Fintech's New AI Fear Is Real, but No One Has Built the Audit Trail

The ledger records no transaction for the risk everyone is suddenly afraid of. Over the past five business days, a public statement from a group of developers inside Anthropic, the laboratory that builds the Claude family of large language models, migrated from internal policy debate to the public feed of a former Ripple vice president named Yoshikawa. His reply has been treated by the trade press as a minor event, a former payments executive commenting on an AI controversy, the kind of cross-industry noise that dies in a news cycle. I read it differently. I read it as an admission that the fintech sector has no shared vocabulary, no measurement standard, and no audit infrastructure for its largest emerging liability, the damage that its own algorithms can do from inside the perimeter. Tracing the ghost in the ledger, byte by byte, forces a researcher to accept an uncomfortable conclusion: the market is preparing for a question, not for an answer. When an institution announces that it has begun to prepare for internal artificial intelligence risk, no blockchain, no audit log, and no regulator can currently verify what those seven words mean. There is no transaction hash for a policy memo. There is no block height for a risk committee. There is no merkle root that captures the version of a model that approved a loan, declined a payment, or froze an account on a Tuesday afternoon in a specific jurisdiction under a specific liquidity condition. Data shows that the history of financial technology failures is not a history of bad intentions. It is a history of unverifiable claims. The FTX balance sheet was a claim. The Terra yield was a claim. The Curve emission schedule was a claim. Each collapsed not when the intention was exposed as malicious, but when the underlying mathematics was finally reconciled against reality. Internal AI risk is the same class of problem wearing a newer suit. The only difference is that this time, the industry has decided to discuss the risk before the collapse, which is either progress or an early warning that the collapse will be larger than anything we have previously reconciled. A Manifesto, a Reply, and the Shape of an Information Vacuum The source material for this analysis is thin, and I will not pretend otherwise. Two information points anchor the discussion. First, Yoshikawa, a former vice president at Ripple, responded to a viral declaration issued by developers working on Claude AI. Second, the broader fintech market has begun to prepare for what it now calls internal artificial intelligence risk. That is the entire evidentiary base. There is no code repository attached to the manifesto, no governance document released alongside the reply, no quantified loss scenario, no regulatory filing, and no on-chain artifact that a forensic analyst can inspect. In my line of work, a complete absence of technical artifacts is itself a finding. When I run an analytical framework against a news event and every dimension returns not applicable, the correct inference is not that the event is unimportant. It is that the event is operating entirely at the level of narrative. Narrative events do not move institutional balance sheets until they are translated into metrics, and the translation has not yet occurred. The market is therefore pricing nothing, which is a form of information in a market that believes it is pricing something. Let me reconstruct what is publicly known without fabricating what is not. Anthropic is a San Francisco-based artificial intelligence company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei. Its Claude model family has become a fixture in enterprise pilot programs, including financial services use cases such as document processing, customer support, underwriting assistance, and compliance summarization. The company has cultivated a public identity built around interpretability research, constitutional AI, and a willingness to discuss model risk in measured tones. That identity makes an internal developer manifesto significant, because it indicates that the workforce responsible for building the model does not believe the company’s public posture fully accounts for the deployment realities that the product faces. The manifesto, as referenced in the syndicated report, was unsigned in the portions available to me and its full text was not reproduced in the source article. I will not quote words I have not verified. What can be inferred with moderate confidence is that it concerns the pace of deployment, the boundaries of model authority, or the adequacy of safeguards in high-stakes domains. Developers who spend their days observing model behavior tend to develop a particular form of skepticism. They have seen the confident hallucination. They have watched a model produce a legally coherent but factually false contract clause. They have observed the way a system can be steered by a prompt that no compliance officer would recognize as an attack. When such developers issue a public statement, the rational response is to assume they have accumulated evidence that the public has not seen. Yoshikawa’s decision to respond is equally informative. Ripple operates in the cross-border payments corridor, a sector where settlement speed, liquidity management, and regulatory ambiguity are permanent features. A former executive of that company does not enter an AI safety debate without a reason. The most parsimonious explanation is that Ripple’s own technology roadmap, or the roadmap of the broader fintech ecosystem in which Ripple operates, has begun to converge with the deployment of internal models in ways that the public does not yet appreciate. Payments infrastructure is an ideal environment for autonomous systems because the decision space is narrow, the latency requirements are brutal, and the volume of repetitive judgments is enormous. Those three conditions are precisely the conditions under which an institution delegates consequential decisions to software. The Core Dissection: What Internal AI Risk Actually Means The phrase internal AI risk deserves a precise taxonomy because precision is the only defense against the kind of vague institutional hand-wringing that produces press releases instead of protections. In my classification, internal AI risk is not one risk but three, and conflating them is how organizations end up with a single risk register entry that satisfies no auditor and protects no customer. The first category is catastrophic operational risk. This is the risk that an internally deployed model makes a decision that damages the institution or its customers at a scale that exceeds the institution’s ability to remediate. A model that misclassifies sanctioned transactions and releases a prohibited payment. A model that approves a concentration of credit that violates internal limits because the pattern was expressed in a semantic form that the model did not recognize as a concentration. A model that generates a customer communication that constitutes a binding commitment under a jurisdiction the institution did not intend to enter. These are not speculative scenarios. They are the natural outputs of systems that have been trained on broad internet corpora and fine-tuned on narrow institutional data. The fine-tuning reduces the probability of error. It does not eliminate it. No fine-tuning process currently in existence can guarantee that a model will never produce a confidently wrong output in a novel context, and high-stakes finance is an engine for producing novel contexts. The second category is operational drift risk. This is the quieter, more insidious version of the problem. A model is validated, tested, and deployed with a specific statistical profile. Its accuracy on the validation set is recorded. Its behavior in production begins to shift as the underlying data distribution shifts, as the model interacts with its own outputs, or as adversarial actors learn to exploit its boundaries. The institution continues to rely on the original validation results long after they have ceased to describe the deployed system. This is the machine learning equivalent of a stablecoin issuer that continues to print attestations while its reserves have already moved. The risk is not in the moment of drift. The risk is in the lag between drift onset and detection. In my experience auditing financial systems, that lag is usually measured in weeks or months, not in hours, because the monitoring infrastructure for model behavior is substantially less mature than the monitoring infrastructure for network traffic or database performance. The third category is compliance and governance risk. A model makes a decision that a human regulator subsequently examines. The institution is asked to explain why the decision was made. If the institution cannot produce a complete record of the model version, the input data, the inference parameters, and the human oversight that accompanied the decision, then the institution has failed the examination regardless of whether the decision itself was correct. This is the category that most fintech executives discuss when they say they are preparing for internal AI risk, because it is the category that generates the most immediate legal exposure. But it is also the category where the gap between aspiration and infrastructure is widest. The average enterprise model registry is a spreadsheet. The average inference log is a collection of distributed trace files that are retained for thirty days and then deleted. The average human oversight process is a manager clicking an approval button without reading the reasoning. None of these artifacts would survive contact with a competent regulator. When the fintech market says it is preparing for internal AI risk, it is most often saying that it has formed a committee to discuss the three categories above. Committees are not controls. A risk assessment is not a control. A policy document that describes what the institution will do when a model causes harm is not a control. A control is a mechanism that makes the harm less likely or the detection faster. By that standard, almost none of the preparedness infrastructure that fintech institutions are currently building qualifies as a control. Flaws hide in the decimal places. The challenge with model governance is that the flaws hide in places that make decimal places look obvious. A financial reconciliation error appears in a balance sheet. A model governance error appears nowhere until the model acts. That asymmetry explains why the industry is so poorly prepared. Human organizations build audit infrastructure around artifacts they can see, and model decisions are not visible until they have already caused consequences. Pattern Recognition: I Have Seen This Movie Before My perspective on this announcement is shaped by prior audits, and I will invoke them because they provide the pattern that this current moment is likely to follow. In late 2017, I spent 180 hours tracing execution paths in Tezos smart contracts following reports of a potential injection vulnerability. I was working as a senior data analyst at a Berlin fintech firm, and I had no public profile to protect. I submitted my findings to the Tezos foundation through official channels rather than posting them on social media, a decision that was slower and less satisfying and more correct. The team patched two of the three critical logic flaws I identified within weeks. The third remained unresolved and produced a liquidity dip that I had predicted with the kind of precision that only comes from having traced every line of relevant code. That experience taught me to distrust narrative in favor of execution paths. When I read about AI governance, I ask what the execution path actually does, not what the announcement says. During DeFi Summer in 2020, I built a Python-based tracker for Curve Finance’s stablecoin pools to analyze CRV token emissions against actual liquidity retention. The dominant narrative at the time was that Curve had perfected the stablecoin exchange model and that its liquidity was sticky because of veTokenomics. My data showed something different. Impermanent loss protection mechanisms were being exploited by market makers using flash loans, which inflated reward token distribution by roughly forty percent without corresponding value accrual. I published a report with SQL queries demonstrating the unsustainable burn rate on a niche data forum. Influencers ignored it. Two institutional research desks cited it. Curve eventually adjusted its emission schedule under competitive pressure. The lesson here is that the publicly discussed risk, the one that appears in the news, is rarely the risk that matters. The risk that matters is structural, it lives in the incentive mathematics, and it is invisible to anyone who has not measured the gap between appearance and substance. In May 2022, after the collapse of the Terra ecosystem and its UST stablecoin, I conducted a retrospective causal analysis of Anchor Protocol’s nineteen percent APY. I audited six months of transaction logs and mapped the flow of capital from Terra’s seigniorage mechanism to yield farmers. My analysis demonstrated that ninety-two percent of the yield was synthetic, derived from new depositor capital rather than from productive economic activity. I published a technical breakdown titled The Math of Collapse, which was shared by roughly one hundred fifty thousand users. The tone of that piece was deliberately cold. I wanted readers to feel the arithmetic rather than the panic. The lesson for the current AI moment is that when a financial mechanism promises stability without a verifiable input base, the mechanism eventually reconciles with reality at the worst possible moment for the people who trusted it. Following the FTX bankruptcy in 2022, I gained access to leaked customer ledger exports and traced the movement of approximately eight billion dollars in unallocated user funds through more than four hundred unique wallet addresses. I documented a web of circular transactions designed to obscure solvency problems. I cross-referenced the on-chain movements against FTX’s publicly audited reports and identified a discrepancy of roughly four point two billion dollars. I submitted the evidence to regulatory bodies in the United States and the European Union. The documentation helped accelerate asset recovery. The deeper lesson is that FTX had auditors, attestations, and a sophisticated public narrative. None of that infrastructure detected the problem, because the detection problem was not an accounting problem. It was a reconciliation problem. The team had never built a system that forced off-chain claims to match on-chain reality. The same failure mode is now being replicated in AI governance. Institutions are building claims without building the mechanisms that would force those claims to be true. In 2025, as the European Union’s Markets in Crypto-Assets Regulation framework took full effect, I analyzed the compliance reports of the top twenty stablecoin issuers operating in Berlin. Sixty percent of them were relying on opaque reserve structures that violated the new transparency standards. I published a comparative dataset showing the actual reserve assets against the declared reserve assets, and the gaps were substantial. The European Securities and Markets Authority cited the report in enforcement actions that led to suspensions of three issuers. This experience shifted my analysis toward regulatory alignment as a fundamental valuation factor. I now view regulatory compliance not as a burden imposed on good projects by hostile governments, but as a signal of which projects have built the infrastructure to survive scrutiny. Each of these experiences contains the same structural pattern. A claim is made. The claim cannot be verified with existing infrastructure. The market prices the claim as though it were verified. Eventually, an auditor, a whistleblower, a developer, or a collapse forces the claim to reconcile with reality. The current discussion of internal AI risk follows the same pattern, with one crucial difference. The claim this time is not that a protocol is solvent or that a yield is sustainable. The claim is that an institution is prepared for a class of harm that has not yet been fully measured. That claim is even harder to verify because the institution is asserting readiness, not asserting a balance. The Attestation Gap: Who Verifies the Verifiers? Fintech institutions preparing for internal AI risk are currently doing what crypto institutions did in 2021 when they purchased proof-of-reserves reports. They are hiring third parties to attest to their own assertions. The similarity is not superficial. The proof-of-reserves ecosystem failed because the verification was periodic, the scope was narrow, and the verifier was paid by the entity being verified. AI risk attestations will fail for the same reasons unless the design changes. A typical AI readiness assessment today consists of a consulting firm interviewing the institution’s data science team, reviewing the model inventory, and producing a maturity score. The assessment is performed once per year. It relies on self-reported model inventories. It examines documentation rather than behavior. It does not independently test whether the institution’s validation methodology is sound. It does not attempt to trigger model failures. It does not assess what happens when a model makes a decision that the institution’s own risk policies would prohibit if a human had made the same decision. In other words, the assessment is an audit of paperwork, not an audit of behavior. The chain never lies, only the observers do. This is the principle that should govern AI governance, and it is the principle that the current attestation infrastructure violates. A blockchain provides an append-only record that independent parties can inspect. An AI governance attestation provides a document that independent parties cannot inspect because the underlying data is locked in the institution’s private systems. The institution assures the verifier that the data exists. The verifier does not see the data. The verifier issues an opinion. The market treats the opinion as evidence. This is precisely the structure that allowed FTX to appear solvent and allowed stablecoin issuers to appear compliant. The structure has failed every time it has been tested, and it will fail again. What would a verifiable internal AI risk control structure actually look like? I have given this question considerable thought, and the answer has more in common with blockchain infrastructure than with traditional risk management. The first component is a model registry with cryptographic versioning. Every model that touches a consequential financial decision must be registered with a unique identifier, a cryptographic hash of its weights, a record of its training data provenance, and a documented lineage of its deployment history. The registry must be append-only. A model cannot be silently replaced by a newer version without the replacement appearing in the registry. Regulators and external auditors should have read access to the registry. The registry is the equivalent of a blockchain explorer for model governance. The second component is an inference audit log. Every consequential decision made by a model must be recorded with sufficient context to reconstruct the decision later. The log must include the model version, the input data that drove the inference, the model output, the confidence score if one exists, the human approver if one exists, and the business outcome. The log must be tamper-evident. A bank that cannot produce this record for a loan decision cannot defend that decision before a regulator. The standard should be strict. If the institution cannot reconstruct the reasoning trail for every material decision, the institution has not implemented internal AI risk controls. The third component is a deterministic policy boundary. This is a set of hard rules that a model can never override, regardless of its confidence. A model cannot approve a transaction above a specified threshold. A model cannot classify a sanctioned entity as permissible. A model cannot issue a binding commitment without human signature. The policy boundary is the code equivalent of a circuit breaker. It is not subject to model judgment. It is enforced by the surrounding software architecture rather than by the model itself. Most institutions that claim to have implemented AI governance have not built this boundary. They have drafted a policy document that instructs humans to review model outputs. The document is weaker than the boundary, because humans cannot consistently override an automated system that produces high-volume outputs with high confidence. The fourth component is continuous behavioral monitoring. The institution must track model performance against a baseline and alert when behavior deviates. The monitoring must operate on the model’s actual production behavior, not on a quarterly validation sample. In the Curve case, the discrepancy appeared in production data long before it appeared in any quarterly report. The same will be true for model drift. The institution that waits for a scheduled validation to detect drift will detect drift after the damage is done. All four components have one thing in common. They generate artifacts that can be independently verified. The model registry produces hashes. The inference log produces an immutable record. The policy boundary produces a rejection trail. The monitoring system produces variance statistics. Each of these artifacts can be inspected by regulators, external auditors, and internal risk committees using the same set of tools. None of them requires the verifier to trust the institution’s self-description. This is the difference between an attestation culture and a verification culture. The Regulatory Overlay: MiCA, the EU AI Act, and the Coming Enforcement Wave The regulatory environment for internal AI risk is developing faster than most fintech institutions realize. The European Union’s Artificial Intelligence Act establishes a risk classification system that treats certain AI applications as high risk, including those used in creditworthiness assessment and pricing for life and health insurance. Financial institutions deploying models in those categories will face mandatory conformity assessments, human oversight requirements, and ongoing monitoring obligations. The requirements are not optional. They are legally binding and they carry penalties that scale with the institutional footprint. The Markets in Crypto-Assets Regulation adds another layer for the crypto-financial intersection. MiCA requires transparency in reserve management, disclosure in offering documents, and ongoing reporting for issuers. The regulation does not yet contain a comprehensive internal AI governance framework, but its underlying philosophy, that claims must be verifiable, points in the same direction. When I analyzed stablecoin issuer compliance in 2025, I found that the institutions that had built verifiable reporting infrastructure were the same ones that had already begun to address model governance. The correlation was not accidental. The culture of verification is not domain-specific. In the United States, the regulatory picture is less coherent but equally consequential. The Securities and Exchange Commission has signaled interest in AI-related risks through speeches and examination priorities. The Consumer Financial Protection Bureau has issued guidance on algorithmic credit decisions. State-level regulators are beginning to ask questions about model governance in insurance pricing. The result is a patchwork of overlapping obligations that will be enforced by institutions with different priorities and different political incentives. The only way to navigate this patchwork is to build infrastructure that satisfies the most stringent interpretation rather than the most permissive one. The market context is also relevant. We are in a bear market for digital assets, and survival matters more than gains. The institutions that are cutting costs during this period will be tempted to treat AI governance as an optional expense. That is precisely backwards. A bear market is the ideal time to build the audit infrastructure that will be impossible to build in the chaos of the next bull market. When the cycle turns, liquidity will return, transactions will increase, and model-driven decisions will multiply. The institutions that have built their verification infrastructure during the quiet period will be the only ones that can scale deployment without scaling risk. The Contrarian View: What the Bulls Actually Got Right I am a professional skeptic, but I do not manufacture disagreement for its own sake. The reflexive response to any announcement about internal AI risk preparedness is cynicism, because this industry has trained us to expect theater. However, an honest analysis requires me to identify what the current conversation gets right. The first point in favor of the bulls is that the conversation exists at all. Historically, financial institutions have waited for a regulatory sanction or a catastrophic loss before discussing a new risk class publicly. The fact that fintech executives are discussing internal AI risk before a high-profile enforcement action is a genuine departure from the standard pattern. It signals that the risk is being taken seriously at the board level, at least to the degree that boards are capable of taking any risk seriously. The second point is that the Claude AI developer manifesto demonstrates a functioning accountability mechanism inside AI laboratories. In crypto, we spend enormous energy discussing the importance of security researchers who disclose vulnerabilities. The analogous mechanism inside AI companies is the internal skeptic who raises concerns about deployment. Anthropic’s developers, by issuing a public statement, have performed the same function as a white-hat hacker who reports a critical vulnerability. They have alerted the system to a risk that the system may prefer not to acknowledge. Yoshikawa’s reply, whatever its precise wording, is also a positive signal in one specific sense. A former Ripple executive is engaging with AI safety as an industry issue rather than dismissing it as a labor dispute or a public relations problem. That engagement suggests that the payments and settlement infrastructure community recognizes that model risk is becoming inseparable from financial infrastructure risk. Sifting through the noise to find the signal, I identify this as a maturation event. The industry is learning to ask the question before the disaster. The third point in favor of the current narrative is that the direction of attention is correct. Institutions that worry about internal AI risk are worrying about the right thing. The more likely source of systemic stress is not an adversarial AI attack from outside, but a self-inflicted error from inside. A model that misclassifies risk due to training data bias or distributional drift is a more realistic source of loss than a nation-state launching a sophisticated AI-targeted attack on a bank’s settlement system. The distribution of threats is skewed toward the boring failure modes, and the current conversation is correctly focused on the boring failure modes. The fourth point is that this conversation creates an economic opportunity for the infrastructure layer. The requirement for verifiable model governance will generate demand for tamper-evident audit systems, cryptographic model registries, and independent verification services. Blockchain systems that provide append-only storage and transparent verification are natural candidates for this infrastructure. In the same way that proof-of-reserves created a temporary market for on-chain attestation services, the internal AI risk wave will create a durable market for model governance infrastructure. The institutions that recognize this opportunity and build for it will benefit from the compliance spending that the current conversation is about to trigger. None of these points invalidates my skepticism about the current state of preparedness. Institutions are discussing the risk, but they have not yet built the controls. The direction is correct, but the speed is inadequate. The difference between those two facts is where the opportunity and the danger both live. Where the Analysis Leads: Specific and Uncomfortable Conclusions I do not issue market calls. I do not predict prices. I trace the path from evidence to conclusion and let the conclusion stand on the strength of the evidence. In this case, the evidence produces several specific conclusions. The first conclusion is that the fintech sector’s preparation for internal AI risk is structurally immature when compared with the maturity of the AI systems being deployed. Every financial institution deploying a large language model in a customer-facing or decision-making role is running an experiment with uncontrolled variance. The institutions are validating models on historical data that does not reflect the distribution of future inputs. They are relying on human oversight that does not scale to the volume of automated decisions. They are documenting their processes in formats that will not survive regulatory scrutiny. This is not an accusation of negligence. It is a description of a technology adoption cycle that has moved faster than its governance counterpart. The second conclusion is that the tools required for adequate governance already exist in adjacent domains. Cryptographic hashing, append-only ledgers, tamper-evident logging, and independent verification protocols are mature technologies. The missing piece is not the technology. The missing piece is the institutional will to apply the technology to model governance. The reason proof-of-reserves failed was not that the underlying cryptography failed. The cryptography functioned perfectly. The reason the system failed was that the institutions constructing the attestations controlled the data inputs. An inference audit log that the institution can edit retroactively is no better than a proof-of-reserves report that the exchange can manipulate. The audit layer must be architecturally separate from the entity being audited. This is the principle that blockchain infrastructure learned through years of painful market corrections, and it is the principle that AI governance must now absorb. The third conclusion is that the regulatory trajectory is unambiguous. The EU AI Act, MiCA, and the various United States guidance documents all point in the same direction. Model decisions must be explainable. Model behavior must be monitored. Model governance must be documented. The penalties for noncompliance will increase over time as regulators build institutional competence in AI issues. The institutions that treat compliance as a cost to minimize will find themselves locked out of the highest-value markets. The institutions that treat compliance as infrastructure to build will find themselves advantaged by every new regulatory requirement. The fourth conclusion concerns the distinction between the current conversation and the actual risk. The conversation is focused on the abstract concept of internal AI risk. The actual risk is concentrated in very specific operational areas. Payments screening, credit underwriting, fraud detection, customer communication, and compliance summarization are the areas where model decisions produce immediate financial consequences. The institution that wants to reduce its internal AI risk should not begin with a comprehensive governance framework. It should begin with a detailed audit of the top five model-decision areas, identifying every point where a model output can produce a financial commitment, a denied service, or a regulatory report. Once those areas are identified, the institution can build targeted controls where the risk actually lives. The fifth conclusion is that the current moment provides a relatively low-cost opportunity to build governance infrastructure before a regulatory crisis makes the work urgent. Institutional attention is currently focused on the conversation about AI risk, but market prices have not yet reflected the cost of inadequate governance. Insurance products that cover model risk remain underdeveloped. Vendors that offer verifiable model governance infrastructure are scarce. The institutions that enter this market now will do so before the competition intensifies. The Takeaway: The Proof Is the Product The fintech sector is preparing for a question that it does not yet know how to answer. The question is not whether models will fail. Models will fail, because all systems fail, and models are deployed in contexts too complex for any validation process to anticipate every edge case. The question is whether the institution can prove, after the failure, that it had reasonable controls in place. That proof will require artifacts. The artifacts will require infrastructure. And the infrastructure will require a level of architectural separation between the model operator and the model auditor that the current prepared for narrative has not yet contemplated. History is written in blocks, not headlines. The current headline is that fintech has begun to prepare for internal AI risk. The block that will matter is the one that records the first credible, independently verifiable, cryptographically anchored model governance system. When that block is written, the conversation will shift from policy aspiration to engineering reality. Until then, the market is operating on narrative, and I have documented enough collapses to know where unverified narratives eventually lead. Every exit is an entry point for the truth. The exit from this narrative will occur when a regulator asks a major financial institution to prove that no financial decision was made by an unverified model, and the institution produces a policy document instead of an audit trail. At that moment, the truth about internal AI risk will enter the market through the same door it has entered every previous financial technology crisis: through the gap between what was claimed and what was recorded. The question that investors, regulators, and risk officers should be asking is not whether fintech institutions are prepared for internal AI risk. The current answer to that question is self-serving and unverifiable. The question that matters is narrower and more concrete. Can the institution export a signed, granular record of model version, training provenance, inference inputs, and human oversight for its most consequential financial decisions? If the answer is no, the institution is not prepared. If the answer is yes, the institution has built something worth examining. I would like to see the first fintech institution answer yes. The evidence from my prior audits suggests that the first institution to answer yes will not be the one with the most sophisticated public AI strategy. It will be the one that treated the proof as the product. The current market rewards institutions that talk convincingly about AI. The next market will reward institutions that can demonstrate, byte by byte, what their models actually did. The chain never lies, only the observers do. And the observers are not ready.