Answers to Critics: Where the AI Governance Framework Holds and Changes

A direct response to serious criticism of the framework, including what changed, what was clarified, and what still holds.

An adversarial review identified weaknesses that cannot be answered by saying the critic misread the framework. Some require refinement. Others concern precision without changing the substance, and some concern tone rather than argument.

The AI-First / EKA Tension

AI-First Software says the system can “dynamically select algorithms and approaches” and that “the user describes outcomes; the system determines approach.” Executable Knowledge Architecture insists that expert judgment stays in the method, and the AI translates that method into execution. A hostile reader calls this a contradiction.

The apparent conflict is a valid tension that requires reconciliation.

AI-First Software was written before agentic frameworks automated the routing decision. It describes interface design: systems acting on natural-language requests. EKA specifies the governance requirements for defensible professional work.

I should have made the compatibility condition explicit from the beginning: the human must have already decided to require executable artifacts. This refines the proposal without retracting it.

“System determines approach” is acceptable when it means: the AI selects implementation details within a scope the user has authorized. For consequential work, that scope must be: produce inspectable methodology that I can verify, not direct answers I must trust.

Update for February 2026 explains how agentic frameworks can automate the routing decision and remove that human control. EKA requires users to retain the choice of whether the system must produce code.

The Verification Protocol

The framework distinguishes reviewing a method from verifying a direct answer, but also suggests that experts can rely on AI to explain code. That reliance reintroduces oracle trust: the expert would be checking a generated account of the code rather than the code itself.

I should not have written that experts “don’t even need to read code.” That claim was too strong. Experts need not author code, but verification requires more than an AI explanation. The critique correctly identified a weakness in the original verification requirement.

The research shows that reasoning traces can be unfaithful. Anthropic’s paper “Reasoning Models Don’t Always Say What They Think” directly challenges whether exposed chain-of-thought represents actual model cognition. Models achieve only 81-83% factual accuracy in intermediate reasoning steps. If this unfaithfulness applies to AI explanations of generated code (and there’s no reason to think it doesn’t) then verification through AI explanation alone is insufficient.

Experts can review a method without being able to author it, but review must include evidence independent of the generating model.

The verification protocol depends on domain, stakes, and available tools. It must include a check independent of the generating model’s description. Organizations can implement that through human code review, automated tests, secondary models, formal verification, or domain-specific validation. The implementation is a design choice; independence is a requirement.

Determinism and Reproducibility

The claim that identical inputs produce identical outputs needs to account for production conditions. Environment and data drift, concurrency, and approximate solvers can invalidate assumptions of determinism.

The original determinism claim needed qualification for production use.

Deterministic execution means deterministic given a captured environment, versioned dependencies, and a fixed input snapshot. Production reproducibility requires these controls. Without them, “same code, same output” is aspirational rather than actual.

With those controls, assistant artifacts offer a way to reproduce results that direct stochastic oracle outputs do not.

The Strawberry Experiment

Critics argue that letter counting favors the proposed approach: models often fail at direct counting while code handles it trivially. They ask for evidence behind the 100% result and note that better prompting might improve direct answers while the code prompt already specifies the solution procedure.

The strawberry experiment illustrates a mechanism; it does not benchmark general performance.

I chose letter counting because the task is unambiguous and clearly demonstrates the failure mode. The experiment shows how different generated implementations can produce the same deterministic result.

The experiment establishes how the mechanism operates on that task, without establishing universal applicability.

Update for February 2026 cites broader research: reproduction attempts in which accuracy fell from 79.9% to 45.6%, 99% confidence intervals that contained the truth only 65% of the time, and studies of which only 34% reported temperature settings. These findings provide stronger general evidence than the 700 strawberry runs.

Absolute Language

The absolute wording of “ungovernable” and “solves” invites objections based on partial mitigations such as sampling audits, confidence calibration, and source grounding.

The criticism concerns semantics, but it correctly identifies a need for more precise scope without changing the underlying claims.

“Ungovernable” means specifically: cannot satisfy the three requirements of reproducibility, verifiability, and accountability that professional work under risk demands. It doesn’t mean “no mitigation is possible.” It means “no mitigation achieves a defensibility standard adequate for professional accountability.”

“Solves” means: creates a viable path to defensibility. It doesn’t mean “eliminates all risk” or “requires no further effort.”

Use these narrower meanings:

• “Ungovernable” → “doesn’t meet a defensibility standard for accountable professional work”
• “Solves” → “creates a viable path to defensibility”

“Solved Language” and “Hallucinations Are Manageable”

AI-First Software uses “solved language” and “formal competence,” claims reliable natural-language action at scale, and calls hallucinations manageable. Linguists, engineers, and safety reviewers can challenge those claims, which the governance argument does not require.

The claims concerned interface design, rather than metaphysical truth, but my wording invited misreading.

By “solved language,” I mean that systems can map natural language to structured actions well enough to change interface design. The intended claim was operational. If the wording suggests a maximal claim about understanding, that is a writing failure.

Hallucinations can be managed when high-stakes outputs expose their methods, can be tested, and require explicit review. Those engineering conditions reconcile the claim with the unreliability of direct oracle answers.

The governance approach can work with imperfect language understanding if the artifacts are inspectable and testable. The stronger language claims were unnecessary rhetorical overreach.

Security and the Universal Interface

Agentic AI as Universal Interface describes agents exploring interfaces, reading documentation, and connecting services. Unconstrained trial and error against production systems creates security risks that require a threat model.

The original essay gave inadequate attention to security risks because it concentrated on capabilities. The criticism correctly identified that omission.

For agentic AI to function as universal interface in professional contexts, constraints must be architectural:

• Least-privilege permissions: grant only what the task requires
• Read-only defaults for exploratory work
• Explicit approval gates for any write action
• Logging of tool actions
• Separation between confidential contexts
• Secure handling of secrets

Deployments must enforce these controls. An agent that persists in exploring how systems work can be dangerous without them.

Concentrating on capability without threat modeling left a gap in the original essay.

“Obfuscation Buys Time”

The argument that obfuscation isn’t durable and that “the effort to circumvent is worth it when value is high” can be read as endorsing bypassing controls. That’s legally and ethically fraught.

The argument concerns customers choosing alternatives and competitors building replacements.

Organizations that rely on obfuscation to protect value create incentives for customers to choose more open alternatives and for competitors to build replacements. Obfuscation creates pressure toward openness, not durable protection.

“Effort to circumvent” was poor wording. The argument does not authorize bypassing controls. Obfuscation invites competition from more accessible systems.

“Most” Tacit Knowledge

The claim that most tacit knowledge could be expressed but has not been implies a measured proportion without supporting data. Organizational-learning researchers, skilled tradespeople, and ethnographers have grounds to challenge it.

The evidence does not justify a claim about most tacit knowledge.

More organizational knowledge than organizations assume can be expressed with the right vocabulary, interlocutor, or incentive. A substantial fraction of knowledge called tacit has these representational limits.

The distinction between structural inexpressibility and representational limits remains useful, but does not establish a numerical proportion.

The Measurement Gap

The framework identifies benchmarks’ failure to measure defensibility and process but offers no complete solution. That criticism calls for a concrete measurement proposal.

The framework needs measurements of assistant-pattern performance.

A measurement agenda aligned with the framework would track:

• Reproducibility: Can a method-bearing artifact reproduce outputs across time given captured inputs and environment?
• Failure-mode distribution: When failures occur, what type? Wrong method selection, wrong data mapping, wrong assumption, execution error?
• Calibration: For any component that still produces conclusions, what’s the alignment between claimed confidence and actual correctness?
• Defensibility indicators: Not a scalar benchmark, but a rubric: how inspectable is the method, how complete is provenance, how easy is independent replication?

These measures provide a starting point for a measurement program, with further work required.

Capability Governance and Legal Precision

Professional responsibility includes the legal standard of care, and clients may interpret professional stamps as claims of correctness. A normative governance proposal needs to distinguish those legal issues.

The framework addresses artifact governance and does not provide legal advice.

Professional liability, regulatory compliance, and insurance requirements vary by jurisdiction and domain. The framework provides a way to think about artifact defensibility. It doesn’t replace qualified legal and regulatory advice for specific situations.

Update for February 2026 provides citations for factual claims such as regulatory requirements. Recommendations about how governance should work remain normative arguments, distinct from legal conclusions.

The 33% Penetration Claim

EKA in Practice reports “33% penetration across North America” without defining the metric. Critics will ask: 33% of what?

Within three weeks of launch, 33% of eligible professional staff used the system to produce at least one client-facing artifact.

The figure measures adoption, not frequency of use or the tool’s importance in daily work. The compounding effects described in the deployment essay concern what happened afterward; the initial adoption number does not prove them.

Rhetorical Choices

Lines like “Software is dead. Long live software,” “If it can do the job reliably, it has the knowledge. That’s the whole claim,” and the “Amish hand-crafted furniture” aside are punchy but risky. They invite critics to focus on tone rather than substance.

I defend these as style choices.

“Software is dead. Long live software” is deliberately provocative and intended to be memorable rather than measured. The surrounding paragraphs supply the argument; the slogan attracts attention to it.

“That’s the whole claim” states the scope of my argument directly. A substantive objection needs to challenge the argument rather than the confidence of the wording.

The Amish furniture comparison concerns competition. Someone who finds it dismissive can argue that handcraft will remain economically viable against AI-assisted competitors. I am skeptical.

Critics who concentrate on tone rather than substance reveal how they approach the work. The arguments need to be judged on their merits, irrespective of how much I hedge.

The Consciousness Tangent

Knowledge as Capability goes into consciousness and solipsism territory, suggesting that denying AI mental states requires a leap comparable to affirming human ones. This is unnecessary for the operational definition and introduces a controversial topic that doesn’t advance the governance framework.

The philosophical discussion addresses an objection to the operational definition, but readers can use the definition without it.

Dismissing AI as unable to understand assumes criteria for knowledge-constituting mental states that no discipline has established. That uncertainty explains why the operational definition leaves the question unresolved.

The definition of knowledge as reliable capability does not depend on resolving metaphysical questions about consciousness.

The section includes a link to the operational definition for readers who want to skip the philosophical discussion.

Summary

The review identified requirements and qualifications:

Genuine refinements:

• AI-First / EKA reconciliation: human must control the routing decision
• Verification protocol: requires independence from the generating model
• Determinism: qualified to “given captured environment and inputs”

Precision improvements:

• “Ungovernable” → “doesn’t meet defensibility standard”
• “Most” tacit knowledge → “more than organizations assume”
• Adoption metrics defined

Acknowledged gaps now addressed:

• Security considerations for universal interface
• Legal disclaimer for governance framing
• Measurement starter kit
• Obfuscation reframing

Defended:

• Strawberry experiment as a didactic demonstration
• Rhetorical choices
• Core oracle/assistant distinction

The oracle/assistant distinction remains valid with these requirements and qualifications. The regulatory direction is consistent with the framework’s predictions.

→ For the research findings, see Update for February 2026 → For the original framework, see AI as Oracle vs. Assistant