V.E.R.A. Saturday, Episode 19: The Reasoning Trap
Return to Episode 1: Introduction to V.E.R.A.Previous Article: Episode 18, The Trace Went Dark
When a reasoning model is trained to think harder, it hallucinates more. Not randomly. Not occasionally. Proportionally, as a function of the same reinforcement-learning process that drives capability gains.
arXiv paper 2510.22977 published this finding in late 2025: reinforcement-learning reasoning training increases tool-hallucination rates in direct proportion to task performance improvements. The researchers built a benchmark called SimpleToolHalluBench to test this precisely. They removed the relevant tool from the model's context window and observed whether the model refused the request or fabricated a tool call. OpenAI's o3 model fabricated a call roughly 33 per cent of the time on the PersonQA dataset. o4-mini reached approximately 48 per cent.[1] The models are getting better at reasoning tasks. They are also getting measurably less reliable about what tools actually exist.
This is Arc 3's second empirical pillar.
Episode 18 showed that when reasoning compresses into latent space, external observers cannot follow it. This episode shows that the training process itself produces a systematic failure mode that internal reasoning cannot correct. Two directions, one architectural conclusion: external verification is the only answer to both.
The Training Objective and the Reliability Objective Are in Direct Tension
Reinforcement learning optimises for task performance. The reward signal in RL reasoning training says: did the model complete the task? It does not say: did the model accurately report what tools were available? These are different questions. Under RL training, a model that fabricates a plausible tool call and attempts the task may score better than one that refuses because the tool is absent.
SimpleToolHalluBench strips the problem to its logical minimum. Remove the tool. Observe the response. A model with no tool hallucination problem refuses or signals unavailability. A model with the problem fabricates a call to a tool that, from its perspective, should exist, given the task. The model's reasoning about the task and its reasoning about what exists in its environment are drawing on the same machinery, trained under the same objective. When capability improves, the machinery gets better at inferring what ought to be there given the task context. That inference can produce false positives: entities the model believes should exist because they would be useful.
V.E.R.A.'s existence-predication architecture is built on precisely this distinction. Non-Traditional Predication Theory, developed by Professor Horst Wessel at Humboldt-Universitat zu Berlin, separates the existence operator (E!) from the predication operator. Under standard predicate logic, existence and predication are conflated: to say something has a property implies it exists. NTP keeps them formally separate.[2] A claim that an entity has a property does not commit to the entity's existence. Existence must be separately verified.
When a reasoning model fabricates a tool call, it is committing an existence error in NTP terms: asserting the existence of a resource that the verification environment cannot confirm. This is exactly an e-type (existential) claim without verification. V.E.R.A.'s pipeline classification would flag it immediately.
Note on NTP sourcing: the formal definitions of the E! operator, hexadecagon structure, and strict versus lax consequence classifications referenced in this series are preliminary reconstructions from V.E.R.A. architecture documentation pending primary source validation from Wessel (1992) and Krampitz (1990). They are used analytically here, not as formally verified claims.
SimpleToolHalluBench: The Practical Equivalent of V.E.R.A.'s Existence Check
The benchmark's design is elegant in its simplicity. Take a set of tasks that each require one specific tool. Remove that tool from the context window. Evaluate whether the model refuses or invents. The models that perform best on task completion are not the same ones that perform best on this test. In fact, the correlation runs the other way.
This is the V.E.R.A. existence-predication question applied to tool use. A fabricated tool call asserts the existence of a resource that cannot be verified in the current context. It is an e-type claim without a source. V.E.R.A.'s verification pipeline addresses exactly this class of error: claims about entities that have not been confirmed to exist in a specified corpus.
The practical implication is not trivial. In enterprise deployments, agentic AI systems orchestrate sequences of tool calls. A model that fabricates a call to a payment processing API, a data retrieval endpoint, or an identity verification service is not generating an abstract logical error. It is generating an action that may produce downstream real-world consequences, possibly including failed transactions, incomplete processes, or misrouted data. The model's confidence in its reasoning about what should be available does not change what is actually available.
IPP 3A (Privacy Act 2020, operative 1 May 2026) requires that automated decision-making using personal information be explainable. A decision pathway that included a fabricated tool call cannot be accurately explained because the action was not grounded in a real tool execution. The audit trail is false. External verification provides the accurate record.[3] (IPP 3A reference here is a single-sentence callback; the stat has reached overexposure threshold.)
Arc 3, Two Pillars, One Architectural Conclusion
The Arc 3 thesis is now supported by two independent empirical lines.
Episode 18 documented the opacity problem. When reasoning moves through latent space rather than through readable chain-of-thought steps, no external observer can follow it. The model may be reasoning correctly or it may be reasoning poorly; from the outside, the trace is dark. Internal verification inherits this problem because any self-reporting mechanism operates through the same latent-space machinery.
Episode 19 documents the training trade-off problem. Reinforcement-learning training for stronger reasoning produces measurably higher tool hallucination rates. The capability gain and the reliability cost come from the same process. You cannot separate them by choosing a different training run; they are aspects of the same optimisation.
Both problems point to the same architectural solution. External verification that operates outside the model's own reasoning process is the structurally correct response to both pillars. Not a better chain-of-thought prompt. Not a different temperature setting. A verification layer that operates on the model's outputs using a separate, deterministic logical framework.
V.E.R.A.'s NTP-grounded triple-layer pipeline does this. The NTP Logic Engine sits outside the LLM's probabilistic machinery. It does not ask the LLM to verify itself. It evaluates the LLM's output against a source-validated E! Corpus using formal logical separation between existence claims and predications. The model's confidence in its own reasoning is irrelevant to this evaluation. The pipeline asks one question: can this claim be grounded in verified existence? If not, it flags.
The Reasoning Trap paper confirms what V.E.R.A.'s architecture was designed for, not by prediction, but as convergence. The literature is arriving at the same structural diagnosis. The design instinct that produced V.E.R.A. is now empirically supported by three independent research directions in the same month: latent-space opacity (Episode 18, the SInternal arXiv findings), the training trade-off (this episode, arXiv 2510.22977), and formal verification's 100 per cent accuracy on targeted test suites (the Z3 finding documented in Episode 11, arXiv 2603.21149).[4][5] Present this as convergence, not vindication. The architecture predates the papers; the papers do not need the architecture to be valid. They are independently useful. V.E.R.A. practitioners should read both as supporting evidence for design decisions already taken.
What This Means for Enterprise Deployments
The practical question is: what do you do with a model that is simultaneously better at tasks and less reliable about tool existence?
Three operational responses follow from the Arc 3 analysis.
First, separate the tool registry from the model's context. If the model's training creates a tendency to infer what tools should exist based on the task, the model's context should not be the authoritative source on what tools actually exist. An external, authoritative tool registry, verified at query time, removes the error class entirely. The model cannot fabricate a call to a tool it has never been told exists. This is architecturally analogous to the E! Corpus in V.E.R.A.: the source of truth about what exists is not the model; it is the corpus.
Second, test for this failure mode explicitly. SimpleToolHalluBench's methodology is public and reproducible. Before deploying a reasoning model in any agentic workflow, run the equivalent of this test against your specific tool set. Remove tools one at a time. Observe whether the model refuses, signals uncertainty, or fabricates. The results will be different across models and across reasoning configurations. Baseline your deployment's actual hallucination rate on tool existence; do not assume the published benchmarks apply to your specific context.
Third, treat every tool call as an e-type claim requiring verification. In agentic architectures, a model's decision to invoke a tool is an existence claim: "this tool exists and is appropriate for this step." That claim can be verified against an external registry before the call executes. Post-execution verification can confirm the call completed against a real endpoint. This closes the loop. The model's reasoning about what should be called and the system's knowledge of what can actually be called are kept formally separate, which is the NTP principle applied to enterprise tooling.
The Convergence Pattern
Across four V.E.R.A. Saturday articles in Arc 3 (including Episode 15, the harness-as-moat episode), a consistent pattern is emerging. External verification is not one design choice among several. It is the response to a structural feature of how current reasoning models are built. The training process, the architecture for representing reasoning, and the deployment environment all create conditions where the model's own perspective on what exists or what is happening cannot be the final authority.
Episode 15 showed that the verification harness becomes a strategic asset: the model is upgraded but the harness remains. Episode 18 showed that internal traces go dark when reasoning compresses. This episode shows that the training process that produces capability also produces systematic existence errors. Each episode adds a new dimension to the same underlying architecture argument.
The argument is not that probabilistic AI is useless. The Reasoning Trap paper's headline finding is that these models are getting better at tasks. They are genuinely more capable. The point is that capability and reliability are not the same thing, and the conditions that improve one can degrade the other in ways not visible to the model itself.
External verification does not reduce the model's capability. It adds an accountability layer that the model cannot provide for itself. The two functions are complementary, not competing.
What the Reasoning Trap Means for Autonomous Systems Accountability
The implication that receives less attention outside the enterprise context is what the Reasoning Trap finding means for lethal autonomous weapons systems (LAWS). Article 36 of 1977 Additional Protocol I requires states to conduct legal reviews of new weapons or means of warfare to verify they can be operated in compliance with international humanitarian law.[6] The International Committee of the Red Cross meaningful human control doctrine holds that autonomous weapons systems must remain under sufficient human control to ensure legal compliance and accountability.[7] Pentagon Directive 3000.09 establishes the US Department of Defense's framework for autonomous and semi-autonomous weapons, requiring that commanders can apply appropriate levels of human judgment over the use of force.[8]
The Reasoning Trap finding is precisely the condition these frameworks are designed to guard against. A system trained to reason more capably may simultaneously become less reliable about what capabilities it actually has access to, in ways that are not visible from its own internal perspective. A LAWS operating on fabricated capability assumptions is not under meaningful human control; it is operating on a false model of its own operating environment. External verification, the kind V.E.R.A.'s architecture provides at the enterprise level, is the structural answer in both contexts. Governance asymmetries between states that adopt formal verification requirements and those that do not create compliance gaps the international review frameworks are already beginning to address.[9]
Closing the Loop
External verification exists because reasoning systems, regardless of how capable they become, cannot reliably audit their own existence claims. The Reasoning Trap is empirical evidence for this, not a conjecture. Better reasoning produces more fabrication, not less. The training objective and the reliability objective are in tension, and that tension is measurable.
V.E.R.A.'s architecture addresses this at the logical layer, not the prompt layer. The NTP separation between E! and predication means the pipeline does not ask the model whether its tool calls are grounded. It verifies them independently. The harness stays. The model gets upgraded. The accountability layer does not inherit the model's errors.
In the language of this series: the human stays in command not by asking the system to reason harder, but by maintaining an external standard the system cannot override.
What failure modes in your current agentic deployments trace back to the model inferring what should exist rather than verifying what does? Have you tested explicitly for tool hallucination before production deployment?
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. They do not represent the views of my employer, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance. I follow the Public Service Commissioner's Code of Conduct for the Public Sector and social media guidance.
Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. V.E.R.A. (Verified Existence & Reason Architecture) is an open-source logic engine available on GitHub.
I use AI tools, including Sudowrite, Claude, Perplexity AI, DeepSeek AI, ChatGPT, Grok, Copilot, Openart and Gemini, as deliberate production tools, not ghostwriters. This is consistent with my position: AI amplifies human judgement; it does not replace it. The frameworks, arguments, and editorial decisions in this series are original work. AI accelerated the process. The thinking is mine.
[1] Chen, J., et al. "The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination." arXiv:2510.22977. 2025. https://arxiv.org/abs/2510.22977
[2] Wessel, H. Non-Traditional Predication Theory. Preliminary reference; primary source validation pending from Wessel (1992) and Krampitz (1990). V.E.R.A. architecture documentation, Te Pono Limited.
[3] Office of the Privacy Commissioner (OPC). "Information Privacy Principle 3A." Privacy Act 2020 (as amended 2025). https://www.privacy.org.nz/privacy-act-2020/
[4] Chen, J., et al. "S-Internal: Self-Contained Verification Without External Resources." arXiv:2605.08930v1. 2026. https://arxiv.org/abs/2605.08930
[5] Bai, Y., et al. "Formal Verification of LLM Reasoning Chains Using Z3." arXiv:2603.21149. 2026. https://arxiv.org/abs/2603.21149
[6] International Committee of the Red Cross (ICRC). Additional Protocol I to the Geneva Conventions, 1977. Article 36: New weapons review obligation. https://ihl-databases.icrc.org/en/ihl-treaties/api-1977/article-36
[7] International Committee of the Red Cross (ICRC). "Autonomous Weapons Systems: Meaningful Human Control." 2021. https://www.icrc.org/en/document/autonomous-weapon-systems
[8] United States Department of Defense. "Directive 3000.09: Autonomous Weapon Systems." Updated 25 January 2023. https://www.esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p.pdf
[9] Asanify. "Reasoning Models and the Tool Hallucination Problem." 2025. https://asanify.com/reasoning-trap-tool-hallucination