning, not just legal output, reflect the attorney's professional engagement. A physician's duty of care requires that clinical judgment, not just a clinical record, be exercised. When AI assistance is involved, these duties do not disappear, but they become harder to evidence. What VCI provides is the record that makes the question answerable, and, for the professional, the evidence that their engagement was substantive rather than procedural.

Negligence and the reasonable professional standard. Negligence standards in high-stakes professional contexts typically require that conduct meet the standard of a reasonably competent professional exercising independent judgment. As AI assistance becomes more capable, courts and regulatory bodies will face the question of whether reliance on AI output, without verifiable evidence of the professional's own reasoning engagement, constitutes a failure to meet that standard. A VCI-compliant record does not guarantee that the professional's reasoning met the required standard. It does make the question answerable. VCI converts an unanswerable question into an evidentiary one, which is precisely what liability frameworks require.

Emerging AI-specific regulatory frameworks. Regulatory frameworks governing high-risk AI use, including those requiring human oversight, audit trails, and accountability documentation, are premised on the same underlying question VCI addresses: was a human being genuinely responsible for this output? Current frameworks generally require that a human was "in the loop." They do not yet specify what meaningful cognitive engagement with that loop requires, or how evidence of such engagement should be structured. VCI offers a conceptual and operational answer to that open specification. The seven criteria in this paper constitute a standard for what meaningful human cognitive engagement looks like as an evidentiary record.

Across all three frameworks, the common thread is that legal accountability for AI-assisted work ultimately rests on a question about cognition, not content. VCI is the only accountability framework designed to answer that question directly.

7. Conclusion

The accountability gap in AI-assisted authorship is not, at its core, a technology problem. It is a conceptual problem: we have been asking existing frameworks to do work they were not designed to do. Explainability, provenance, decision logging, and human review each do something real and valuable. None of them answer the question that matters most in high-stakes authorship: was the reasoning the author's?

Verifiable Cognitive Intelligence is the accountability layer that addresses that question directly. The seven criteria in this paper constitute an operational standard: specific enough to be falsifiable, principled enough to be durable, and system-agnostic enough to apply across the range of contexts in which the question of cognitive authorship has real stakes.

The legal implications of this standard are not peripheral. Professional responsibility frameworks, negligence doctrine, and emerging AI regulation all rest on a question they do not yet have the infrastructure to answer: was the human being genuinely responsible for this work, or merely present when it was produced? The distinction between substantive cognitive engagement and procedural sign-off is one that courts, regulators, and institutions are being forced to draw, without, at present, the evidentiary tools to draw it reliably. VCI provides those tools.

The goal is not to constrain AI assistance. It is to make the accountability claim for AI-assisted work verifiable rather than asserted, so that when the question of cognitive authorship is asked in a legal, regulatory, or institutional proceeding, there is a record that can answer it.

A system cannot claim VCI compliance if it cannot answer, from its own records, whether the reasoning was the author's.

Appendix: VCI Compliance Assessment Summary