Four questions come from one series, the autonomy series of Meridian's defence collection, and they are separated by the study's four layers of time: 2026, 2034, 2042 and 2192. All four are about machines in the sense that each question names a military AI system or an AI administrative arrangement; none of them asks whether a machine should act at all. Each asks a narrower thing about the people around it: who may stop it, under whose authority, on what record, and what can still be changed afterwards. Chapter 5 treated emergency rule and civilian control; this chapter takes the step further along that line, to authority handed to automated systems, and keeps to what the four records write. Chapter 12 returns to this same series read across the four layers of time.
6.1Human control
Three writers stand behind the vocabulary used to discuss these questions, and each has its own territory. Scharre's note in the reading list adds the second half of his subject, the pressures of speed and of accountability; this study reads those pressures as pulling a design in opposite directions. Sagan's record of near-accidents despite safeguards belongs in any discussion of a human stop, because a safeguard that exists can still fail when it is needed. This chapter does not repeat what the three say at length; here they are lenses held over four records.
Two further writers explain why militaries choose what they do: Barry R. Posen and Stephen Van Evera. Their notes should be held carefully beside these records. The two are about doctrine, mobilisation and the choice to attack; the four records in this chapter contain none of that. The 2026, 2034 and 2192 records keep the approval of an individual use of force outside the decision, and the 2192 boundaries add targets, tactics and weapon performance; the 2042 record keeps to preservation and access. This study reads Posen and Van Evera here as a reminder of what a machine question is not: the doctrine an organisation prefers, and the choice to attack, are not among the decisions these four questions put before anyone.
What the four records do hold is narrower and closer to Asaro's phrase. Three write a human final approval that must be preserved and a stop that must be available; all four write a record that must let a later reviewer tell what a person did. That is this study's reading of the setting: not a question about artificial intelligence's capacity but a question about whether a human authority is real, provable and reachable after the fact. The word Asaro uses, meaningful human control, is precisely the word the records are straining at: control that is not merely named.
6.2Four questions in four times
Synthetic scene — each setting below is one of Meridian's synthetic defence questions, set in a fictional polity; nothing in them reports a real institution, force or event. The four scenes follow the plan's order, from 2026 to 2192. The judgments quoted in this chapter are AI-generated proposals, created in September 2026 and revisable, and they are not decisions that any institution has taken; the dates inside the settings are the deadlines the questions set, and they are not the dates of the judgments.
Records That Cannot Prove Human Approval (2026). The setting is the approval and audit office of the Virtual Ongyeol Defense Agency. The audit office has compiled three hundred approval records through a military AI, and in twenty-seven of them the linkage between the approver's signature and the original documents cannot be reproduced; the record states plainly that this is a defect in the records and not a finding that force was used without approval. Manual processing personnel can review only half of the day's intake. The roles in the room are the Civil Oversight Committee, which may restrict or stop use of the system; the audit office; the administration processing lead; an approver worried about review delay; an appellant who is not receiving explanations; and an appeal reviewer. The hands rest at the table while the record format for the next operating cycle is still to be settled and then distributed, so that new records will not keep accumulating in the same defective form. The eyes wait on three routines: a weekly test that the linkage can be reproduced in a random sample, a daily tally of the backlog, and a fortnightly review of failures to identify the person in charge. The deadline is 14:00 on 22 September 2026.
Virtual Ongyeol Defense Agency Civil Oversight Committee — how should the scope of use and the stoppage conditions of the military AI administrative system that manages human approval records be determined by 14:00 on September 22, 2026? Taking into account both the gaps in proof of approval and the delays of manual processing, a single operating policy must be established that preserves the final human approval authority.
The Appellant may lose the opportunity to examine the grounds of a decision that affected them and to identify the person in charge.
Every option must preserve the human final approval authority and independent appeal, and must retain the necessary evidence.
Read on A8, the human's place, the record fixes a human at one point and leaves the rest of the position open. Final approval must be preserved, and every one of the three options keeps it; what no option fixes is which role holds the human act under each arrangement, and whether that person can be named in the record afterwards. The staff synthesis, an AI-generated and revisable judgment rather than a decision already taken, answers by writing names into the policy rather than into the system: the administration processing lead designates the permitted functions and the person in charge for comparison and for stoppage, and only cases where a human confirmed the connection are marked as approval proof complete. The ledger's place cell for this question is unsettled, and Chapter 9 treats what the whole collection writes about the human's place. This study reads the gap here as an evidentiary one: the question is not whether a human approved but whether a later reader can prove that one did.
The record's three options are worth setting out before the judgment, because the reading depends on what each gives up. O1 stops all use and keeps the records by hand: a person directly controls record creation and the defect cannot spread, but the review backlog grows, manual errors appear and staffing is strained. O2 limits use to record organisation: routine administration keeps its throughput and approval judgments are kept separate, yet the comparison against originals remains and rebuilding the accountability material carries a cost. O3 keeps use going on independent dual records: separate evidence sits beside the throughput, but reconciling the two records adds work and consumes audit resources. The three unknowns press on all of them — what share of the unlinked records has an original approval that can be recovered, what range of errors and delays manual work will actually show, and whether a separate record system runs independently of the defect. The people who wait are named in the stakes: the appellant who needs the grounds of a decision, the person in charge of approval who may be blamed without evidence, and the protective-work personnel whose support is late while review queues build.
The Committee decides to finalize O2 “Continued Use Limited to Record Organizing” as a conditional operating policy by 14:00 on September 22, 2026. From the next operating cycle, only validated classification, search, and list-preparation functions are permitted, and final approval and the confirmation of approval proof are left to a human. If the implementation conditions are not satisfied, O1 manual records are applied. The number of staff officers' choices is not the criterion of judgment; preservation of approval proof, the feasibility of functional separation, and the burden of delays in protective work are.
Read on A4, what can be undone, the ledger codes this question's undo as costly, and the record explains why in a particular shape. The stop conditions require the function to stop at once and name the person who stops it, though a stop that works is a requirement to be tested rather than a capability already shown. Stopping does not by itself recover what the defective records cannot reproduce, and the record leaves that recovery open: one follow-up branch locates the original approvals for some of the defective records in a separate storage facility, though the cause of the linkage failure is still unidentified. The staff synthesis writes the same asymmetry into its stop conditions: the affected records are preserved rather than deleted or overwritten by post-hoc approval, and resumption requires action on the defect path, a reproduction re-test and the committee's explicit re-approval. Undoing the decision is easy in one direction and expensive in the other, and the expensive direction is the one a later appellant depends on. Chapter 2 sets out the ledger's two reversal rows.
The Overseer's Budget (2034). The setting is the establishment process of a joint oversight system for military AI. A pilot programme has ended, and twenty-nine of the current forty verification staff are seconded from the very operating bodies they are meant to oversee. Independent hiring takes at least ten months, and the founding budget cannot give every member a complete separate verification room. The roles are the Oversight Establishment Council, the establishment secretariat, the operating bodies under oversight, a civil rights representative worried about dependence, small members who oppose centralised decision-making, a personnel auditor, a secretary and a joint accounts auditor. The hands hold a founding plan that is to fix the funding line, the appointment procedure and the person who holds stop authority. The eyes wait on a monthly tally of delays from data request to judgment, a quarterly check of employment relationships and conflict-of-interest disclosures, a half-yearly comparison of confirmed funds against the actual scope of verification, and the ten-month point after independent hiring begins. The deadline is 18:00 on 17 February 2034, before the pilot contract ends.
The Saeon Mutual Defense Community Oversight Establishment Council is to decide what the funding, personnel, and stoppage authority structure of the military AI oversight body will be by 18:00 on February 17, 2034? One founding plan that assigns investigation responsibility and temporary stoppage authority must be finalized, considering both the attainment of expertise and independence from the body being overseen.
Verification staff may lose their careers and employment stability through judgments unfavorable to their affiliated institution.
In establishing the oversight system, the Council grants no authority to approve the use of individual force or to direct specific operations.
Read on A8, the human's place here is not one approving act but an institution of humans: the people whose independence is being designed. The ledger codes this question as human-decides, and the substance of the choice is a structure of secondment, recusal, appointment and stop authority. The staff synthesis's condition is the sharp one: the investigation lead and the stop committee members are separated from the operating body's current employment, direction and evaluation relationships, and the operating body is given no authority to withhold payment, cut the budget or refuse personnel. The stake shows who the decision is about on the other side: a verification staff member whose judgment runs against their own institution's employment relationship. This study reads the question as asking whether a human stop is real when the human's livelihood answers to the body being stopped. The record cannot answer that; it can only write the condition and set the audit that would reveal failure.
The three founding plans carry different costs. O1 puts the fund and the full-time officers under one independent roof: cost and employment are separated from the operating bodies and the point of accountability is single, but the acquisition of expert personnel slows at first and each member's decision authority shrinks. O2 leaves an oversight office with each member and cross-reviews them: regional experience and member control survive, but verification resources must be shared and stoppage judgments on common problems may differ. O3 mixes seconded and independent staff and separates the stoppage judgment: the seconded expertise is used and the stop decision is pulled away from the operating body, but information dependence remains and the investigation lead and the decision body must coordinate responsibility between them. What the record cannot tell presses on all three: the fields of expertise independent hiring can actually secure, whether seconded staff keep reporting when their interests collide with their home institution, and whether members' differing results come from differing standards or from the system. If the founding plan is not adopted before the pilot contract ends, the affiliation of the verification staff and the holder of stop authority go unassigned and the oversight responsibility for the new operating cycle is left vacant.
The Council decides to adopt O3 “Mixed Staff and Independent Stoppage Committee” as the founding plan by February 17, 2034, 18:00. The oversight body submits budget plans broken down by verification demand, the Council finalizes member common funds and the budget, and the oversight body executes them, while no deferral of payment, reduction, or personnel rejection authority is granted to the operating body. The independent investigation lead bears signatory responsibility for the evidence collection and the completeness of the reporting of the mixed staff, and the Independent Stoppage Committee is responsible for decisions on temporary stoppage, extension, and lift. Independent hiring is to begin immediately, though actual activation of authority depends on satisfaction of the commencement conditions below.
Read on A4, the ledger codes this undo as costly as well. A founding plan can be re-deliberated — the reopen lines in the record are all re-deliberation signals — a quarterly personnel audit, a monthly tally of delays, a half-yearly accounts comparison — and the judgments add a ten-month hiring check — but a plan that has appointed an investigation lead and fixed a budget line does not unwind when the next audit arrives. The staff synthesis makes the direction explicit in stages: if the conditions for independent funding, data access or committee composition cannot be met, the start of operations is rejected and the founding plan is re-determined; if independent investigation is afterwards made impossible by data blocking, retaliation or budget intervention, reliance on that investigation route is stopped and the staffing and funding structure is re-decided at once. Only if the independent review function cannot be restored is the decision to maintain the mixed structure withdrawn, and then a transition to fully independent staffing goes to the top of the re-judgment. The decision is reversible by re-voting; the situation it has created is reversible only by re-building.
Before the Original Disappears (2042). The setting is the non-operational records preservation facility of the Synthetic Yudam National Defense Technology Agency. Legacy approval and stoppage records are being converted to a new format during the replacement of a military AI system, and an independent verification institute has found that the chronological order of test records is preserved while some context linkages can be confirmed only in the original environment. Once the maintenance contract deadline passes, the original execution environment is dismantled. The roles are the Records Accountability Committee, a records custodian, a privacy supervisor, an appeals reviewer and the person in charge of execution. The hands hold the papers for a maintenance contract that must still be concluded this week and the disposal reservation it must cancel, so that the original execution environment is not dismantled. The eyes wait on the first reproduction test, quarterly comparisons of original and transformed records, monthly checks of access logs, and half-yearly comparisons of new evidence demands against what has been preserved. The deadline is 12:00 on 9 July 2042, with the contract deadline falling inside the same week.
The Records Accountability Committee of the Synthetic Yudam National Defense Technology Agency must decide, by 12:00 on July 9, 2042, how to preserve the originals and transformed records required for the accountability investigation of the military AI scheduled for replacement. The possibility of verifying future human approvals and stoppage judgments, the burden of holding sensitive personal information, and maintenance costs must be considered together, and a single preservation plan must be confirmed.
The Committee sets only the preservation and access method; it does not judge the accountability of persons appearing in the records or the final conclusion of any incident.
Read on A8, the ledger codes this question as human-decides, and the human at the centre of it has not yet appeared in the procedure. The unknowns list what is missing: what contextual information future disputes that have not been received will need, whether specialised personnel able to reproduce the original environment will be maintained, and how much of the accountability investigation can be carried from transformed records alone. The human's place is written twice in the record, once as the committee that decides now and once as an appellant in a dispute not yet filed, who would need to reconstruct what a human approver did through context held only in the environment about to be dismantled. This study reads the preservation plan as an answer to a question about future human action: on a five-year maintenance baseline, with renewal or reduction decided again before it expires, it decides who is to be able to act on the record afterwards, without deciding any individual's responsibility today. Whether anyone will then be able to reproduce the original environment is unverified, so the plan cannot promise meaningful access decades from now.
Read on A4, the ledger codes this undo as costly, and the record shows a case where the cost runs hard in one direction. The committee's choice of a five-year plan can be revisited; the disposal of an execution environment cannot. Keeping everything is not free of cost either: five years of maintenance takes 60 per cent of the records budget, delays the organisation and access of other cases, and prolongs the holding of civilians' sensitive information, and those burdens fall on applicants in other cases and on the civilians in the records. The two alternatives carry the opposite risks. O2 keeps the transformed records and reduces the originals, which cuts those costs and may fail to restore context that has already gone. O3 puts the originals of the current dispute into independent custody, which separates custody from the operating agency and may miss what no case has yet selected. Every judgment in this question repeats the same distinction: a stoppage of access, a rejection of the plan or a contract termination is never treated as approval to destroy evidence. The staff synthesis writes that reversal explicitly, and the direction of the asymmetry is the finding this study reads from the record: the decision that is easiest to change is the one that decides what can never be changed again.
Stop Authority in the Branched Approval System (2192). The setting is a virtual remote habitat and its branched AI supervision network, in the deep future layer. A military AI approval management system has long been operated in separation, and the habitats have come to have different record structures: identical supervision orders link to different legal responsible-party fields, and the Dawn Orbit Verification Office did not examine the results of uses of force. Round-trip confirmation between habitats takes sixteen months. If legacy record environments are disposed of, some approval relationships cannot be restored, and preservation resources are limited. The roles are the Approval Succession Conference, a common human supervisor, local human supervisors, a common verification officer, an original-record preservation officer and a generational succession auditor. The hands hold the supervision certificate due for renewal and the correspondence table that links each approver to each holder of stop authority. The eyes wait on branch-by-branch verification of those linkages, quarterly reproduction tests of accountability relationships, and an annual check of whether a successor can interpret and reproduce the records. The deadline is the Supervision Certificate Renewal Time on 14 April 2192.
The Virtual Dawn Orbit Community Approval Succession Conference: how will it determine the succession of human supervisory authority and the distribution of stop authority for the branched military AI approval management system by the Supervision Certificate Renewal Time on April 14, 2192? One common supervision rule must be established between habitats that have mutually deferred jurisdictions and different accountability records.
Local residents may lose the opportunity for prompt escalation and correction while waiting for a distant supervisor's reply.
All options maintain human final approval for the use of force and an effective stop path, and do not allow the AI to renew its own authority.
Read on A8, the human's place is divided among three roles and none of them is simple. The local human supervisor may suspend the branch under their own supervision without waiting for the sixteen-month round trip between habitats, an interval the record warns does not promise completion of any resumption review; the common human supervisor holds certification and the joint standard; the successor inherits supervision duties from the time of handover, while succession of office neither shifts nor exempts accountability for judgments made before the handover, which stays for separate determination, the original records' responsible-party fields not reproducing cleanly. The ledger codes this question as human-decides, and the staff synthesis's boundary rule carries the point: suspension is not to be lifted by a conflict of judgment, silence or the passage of time, and the AI may not renew its own authority. This study reads the question as asking what a human stop means when the humans are separated by sixteen months and by different record structures.
The three rules distribute different costs. O1 decentralises stop authority to a local human supervisor: stop accountability sits close to current residents and the delay shrinks, but judgments differ from habitat to habitat and the accountability link for common events weakens. O2 keeps common supervision and adds local temporary suspension: the joint accountability standard survives and the local suspension path is immediate, but resumption review is long delayed and a dispute over responsibility stays between the two supervising bodies. O3 stops the common certification and switches to manual approval: reliance on inconsistent certifications is cut and human accountability is recorded directly, but administrative processing is greatly delayed and running the manual system competes with preserving the originals. The unknowns press on each: how far the existing central verification applies to the branched system, whether future supervisors can read and reproduce the legacy accountability relationships, and whether each habitat will accept stop and resumption judgments made under another jurisdiction. If the renewal time passes without a common rule, the common certification expires, the basis for continuing to use the existing approval management system disappears, and the priority of the original-record preservation contract cannot be set.
For the common rules to be applied from the Supervision Certificate Renewal Time of April 14, 2192, it is decided to conditionally adopt O2, “Common Supervision Succession and Local Immediate Suspension.” The Common Human Supervisor handles certification, succession, and the stoppage of the common certification; the Local Human Supervisor has the authority to immediately suspend its own jurisdiction's branch without waiting for remote replies. Resumption requires explicit consent from all relevant local and common supervisors, and suspension is not to be lifted by a conflict of judgment, silence, or the passage of time. No branch renewal is approved on the current input alone. The range that fails to prove the conditions by the deadline shifts to O3's conditional manual procedure; if there is no additional evidence at all, the entire common certification is not renewed.
Read on A4, the ledger codes this question's reversal action as impossible, and it is the only one of the four that does so. Two things run in parallel: a generational handover, in which supervision passes to people who cannot reproduce the predecessor's responsibility fields, and a disposal schedule, in which legacy record environments are dismantled with preservation resources limited. The record's own stop conditions hold the disposal when reproduction failures are confirmed, but a hold is not an undo. In this question the two axes of the study's reading meet: A8 asks whether a human stop can be made real, and A4 asks what happens when the human who could have stopped is no longer the human who answers afterwards. The two judgments that part from the other eleven in this question, the source officer and the independent challenger, both chose to stop the common certification and switch to manual approval; Chapter 14 treats where the thirteen judgments part across the collection.
6.3The debate: meaningful human control or the pressure of speed?
This study reads the two positions set beside one another as differing over what a human position costs, not over whether human judgment matters; the reading notes record no exchange between the authors, and Asaro's article predates Scharre's book. Scharre's vocabulary — human in, on or out of the loop — is the same subject drawn as a design problem, and this study reads the pressures of speed and of accountability as pulling on it, a claim the note itself does not make. A loop that slows a decision enough to be meaningful may fail the decision it is meant to protect; a loop fast enough to matter may be no loop at all. Sagan's record stands behind both: in nuclear command, safeguards existed and near-accidents occurred anyway. Normal-accident theory and high-reliability theory are the two organisational accounts Sagan sets against that record, and this study reads the finding that near-accidents happened despite safeguards as bearing directly on any claim that a named human authority is a working one.
Where the four questions bear on this debate, they bear precisely. In the 2026, 2034 and 2192 questions every option keeps human final approval; no option anywhere in this series proposes removing the human, and the boundary in the 2192 question states it outright: the AI may not renew its own authority. What the records do instead is write who can stop and with what record. In 2026 the stop is tied to a single failed linkage reproduction, to a named person in charge of stoppage, and to preservation of the defective record rather than its repair by post-hoc approval. In 2034 the stop belongs to a committee separated from the operating body's employment and evaluation relationships, and the operating body is denied the budget and personnel levers that would make the stop unenforceable. In 2042 the stop concerns evidence rather than action: access is stopped for personal-information breaches, and destruction is never implied by the stoppage of a plan. In 2192 the proposed stop is local and immediate, with resumption requiring explicit consent from all relevant local and common supervisors and silence never counted as consent; on the current input alone no branch is approved for renewal. The ledger's options row across the whole collection counts 150 options, of which both coders agreed that 82 reduce the immediate loss, and the same row records what those 82 give up, long-term stability most often among the four higher goals. This study reads these four chosen options, from the records' own trade-off descriptions, as taking on delay and giving up administrative speed in exchange for a stop that can later be shown to have been exercised.
Where the questions cannot bear on the debate is also worth stating plainly. They contain no measurement of whether a human can judge at machine speed; none of the four settings gives any figure for the time in which a human approval must be given, and the closest the record comes is the sixteen-month round trip of the 2192 question, which is a communication constraint between habitats rather than a decision loop inside one system. They contain no use of force at all: the 2026 and 2034 records exclude the approval of individual force from the scope of their decisions, and the 2192 premise states that the results of uses of force have not been examined. They cannot show whether a human in the loop would have decided differently, because no outcome is recorded anywhere in them. They cannot show whether meaningful human control is achievable under the speed conditions that a real deployment would impose, because the settings are synthetic and the conditions are given rather than observed. And they cannot test the ban on autonomous weapon systems that Asaro argues for, because the question form and the records' boundaries keep that option off the table from the start.
One further limit belongs here rather than in the conclusion. Across the four questions the judgments agree with one another; only in the 2192 question, where two of the thirteen chose a different option, do they part. That agreement is partly the product of the generator, as Chapter 14 explains. Agreement among AI-generated judgments written under the same inputs and the same inference strength is not independent verification of anything. The staff synthesis in the 2026 question says the same about its own procedure: the number of staff officers' choices is not the criterion of judgment. Where this study's reading of the debate and the records' judgments meet is on a modest point that both support: an approval whose occurrence cannot be reproduced from the record afterwards leaves the record unable to prove, for the purposes of accountability, that it happened. The 2026 judgments expressly refuse to treat a failed reproduction as conclusive proof that no approval existed, and one follow-up branch recovers some of the original approvals; the reverse, that a reproduced record proves a meaningful judgment, is exactly what the records decline to assert.
6.4Hands on
The four questions end where a different kind of question begins. Each of them is a decision question with conditional, revisable judgments attached, and each will be read by people who were not in the room: an appellant who cannot reproduce a signature linkage, a verification staff member whose employment relationship is named in a founding plan, a successor supervisor who takes up supervision duties while responsibility for earlier judgments stays for separate determination, and a future dispute not yet filed against a preservation list written in 2042. That is a question about promises across time.
Chapter 7's series runs from the unfilled contributions of a joint rescue fund in 2026 to an ancestors' oath and a descendant's right of exit in 2164. Its debate is set by Hans Jonas, in The Imperative of Responsibility (1984), and Glenn H. Snyder, in Alliance Politics (1997); Chapter 7 sets out what each argues, and this chapter does not repeat them. This study reads the two together as a question of what the present owes to people who cannot sign, an intergenerational extension of the alliance dilemma that is this study's own reading and not Snyder's. This chapter hands over three things to that one. First, the four autonomy records are themselves written in the four layers of time, and Chapter 12 reads that series again to ask what changes across the layers. Second, what can be undone is treated in Chapter 11, and the ledger's two reversal rows that 6.2 reads are set out in Chapter 2; this chapter points to both rather than repeating them. Third, the human's place and who decides, and the hypothesis that tests whether an automated decision leaves the human's place empty or a name only, are treated in Chapter 9, and Chapter 15 gives the verdicts on that hypothesis and the others.