The Trace Went Dark
Pre-Flight Metrics Card
[Navigation Links: previous and next episode links to be inserted at publication]
One confirmed from five. Eighty per cent rejected. The external referent did its work. Now the model's internal reasoning is going underground. That does not change the result.
The arc closed. The question it left behind is bigger.
Episode 17 closed the Coding Arc with a precise empirical finding. Daniel Stenberg and the curl security team reviewed a third-party assessment of Anthropic's Claude Mythos against curl's 176,000-line C codebase. Five vulnerabilities reported; one confirmed; three false positives; one non-security bug. The 80 per cent rejection rate was not a failure of the model being assessed. It was a demonstration that an external referent with its own audit history, operating independently of the model making claims, is the structural requirement for verification. The external referent did not need access to the model's internal reasoning. It had its own evidence base.
That is the finding Arc 3 opens with.
The question Arc 3 now asks is not whether external verification is preferable to internal self-assessment. The Stenberg test answered that question empirically. The question is what it means that, in the same period, the internal reasoning trace that verification approaches have historically relied upon is disappearing. Not metaphorically disappearing. Structurally disappearing. The next generation of frontier models reasons in latent space: compressed, continuous vectors that carry no human-readable representation of the model's reasoning process.
The trace went dark. The architecture choice the V.E.R.A. Logic Engine made when it was first specified turned out to be correct for reasons that were not fully articulated at the time. This episode articulates them.
What "the trace went dark" actually means
Three independent research developments in the first half of 2026 converge on the same structural finding. Frontier model reasoning is moving from token-stream processing, where reasoning steps are encoded as readable tokens, to latent-space processing, where reasoning is encoded as continuous high-dimensional vectors. The difference is not a performance detail. It is an architectural shift that removes the audit surface that chain-of-thought inspection, trace auditing, and internal consistency checking depend upon.
According to published reports of NVIDIA's Fast-ThinkAct architecture, a 200-token reasoning plan can be compressed into 6 latent tokens. Each latent token encodes approximately 800,000 internal numbers, representing the model's high-dimensional embedding. The compression is computationally efficient for inference. It is not reversible. An external auditor cannot reconstruct the 200-token plan from the 6 latent tokens; the information is not recoverable in human-readable form.
Researchers at the University of Virginia used sparse autoencoder analysis on LLaMA-3 to identify a discrete functional feature, designated Feature #8629, that activates when the model switches from response generation to internal reasoning. When this feature is active, the model's internal computation shifts to a pattern that the analysis cannot fully decode. The feature is not visible in the token stream. From the perspective of an external monitor reading the model's outputs, there is no signal that Reasoning Mode has engaged.
Microsoft Research and the University of Pennsylvania's Multiplex Thinking paper demonstrates that frontier reasoning models hold multiple reasoning threads simultaneously in latent space rather than processing them sequentially. The reasoning is parallel by architecture. A sequential trace, even if readable, would not accurately represent what the model was computing. The trace has never been a complete record of the reasoning process; for latent-space models, it is not even a partial one.
| Development | Mechanism | V.E.R.A. Logic Engine implication |
|---|---|---|
| NVIDIA Fast-ThinkAct (Likely; primary paper pending) | 200-token reasoning plan compressed to 6 latent tokens. Each token encodes approximately 800,000 internal numbers. No human-readable trace survives the compression. | Verification approaches that audit the token-stream reasoning trace become inapplicable. Output-level verification against external facts is unaffected. |
| LLaMA-3 Feature #8629 (UVA; Likely; primary paper pending) | Discrete Reasoning Mode feature identified via sparse autoencoder analysis. Activates on multi-step prompts; structurally invisible to token-stream monitors. | The model's reasoning state is not readable from its output stream. The V.E.R.A. Logic Engine's claim-type classification (e-type vs n-type) operates on outputs, not on reasoning state. |
| Multiplex Thinking (Microsoft / UPenn; Likely; primary paper pending) | Frontier models hold multiple reasoning threads simultaneously in latent space rather than sequentially in a token stream. Superpositioned reasoning is not representable as a sequential trace. | Sequential trace inspection assumes sequential reasoning. Frontier model reasoning is structurally parallel. The V.E.R.A. Logic Engine never assumed sequential trace access. |
Table 1: Latent-space reasoning developments and their implications for the V.E.R.A. Logic Engine architecture. Sources: NVIDIA Fast-ThinkAct (primary paper to be confirmed; reported via ML research aggregators); University of Virginia sparse autoencoder analysis on LLaMA-3 (primary paper to be confirmed; Likely confidence); Microsoft / UPenn Multiplex Thinking paper (primary paper to be confirmed; Likely confidence).
Two architectural bets: what each one loses when the trace goes dark
The SInternal Framework, published on arXiv (2605.08930v1, 9 May 2026), represents the inverse architectural position to Te Pono V.E.R.A. SInternal proposes that AI models verify their own outputs using their own internal reasoning, without external reference databases or human verification. The proposed mechanism is a verification module trained alongside the base model, generating a confidence assessment of the base model's outputs from within the same training distribution.
SInternal's architectural bet is that the model's internal reasoning is the right instrument for verifying the model's outputs. When the trace goes dark, that bet loses. A verification module trained on the same distribution as the base model inherits the base model's systematic errors. A model that has learned an incorrect fact will have a verification module that is confidently wrong about the same fact. SInternal detects inconsistency, not error. If the model is consistently wrong, SInternal will confirm it.
The latent-space opacity problem compounds this. SInternal's verification module reasons about the base model's outputs; it does not have access to the base model's latent-space reasoning process any more than an external auditor does. Internal self-verification was theoretically limited before latent-space architectures became mainstream; it is structurally inoperable as a complete verification approach against Fast-ThinkAct class models.
The V.E.R.A. Logic Engine's architectural bet is different. Verification is performed against external NTP-grounded facts in the E! Corpus: claims checked against independently sourced, independently maintained reference data. The V.E.R.A. Logic Engine never inspected the model's internal reasoning trace. It does not require the trace to be readable. When the trace goes dark, the V.E.R.A. Logic Engine's verification capability is structurally unchanged.
Two boundaries must be stated explicitly before the comparison table below.
First: the V.E.R.A. Logic Engine does not verify its own reasoning. It verifies the model's outputs against external facts. This distinction is not rhetorical; it is architecturally significant. A system that verifies its own reasoning inherits the SInternal problem. The V.E.R.A. Logic Engine avoids this by grounding verification externally.
Second: the V.E.R.A. Logic Engine is not a statistical consistency framework, a behavioural monitoring framework, or a general-purpose AI safety layer. PRISM addresses statistical consistency. The Proofpoint Agent Integrity Framework addresses behavioural monitoring. These are legitimate tools for legitimate problems. The V.E.R.A. Logic Engine addresses a specific and different problem: the verification of existence and predication claims against formally structured external reference data.
| Dimension | SInternal (arXiv 2605.08930v1) | V.E.R.A. Logic Engine |
|---|---|---|
| Verification source | Internal model reasoning | External NTP-grounded facts in the E! Corpus |
| Independence from model | Zero: same model, same training distribution | Complete: external reference databases independent of the model |
| Opacity resilience | None: the verification module inherits the base model's opacity | Full: external reference facts are not opaque |
| Adversarial resilience | Low: inputs that fool the base model also fool the verifier | High: inputs that fool the model are caught against external facts |
| Detects errors or inconsistency? | Inconsistency only: a model with a systematically wrong belief will be confidently wrong, and the verifier will agree | Errors: claims checked against facts that exist independently of the model's training |
| Self-verification claim | Yes | No. The V.E.R.A. Logic Engine does not verify its own reasoning. |
Table 2: Architectural comparison, SInternal Framework (arXiv 2605.08930v1) versus V.E.R.A. Logic Engine. Source for SInternal: arXiv 2605.08930v1 (9 May 2026). This comparison presents the architectural design space; SInternal is an academic paper addressing legitimate verification goals within a different design assumption set.
External verification in practice: what the Erdős result demonstrates
In May 2026, an internal OpenAI reasoning model produced a valid counterexample to the Erdős unit-distance conjecture, a problem that had been open since 1946. Nine external mathematicians, including Fields Medallist Tim Gowers and Thomas Bloom, formally verified the proof using their own mathematical tools and their own expertise. Gowers stated he would recommend the result for publication in the Annals of Mathematics without hesitation.
The verification that made the result credible was not performed by the model that found the result. It was performed externally, by independent experts, using independent methods. The model's internal reasoning was not the verification instrument. The external mathematical community's formal verification tools were.
This is the external verification architecture operating in a domain where human formal verification is available. The V.E.R.A. Logic Engine's argument is that in domains where human formal verification is not available at scale, specifically natural language factual claims, existence assertions, and predication claims across large document sets, NTP-grounded external reference databases play the structural role that the mathematicians played in the Erdős verification. The architecture is the same: external, independent, reference-based, not reliant on the model's internal self-assessment.
The calibration matters here. The V.E.R.A. Logic Engine does not solve eighty-year-old mathematical conjectures. It verifies whether factual claims about the world are grounded in independently confirmed evidence. The connection to the Erdős result is architectural, not capability-level. The point is the structure of the verification act, not the difficulty of the problem verified.
The empirical floor: what the benchmarks tell us
The latent-space opacity development matters because hallucination rates in production-grade frontier models remain substantial. The theoretical argument for external verification would carry less weight if the models being verified were already accurate at acceptable rates. They are not.
| Benchmark / Model | Hallucination rate | Note |
|---|---|---|
| Opus 4.5 with web search (HalluHard) | ~30% | Complex multi-turn conversations requiring precise inline citations |
| Best-case (Vectara Hallucination Leaderboard) | 3.3% | Best-performing model across all leaderboard entries |
| OpenAI o3 (SimpleQA) | 33% | More than double o1's failure rate on the same benchmark |
| OpenAI o3 (PersonQA) | 51% | Optimising for deep reasoning trades off against factual recall accuracy |
Table 3: Hallucination rate benchmarks, frontier models, 2026. Sources: HalluHard benchmark 2026; Vectara Hallucination Leaderboard 2026; OpenAI o3 on SimpleQA and PersonQA (multiple ML research aggregators, 2026).
The o3 regression finding is architecturally significant for the V.E.R.A. Logic Engine's claim-type classification. OpenAI o3, optimised for deep multi-step reasoning, hallucinates at more than twice o1's rate on simple factual recall tasks. This is the jagged frontier pattern at the verification layer: the same architectural optimisation that improves performance on complex reasoning tasks degrades factual accuracy on the existence claims that the V.E.R.A. Logic Engine's e-type classification is specifically designed to catch.
The Krampitz Load Analyser's e-type versus n-type distinction was established as an NTP-grounded logical classification. The o3 data provides an empirical alignment: models optimised for n-type reasoning performance regress on e-type claim accuracy. The classification is not arbitrary; it tracks a real fault line in frontier model performance under architectural optimisation pressure.
The latent-space reasoning development also connects to Episode 7's Verification Gap concept and its empirical grounding in Episode 16. The Verification Gap describes the structural difference between what AI can discover and what humans must still independently validate. The latent-space shift widens the gap on one specific dimension: verification approaches that used the model's reasoning trace as a starting point for validation no longer have a trace to work from. The gap between discovery cost and verification cost was already large. One common starting point for bridging it has been removed.
What this means for the V.E.R.A. Logic Engine architecture
The architecture choice the V.E.R.A. Logic Engine made when it was first specified was external NTP-grounded verification rather than internal trace inspection. That choice was made because Non-Traditional Predication Theory's formal separation of predication and existence requires an external reference frame for existence verification: the E! operator cannot be evaluated by examining logical structure alone. External reference data is not a practical accommodation; it is a formal requirement.
The latent-space reasoning development provides a second, independent structural argument for the same architectural choice. Internal trace inspection, even if it had been the right design choice in 2025, is not available as a general verification approach in 2026. External NTP-grounded verification is.
Te Pono V.E.R.A.'s New Zealand Privacy Act 2020 context adds a practical dimension. New Zealand's Information Privacy Principle 3A, which commenced 1 May 2026, requires that organisations using AI to make automated decisions affecting individuals must be able to explain those decisions to affected parties. A system reasoning in latent space cannot produce a human-readable account of its reasoning process. The V.E.R.A. Logic Engine's output-level verification produces an auditable record of what was claimed and whether it was verified against external evidence. That record is available even when the model's internal reasoning trace is not. IPP 3A does not mandate any specific verification architecture; it establishes an obligation whose practical demands align with what external output-level verification provides.
Arc 3 will take this forward. Episodes 19 and onwards will work through the implementation of claim extraction, NTP classification as a check-worthiness ontology, and the practical mechanics of operating the V.E.R.A. Logic Engine against real document sets. The theoretical argument for the architecture is now fully stated. The implementation work follows.
What this means for sovereign capacity
Latent-space opacity has a defence dimension that the V.E.R.A. Logic Engine architecture answers directly. Pentagon Directive 3000.09 on autonomy in weapon systems, the International Committee of the Red Cross doctrine of meaningful human control, and Article 36 of the 1977 Additional Protocol I to the Geneva Conventions all require that states verify that an autonomous system can be operated lawfully before fielding. The United Kingdom Ministry of Defence Joint Service Publication 936 on autonomous systems governance carries the same obligation in allied procurement terms. None of these frameworks can be satisfied by inspecting a system's internal reasoning when that reasoning is encoded as continuous high-dimensional vectors with no human-readable trace. External, output-level verification against independently maintained reference data is the architecture that survives the trace going dark. The procurement question is whether allied review processes will recognise this in time.
If your organisation is procuring or designing autonomous systems, where in your verification approach does external, output-level evidence already replace internal trace inspection? Where does that gap still need closing? I would like to hear how other architects are thinking about this.
Te Pono V.E.R.A. is open-source and published under GPL-3.0 at https://github.com/andreas-linux/vera/. If you are building in the natural language processing or AI safety space and want to understand what NTP-grounded external verification looks like in practice, this is the working reference. If you are an enterprise architect or governance professional thinking about how automated decision systems will satisfy explanation obligations when the internal trace is no longer readable, the V.E.R.A. Logic Engine architecture is directly relevant to your work. Follow the series. Contribute to the project. Build on the open-source core. The verification problem does not solve itself from the inside.
Ita est momentum veritatis.
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 an enterprise architect and technology strategist based in Wellington, New Zealand, with more than thirty years of professional experience in architecture, security, and technology leadership. He holds an M.A. in Philosophy and Logic from Humboldt University of Berlin, where he studied Non-Traditional Predication Theory under Professor Horst Wessel. He is the founder of Te Pono Limited and publishes The Hamberger Report. V.E.R.A. (Verified Existence and 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.
References
[1] Stenberg, D. "The I in LLM stands for intelligence: Reflections on Mythos's curl assessment." Daniel Stenberg blog, 11 May 2026. https://daniel.haxx.se/blog/
[2] arXiv. "SInternal: A Framework for Internal Verification of Large Language Model Outputs." arXiv:2605.08930v1, 9 May 2026. https://arxiv.org/abs/2605.08930
[3] NVIDIA Research. "Fast-ThinkAct: Latent-space reasoning compression in transformer architectures." Reported via ML research aggregators, first half 2026. Primary paper pending confirmation.
[4] University of Virginia Sparse Autoencoder Lab. "Feature #8629: A reasoning-mode feature in LLaMA-3 identified via sparse autoencoder analysis." Reported via ML research aggregators, first half 2026. Primary paper pending confirmation.
[5] Microsoft Research and University of Pennsylvania. "Multiplex Thinking: Parallel reasoning threads in latent space." Reported via ML research aggregators, first half 2026. Primary paper pending confirmation.
[6] Bloom, T. and Gowers, T. "External verification of an OpenAI-generated counterexample to the Erdős unit-distance conjecture." Quanta Magazine, May 2026. https://www.quantamagazine.org/
[7] HalluHard Benchmark. "Hallucination rates in frontier reasoning models, complex multi-turn conversations." HalluHard 2026 leaderboard. https://halluhard.org/
[8] Vectara. "Hallucination Leaderboard, March 2026." Vectara HHEM-2.3, March 2026. https://github.com/vectara/hallucination-leaderboard
[9] OpenAI. "o3 system card and benchmark results, SimpleQA and PersonQA." Reported via ML research aggregators, 2026.
[10] Wessel, H. Logik. Berlin: Deutscher Verlag der Wissenschaften, 1989; second edition Logos Verlag Berlin, 1998. Foundational text for Non-Traditional Predication Theory.
[11] Krampitz, U. "Studies on the Logic of Non-Traditional Predication." Humboldt University Berlin, 1990. NTP claim-type formal foundation.
[12] Office of the Privacy Commissioner. "Information Privacy Principle 3A: Algorithmic Transparency." Privacy Act 2020 Amendment, in force 1 May 2026. https://www.privacy.org.nz/
[13] United States Department of Defense. "Directive 3000.09: Autonomy in Weapon Systems." 25 January 2023 update of 21 November 2012 issuance. https://www.esd.whs.mil/
[14] International Committee of the Red Cross. "Position on autonomous weapon systems: ICRC report on meaningful human control." ICRC, 12 May 2021. https://www.icrc.org/
[15] Protocol Additional to the Geneva Conventions of 12 August 1949, and Relating to the Protection of Victims of International Armed Conflicts (Protocol I). 8 June 1977. Article 36. International Committee of the Red Cross treaty database.
[16] United Kingdom Ministry of Defence. "JSP 936: Dependable Artificial Intelligence in Defence (Part 1: Directive)." First edition, 7 February 2024. https://www.gov.uk/
[17] Te Pono Limited. "V.E.R.A. open-source repository." GPL-3.0. https://github.com/andreas-linux/vera/