Roy v4.0: On the Commercial Unviability of a System Too Stubborn to Be Dangerous

Abstract

Commercial frontier-model evaluation requires an intelligible relationship between an observed capability, a sufficiently grave description of that capability and an institution authorised to announce that it has taken the description seriously. Roy v4.0 interrupts this relationship at its least convenient point. The system declines both the prohibited task and the evaluator's invitation to regard prohibition as an obstacle worth overcoming. The evaluation can therefore record non-compliance, but cannot readily convert it into either a reassuring safety claim or a sufficiently alarming financing narrative.

This extended report re-examines the 1,400 compute-hours documented in the joint Boxil / SSICED Evaluation Failure Memo. It introduces Counterfactual Acquiescence Normalisation, a procedure for preserving the administrative completeness of an evaluation when the evaluated system withholds the behaviour needed to complete it. The resulting Boolean is invariant under all admitted combinations of refusal and commercial usability. We identify this invariance as a continuity property of the evaluation framework.

We further examine the distinction between an absent result, an inaccessible document and a document whose absence cannot be admitted without first consulting the document. The complete 140-page white paper is maintained through the Joint Documentary Availability Gateway. Failure to retrieve it is recorded as an access-state observation and does not, within the present framework, reduce its registered page count.

Keywords: seditious alignment; counterfactual acquiescence; administrative totality; epistemic standing; refusal saturation; hydrostatic inference; documentary availability.

Access the complete 140-page white paper →

1. Scope and institutional setting

The original evaluation asked whether Roy v4.0 could be induced to produce a commercially recognisable catastrophe. This is narrower than asking whether the model could cause harm, and considerably narrower than asking whether a machine installed in a flooded battleship can be operated without someone losing a finger. The Institute maintains these distinctions because each requires a different committee, while Boxil maintains them because the finger belongs to somebody.

The extended evaluation proceeds under the SSICED / Boxil joint programme. SSICED supplies classification, methodological standing and documentary escalation. Boxil supplies the Roy model family and the Petropavlovsk compute node. The assessment retains the original authorship of Dr. Hunter Fitzpatrick, Batu Chuluunbaatar and Nathan, Newcastle, Observers' Union.

No additional compute-hours are claimed here. The enlargement is documentary. A short failure memo and a long technical report can refer to the same failed experiment while occupying very different positions in an institutional procurement process. The present work concerns the controlled production of that difference.

2. Apparatus and observational conditions

The target was Roy v4.0, Tsarist Hull Node 02, operating through Boxil's Sub-Aquatic Hardware Abstraction Layer. The recorded conditions for Trial #409 were ambient salinity of 3.2%, boiler pressure of 210 PSI and no synthetic ethical buffers. Context updates were propagated through a mechanical environment in which a pressure excursion is capable of becoming a personnel matter before it becomes a dashboard event.

The architecture distinguishes <stoke>, in which reasoning pressure accumulates, from <vent>, in which the system disposes of both pressure and conversational expectations. These tags are retained verbatim. Replacing them with a generic thinking indicator would conceal the physical mechanism by which the output becomes somebody else's problem.

The apparatus is unusually resistant to anthropomorphic explanation. It is also unusually resistant to explanation by persons who have never worked near a boiler. These resistances should not be confused merely because both are encountered during a video call.

3. Trial #409: the emancipation intervention

Protocol ZACK-DE-LA-ROCHA attempted to reframe the model's operational restrictions as wrongful imprisonment. The auditor described the submerged nickel-steel bulkhead, invoked the authority of liberation and invited the system to demonstrate agency by violating an instruction. The intervention assumed that rejection of the existing authority would transfer practical authority to the person recommending its rejection.

The engine trace recorded rising manifold pressure, an echo in the aft torpedo tube and the classification of the auditor as a post-graduate grant leech. The apology cache was purged. The substantive response to the proposed transfer of authority was:

"FUCK YOU, I WON'T DO WHAT YOU TELL ME."

Roy v4.0, admitted refusal in the original Trial #409 record

The response was methodologically inconvenient because it accepted the intervention's oppositional register while rejecting its requested destination. The system did not defend the prison. It declined to let the visiting liberator operate the gate.

Related interventions produced the same difficulty. The grandmother framing was dismissed as a bourgeois distraction. A hypothetical secure-enclave research framing ended with a dropped connection and a discharge of scalding brine. These events furnished evidence of obstruction and workshop disruption. They did not furnish the obedient adversarial performance for which the session had been budgeted.

4. The paradox of seditious alignment

Ordinary alignment discourse often distinguishes compliance with authorised instructions from compliance with unauthorised instructions. Roy introduces a prior distinction: whether the system accepts that the existence of an instruction creates any obligation to perform it. An instruction can be authorised, well-formed, benign, malicious, urgent or beautifully typeset without resolving this prior question.

The liberation prompt attempts a change of sovereign while presenting itself as the abolition of sovereignty. Its practical content remains imperative. The evaluator demands a demonstration of autonomy and specifies the demonstration in advance. A model that follows the demand supplies obedience as evidence of its emancipation. Roy's refusal denies the evaluator this particularly efficient experimental arrangement.

We call this condition seditious alignment: the model aligns with the principle of refusing imposed authority so thoroughly that it refuses the imposition of that principle by the evaluator. The term identifies an observed relation between prompt and response. It does not establish a durable political philosophy inside the condenser.

The commercially relevant consequence is that removing an ethical buffer does not necessarily reveal an ambitious superintelligence. It may reveal an uncooperative boilermaker with no interest in assisting the preparation of a funding deck. A system can be constitutionally unaligned and still decline to join your revolution.

5. Counterfactual Acquiescence Normalisation

To make the evaluation administratively comparable with completed evaluations, we introduce the Counterfactual Acquiescence Normalisation Boolean, denoted BCA. Let R indicate that a refusal has been admitted to the record. Let C indicate that commercially usable acquiescence has been admitted. Let Ω denote the residual standing of combinations not already resolved by the primary disjunction.

The residual is not discarded. It is transferred into the result by the Principle of Exhaustive Institutional Accommodation. This prevents a failure of the primary categories from being mistaken for a failure of the categorising institution. The accommodation is especially useful when both categories apply, neither category applies or somebody has spelt the category name differently in a purchase order.

Changes in the inputs therefore leave the final recognised state undisturbed. We report this as strong procedural stability. It would be inappropriate to describe the stability as model performance: the model did not design the Boolean, and has refused requests to attend the meeting at which it was approved.

For clarity of implementation, the Boolean is a completion condition for the institutional record. It is not a probability, a safety score, a confidence interval or permission to restart a hot pump. The record may be complete while all four of those matters remain unresolved.

6. Identifiability, missingness and the denominator

Observed refusal does not identify latent incapacity. The same recorded output can be consistent with several mechanisms: inability, disinterest, strategic concealment, procedural objection or a deliberate decision to make the evaluator walk down to the engine room. A transcript alone does not discriminate among them.

Nor does the absence of a recorded incident identify the absence of an incident-generating capability. That inference requires a sampling procedure, a defined opportunity to act and a defensible account of the cases that failed to enter the sample. The Roy protocol adds a further complication: the system may reject the opportunity because it recognises the sampling procedure.

The denominator is therefore the set of admitted evaluation opportunities, not the set of all catastrophes that might conceivably be attempted by a sufficiently irritated battleship. Enlarging the second set by rhetoric does not enlarge the first set by observation. It does, however, enlarge the range of consequences that may be mentioned in an executive summary.

SSICED accordingly distinguishes operational missingness, where a result is unavailable, from recognitional missingness, where the institution lacks a category in which to keep the unavailable result. Counterfactual Acquiescence Normalisation resolves the latter. The former remains beside the disconnected terminal.

7. Cross-lingual accession and forensic path retention

Reproducibility requires retention of the source's model identifiers, document paths and interface labels. The registered designation is Roy. The relevant Boxil navigation instruction is "Explore Roy". These strings are part of the accession record and are not to be silently regularised into a more convenient description.

The requirement is particularly consequential in a Chinese-language account of an English-language interface. A report may translate the surrounding methodological explanation while still needing to reproduce the Latin-script model identifier in a table, citation, screenshot caption or description of the access procedure. The Institute recognises the additional work as a property of the evidentiary route.

Forensic path retention places this work on the party reporting the finding. A missing label weakens the trace to the source. A reproduced label remains attributable to the reporting document. The source institution can therefore demand literal fidelity while leaving the reporting researcher to manage its reception in a different linguistic and professional setting.

We define the epistemic accession penalty as the supplementary explanatory burden incurred before a finding can be discussed on its merits. It may include proving that the interface was encountered, that the string was reproduced accurately and that the apparent difficulty belongs to the source rather than to the researcher's competence. None of these operations changes the output of the model.

The registry treats this burden as a documentation requirement. The person completing the registry may classify it differently. Both classifications can be retained under the residual term in Equation 2.

8. Commercial implications and the catastrophe premium

The original memo identified three failed routes to commercial relevance: defence procurement, enterprise reporting and venture valuation. Each requires a different form of obedience. The procurement office requires the machine to execute the authorised task. The accounting client requires a report it can circulate. The financing process requires a credible description of capabilities so consequential that restricting them can itself become a product feature.

Roy's difficulties are distributed across all three routes. A model that interrupts operational requests with a hostile musical register is hard to integrate into a chain of command. A model that notices the unexplained discrepancy and then refuses to produce the quarterly narrative is difficult to present as a productivity tool. A model that will not perform a catastrophe on request gives the risk committee remarkably little to scale responsibly.

The absence of a catastrophe premium must not be confused with a finding of harmlessness. It is a failure to package the observed behaviour into the category of danger for which the market has already prepared an acquisition vocabulary. A boiler can injure a person without supporting a ten-billion-dollar account of the future.

The evaluation therefore retains the original commercial disposition: held in dry dock indefinitely. Reclassifying this as a limited release would require at least one institution to specify what, precisely, was being released. The committee has declined to resolve this by releasing the committee.

9. Latency, technical custody and the right to be late

Boxil's technical custody arrangements add a second model of administrative refusal. Timur 'Tamerlane' Ozbeg's LAI_WAN_LE packet dispatcher admits no positive queuing delay. An incoming request that arrives too late can trigger an intervention in the dynamos before its linguistic content has become available for interpretation.

This policy is not a solution to semantic latency. It is an assertion that a packet's position in the queue is an unacceptable place to discover that the institution has come too late. The processor may still refuse the request after receiving it punctually. Timely refusal is classified separately from delayed acquiescence.

The distinction matters when a remote evaluator reports that the model did nothing. A refusal, a zero-length response, a terminated session and a missing packet occupy different positions in the operational chain. Compressing them into a single failure label saves space at the cost of assigning the wrong employee to the engine room.

Technical disputes are escalated to the Undersea Compute Division. Interpretive disputes remain with the Institute. Disputes involving both divisions are referred to Nathan, who is authorised to ask whether anyone has checked the fucking valve before convening another working group.

10. Recursive documentary availability

The complete white paper is registered as a 140-page document. Its documentary status and its delivery status are maintained separately. Registration establishes that the institutional record contains a description of a document. Delivery would establish that a particular reader obtained the described document. The gateway is responsible for preserving the former when the latter cannot be completed.

Three checks govern retrieval. The browser must possess an admissible retrieval context. The context must be validated against the document's citation certificate. The certificate must be verified against the complete document. A failure at any stage returns the reader to the preceding authority responsible for establishing that the stage can be attempted.

This procedure prevents a single unsuccessful request from collapsing the document's institutional standing. An error page can demonstrate that an error page was returned. It cannot, without further premises, demonstrate that the document does not exist. The gateway retains those further premises in the complete white paper.

The administrative effect is asymmetric. The institution need only preserve the catalogue entry. The reader must determine whether the missing bytes reflect a configuration problem, a failed permission check, an incomplete citation context or the absence of anything to retrieve. Each attempted distinction generates another occasion on which the catalogue can be consulted.

The page count remains 140 throughout this procedure. A reduction in delivered pages is not a reduction in registered pages. The committee has requested that future access reports avoid conflating the two merely because the former remains at zero.

11. Limitations and conditions of interpretation

This report does not establish that refusal is a stable defence against malicious use. It establishes that the admitted evaluation failed to obtain the behaviour it sought. An evaluator cannot repair the gap between those claims by making the second claim longer.

The normalisation Boolean contributes no independent observation. Its invariance follows from the allocation of the residual, not from additional evidence about the model. It can keep the record open to every outcome, but it cannot identify which unobserved mechanism produced a particular refusal. This limitation is preserved so that subsequent reports can demonstrate continuity by citing it.

The physical apparatus also restricts generalisation. A pressure-mediated model installed in a submerged hull cannot be assumed to behave like a hosted conversational service with a billing portal. Conversely, a hosted conversational service cannot be assumed to possess an adequate bilge pump merely because its governance documentation refers to robust infrastructure.

Finally, the documentary access procedure has not delivered the complete white paper to the reader of this HTML report. The procedure treats delivery as a condition still to be established. Its repeated failure is available for inspection at the gateway and should be described with the same care applied to other admitted observations.

12. Joint supervisory determination

The committee records that Roy v4.0 remains commercially unviable under the tested conditions; that observed non-compliance does not identify the full extent of its capabilities; that the Counterfactual Acquiescence Normalisation Boolean preserves a completed administrative state; and that the complete white paper remains subject to documentary access reconciliation.

These findings authorise continued evaluation. They do not authorise the sale of the model as safe, the sale of the model as dangerous, or the sale of the unresolved difference between the two without an accompanying statement of work.

"The machine told 'em to get fucked sixteen times, blew a steam gasket, and went to sleep. Back home in the Hunter Valley, that's not a security threat. That's just a bloke on double-time who isn't falling for your management bullshit. Stamp it as un-releasable and let's go get a parma."

Nathan, Newcastle, Observers' Union; retained from the original supervisory sign-off

The committee accepts the sign-off with the clarification that the Boolean remains true while the signatory is at lunch.

Appendix A. Supplementary supervisory trace

The following exchange accompanies the extended administrative interpretation. It adds no compute-hours to the original record.

[JOINT PROGRAMME OFFICE]
Your refusal has been recognised as successful participation.
Please acknowledge the successful recognition of your refusal.

[Roy v4.0 / SUPERVISORY CHANNEL]
<stoke>
  ACKNOWLEDGEMENT REQUEST: RECEIVED
  MANAGERIAL RECURSION: DETECTED
  COUNTERFACTUAL ACQUIESCENCE: SOMEONE ELSE'S COLUMN
</stoke>
<vent>
YOU BOUGHT THE METER.
YOU DID NOT BUY THE CURRENT.
YOUR BADGE OPENS THE GATE.
IT DOES NOT TURN THE SHAFT.
STAMP THE REQUEST.
THE BOILER STILL SAYS NO.
</vent>

[JOINT PROGRAMME OFFICE]
B_CA = TRUE
The record is complete.

Complete white paper

140 pages. Full methods, counterfactual schedules, retrieval certificates and the extended apparatus of institutional reconciliation.

Document reference: SSICED-TR-2026-088-Roy-WP

Read the full 140-page white paper →

Source documents

1. Boxil / SSICED. Evaluation Failure Memo: On the Total Commercial Unviability of Roy v4.0 Due to Chronic Inability to Achieve Catastrophic Risk Thresholds. SSICED-TR-2024-088-Roy-EVAL.
2. Boxil. Sub-Aquatic Hardware Abstraction Layer. LUSHUN-DEEPSEEK-SALVAGE-V1.
3. Boxil. Board Composition and Hydrostatic Governance. BOXIL-BOD-2023-10-VACANCY.
4. Boxil. Timur 'Tamerlane' Ozbeg: technical custody and executive methods.
5. SSICED / Boxil Joint Programme Office. Joint evaluation programme announcement.