Chapter 8. Who Decides, and the Human's Place

8.1What the records say about the decider

This chapter reads seven of the 53 records on four of the ledger's axes: A1, who decides, A8, the human's place in the decision, and, in section 8.3, A3 and A4. Chapter 2 sets out the axes and the coding; the verdicts on the hypotheses are reported in Chapter 14. What follows shows what each record writes about authority and what it leaves unwritten.

Begin with the form, because the form sets the ceiling on what can be read. A WSTiger record holds a Scenario, a Question, Options, an Example judgment and an Invalidation line. It has no field for a decider, no field for a deadline, and no field for what is missing from the evidence. Chapter 2 states the rule that follows: what a record form carries in a dedicated field is a fact about the form rather than a finding about the records. The other fields can still hold the same information, since a Scenario, a Question or an Example judgment may name a decider, a deadline or missing evidence. The ledger reports the same fact as counts, and those counts frame everything below.

The direct extraction that fills the A1 count looks for a decider in the record's own entry and finds none in any of the 53 records. The coders, reading the record's wording rather than a dedicated field, found a decider stated in two cases, one team-or-organisation and one individual. The two measurements ask different questions and the chapter needs both: the count reports what the form's fields hold, the coder row reports what the record's words support. Alongside that gap the form produces, with high regularity, a judgment and a reason: the reason recorded row shows coders' agreement on 49 of 53 cases (92%), with values reading yes stated in 40 cases, yes implied in 5, unsettled in 4, no implied in 2 and not applicable in 2. So the same form that reliably captures why a judgment was reached carries who reached it in no dedicated field and in only two sets of words. That asymmetry is the starting point of this chapter, and it is why the decider had to be read out of the setting by code rather than taken from a field.

Chapter 2 reports the method and the agreement per cell. The two rows this chapter lives on are these.

Read the decider row first. Two records out of 53 carry a decider the coders found stated; one of those reads team-or-organisation and one individual. Everywhere else the reading was an inference, and in twenty cases the two coders did not settle on a single value. Where a setting describes a desk, a bank or a vault, the inference runs to an organisation; where a setting puts one person's attention at the centre — a solo desk, a single trader, a lone key holder — the inference runs to an individual. Those inferences are reasonable, and they are still inferences: apart from the two stated cases the record does not say who decides, so the ledger cell records what a careful reader would take the setting to imply.

The place row tells a similar story in a different direction. In most readings the setting shows a human acting: "human-decides (implied) 25". Three cases read "human-approves (implied)", where the human step is a check rather than an act. Two cases read "ai-or-rule-decides (implied)", where the decision is left to an automated agent or an automatic rule, and in both cases the human step that accompanies it is not specified. Three read "unknown", and twenty read unsettled. Almost none of this came from a record's words.

What follows for the reading is worth stating once. A cell filled by implication answers a different question from a cell filled by the record's words, and the two should not be quoted as if they carried the same weight. The place row is the clearest case: 25 of its values read human-decides by implication, which means a reader could see a human acting in the setting, not that any record assigns the decision to one. This study reads those values as evidence about the settings and about how they were written, and stops there. A claim about how an organisation allocates authority would need something the form does not hold: a named role, a rule of delegation, an outcome. The seven cases below show what each setting writes about authority, and the ledger shows how consistently it writes it.

The diagnosis this chapter can already offer is narrow and it concerns the reading, not the market. An unsettled decider cell leaves that one cell open and does not unsettle the others: the ledger codes the bearer and the horizon as separate cells, and a record can name who is affected without naming who authorised the decision. This chapter reads the unsettled decider cells as a limit on attribution, so that a statement about who waits and who pays has to come from elsewhere in the record or not be made at all. Chapter 10 takes up the question of who bears the waiting. Each option below is shown with what it gains, what it risks and who stands in the gap.

8.2Agents and their limits

Four of the 53 records are written from inside an automated desk: five agents share one collateral, an assistant pack proposes a trade, several strategy agents share one account, an analytical agent's judgment improves. Each poses the same underlying question in its own terms — how much authority may an automated system hold — and each does so without naming a person. This section reads them on A1, who decides, and A8, the human's place.

Five agents one budget. Synthetic scene. A fictional solo desk in a 2026 setting receives independent-looking proposals from five assistants, all competing for the same collateral. The desk is one place of attention facing five claimants on one collateral pool, with a single risk figure to cover all five. It has one reader, and the proposals are proposals and not decisions: the record names no commitment and no answer that binds. What the desk waits for is the answer to its Question, how one budget is to be divided among five claimants, and the role that will give that answer is not written into the setting.

Three options. This study reads Agent-level allocation as splitting the single budget into one share per assistant: each would work inside its own number and the desk would be freed from repeated negotiation, and the risk would be that nobody checks the five against one another, since the record names no combined figure and no role to hold one. The option's own text runs to three words, and everything beyond them is this reading. A portfolio gate checks combined exposure before commitment, which matches the setting's problem; the risk is that whoever runs the gate decides what counts as correlated exposure, and the record names no such role. Manual review keeps a person in the loop; the record gives no limit on how long review takes, so the five proposals wait, and the waiting falls on the desk.

The ledger's cell for this case reads decider "individual", human place "unsettled", horizon "unknown". The reading of an individual decider follows from the setting: a solo desk, one person holding all five proposals. But the cell describes the setting, not the option that would be adopted. The record does not say what instrument a portfolio gate would be: it could be a rule checking combined exposure or a person checking it, and neither the setting nor the Question settles which. This study reads the unsettled human place cell as the trace of that silence.

Assistant proposes who commits. Synthetic scene. A hypothetical multi-agent desk in a 2030 setting holds one pack combining historical analogies, model forecasts and execution proposals. The three kinds of material sit side by side, and one of them carries a trade. The desk waits on a question the pack does not answer: between the proposal and the commitment, who holds the authority to bind it. This study reads the setting as one with a person at the desk and the pack open in front of them; the record names no reader and no instrument of commitment.

Shared execution authority removes the wait between proposal and action, and the risk is that a pack combining analogies, forecasts and proposals can then commit without a second person reading it. Separating proposal and approval puts a second step in place, and the cost is the interval between the two steps — an interval the record does not measure. Keeping the pack research-only answers the authority question by withholding execution rights from the pack itself; what it changes outside the pack is not settled here, since a person could still act on the research.

The ledger reads this case's decider as "team-or-organisation" and its human place as "human-approves". Authority sits in the separation itself: the approval step is where the human appears, and the record's own reading requires that boundaries travel with the evidence rather than being granted by the pack. Where the record is silent is on the identity of the approver, the form the approval takes, and the time within which it must be given. A place marked "approves" with none of those three is a marker in a diagram.

Agent order conflicts. This record moves from authority over a trade to authority over an account. Several strategy agents share one account; one can expand a position another reduces, and each can modify the other's protective orders.

Each-agent limits keep the strategies autonomous and free of negotiation, and the risk is collision: two individually safe orders can produce an unsafe account. Order ownership and budget reservations make each order attributable to one strategy and reserve the account's budget at order time, which adds work before the order and may delay it. Verifying simultaneous execution control tests for collisions before they happen, which depends on the control being able to see every order in flight.

Supervisable automation scope. The last of the four turns to scope. Research quality has improved; the assumption behind the record is that order limits, failure detection, manual intervention and separation of authority remain unverified.

Expanding authority with the judgment's performance gains execution speed from improved analysis, and risks growing unverified what the record lists as unverified: limits, detection, intervention, separation. Keeping the limited scope keeps the existing cap on what the agent may do whatever its analytical performance, and leaves the record's unverified controls where they are: order limits, failure detection, manual intervention and separation of authority. Separate approval of authority registration decouples the two, and costs a step each time scope changes.

The ledger reads this case's decider as "team-or-organisation" and its human place as "human-approves", with the horizon unsettled.

Where authority sits in each of the four, in the words the records themselves use: in a risk budget; in the boundary between proposal and approval; in order ownership and budget reservations; in the separation of analytical capability from operational authority. In every case the record locates authority in an instrument — a gate, a boundary, a reservation, an approval — and in no case locates it in a named role. The human appears where a permission boundary is drawn: reviewing, approving, releasing. The record is silent about who that person is, on what schedule the human step must occur, and what the desk does if the step does not occur in time.

What the four show on A1 is worth separating from what the coder row shows. The row gives one value per case; the case gives a value only for its setting. In Five agents one budget the coders read an individual, on the strength of a solo desk; in the other three they read a team or an organisation, on the strength of a desk or a shared account rather than of any named body. This study reads that spread as the shape of the evidence rather than of the markets the settings describe: the more a setting names an institution, the more the coders find an organisation, and the more it names one person's attention, the more they find an individual. On A8 the four split, two unsettled and two approving, and none leaves the decision with a rule outright. The five records the lenient reading counts for H7 are drawn from the whole corpus and not from this selection alone.

That silence has a cost on each option and it falls in a particular place. This study reads the four options as trading a check against a delay: the gate, the reservation and the approval step can each put time between a proposal and a commitment, and no record here says who would pay that time or for how long. The options that hand the decision straight to a share, a boundary or a performance figure remove that kind of delay while leaving the combined exposure or the unverified controls to surface later. Neither half of that trade is settled by the records; it is a reading of the designs the options name. This study reads the four records as writing authority as a limit rather than as a person, and the ledger's unsettled cells as what that writing leaves open.

Each of the four also carries an Invalidation line, and those lines state what would make the judgment be revisited. Five agents one budget asks for a revisit when correlated exposure or total loss scenarios exceed the approved envelope. Assistant proposes who commits asks for one if a new adapter bypasses the risk gate, audit trail or human approval policy. Agent order conflicts makes its change condition the consistency of exposure and protective orders under conflict and partial-failure testing, with the next check aggregating all agents' reservation risk before order approval. Supervisable automation scope makes its change condition the verification of control mechanisms and intervention procedures in a restricted environment, with the scope then adjusted through a pre-established approval process, and its next check is to specify the agent's actions and the responsible party in case of failure. This study reads the two conditions that name a verification or a test as the hinge of the judgments above: the settings describe controls as unverified, and the records make the judgment turn on verifying them.

8.3Who holds the switch

Three of the 53 records are written from the moment something goes wrong: an anomaly detected, an agent halted, a venue breaking. Each asks who may stop the system, and each names the instrument or the step more readily than the person behind it.

Who holds the switch. A hypothetical desk runs three execution agents under one human supervisor who cannot watch every screen at once. The setting states one constraint, that attention is finite; the inference that the breach will not wait for it belongs to this chapter.

One option:

Each option names a different holder. The supervisor switch puts the halt in a person's hands, and its cost is the time it takes that person to notice; the record's own reading asks that halt authority be given to whoever can observe the breach first and that the handoff be rehearsed, which is a cost paid in training time before any breach occurs. Per-agent hard limits act without waiting for a supervisor, and their cost is that the limits encode in advance what the anomaly should look like. The record does not rank the three options by speed, and a hard limit can act before any detection runs. This study reads the automatic halt as making the detector itself the decision: a wrong detection would halt a sound agent, and a silent one would halt nothing. The ledger leaves this case's decider unsettled and its human place unknown, which is the honest reading of three options that point in three directions on who acts. Its horizon reads "unknown", its reversal cells read unsettled.

One anomaly three agents. A shared anomaly detector halts three trading agents during a market-data outage, though only one agent was using the failed feed.

One option:

Keeping the global halt protects all three agents under one safe state, and its cost falls on the two that were not at fault and on whatever their continued operation would have been worth during the pause. Isolating the affected agent frees two, and its risk is that the diagnosis is the decision again: the record's own reading asks that a common data failure be separated from independent strategy breaches while preserving a safe default state, which means the detector's judgment about which feed failed carries the whole choice. Requiring human release for all three keeps the safe state as the resting place, and leaves three agents frozen until a person acts — with no deadline in the record, no named role, and no statement of what the pause costs. The ledger reads this case's decider as unsettled and its horizon as "days-to-months", an implied reading by the coders, since the record writes no horizon. Its reversal cells are unsettled.

Three venues halt apart. Circuit breakers on three venues trigger seconds apart, and each pause leaves one leg unhedged.

One option:

Halting everything at the first trigger applies one decision to all three venues at once. This study reads its cost as stopping the venues whose breakers had not yet fired, and does not take the unhedged interval as avoided, since the record does not say what a desk-wide halt leaves behind. Staggering halts preserves continuity on the venues that did not trigger, and the record's own reading states the consequence: asynchronous halts turn a hedged book into sequential outrights, and the halted state should be defined before the halt arrives. Keeping quoting through avoids pauses altogether, and leaves the desk exposed to the effect of the breakers while quoting. This is the one case of the three whose decider the coders read as "team-or-organisation", though the record does not say which body holds it: the venue's breakers constrain the desk's choice rather than make it. The human place cell still reads unsettled, and the reversal cells unsettled.

On A3, the three records share one feature that matters more than any difference between them: none states a deadline, and the ledger reports that across all 53 records, cases with a deadline in the record stand at 0. The settings speak in seconds, in halted states, in pauses; what is missing is a deadline rather than a unit, since no record ties those seconds to a date or a limit by which a step must be taken. The cost of waiting therefore appears in these scenes as a technical property — an unhedged leg, a frozen agent, a sequential outright — and never as a cost carried by a named role. On A4, all six reversal cells across the three cases read unsettled: the coders could not settle whether these decisions can be undone. A halt can be released, which is the whole content of one option; whether the exposure built up during the pause can be undone is a different question, and none of the three records answers it. This chapter does not fill those cells. What the three give on A3 is a horizon without a clock and a cost without a bearer: one reads a horizon of unknown and two read days to months by implication, and the bearer cell is unsettled in all three. A pause has consequences for the desk, for the two agents not at fault and for the leg left unhedged, and the records do not say which of those is the party who waits and pays. This study reads the pair as the place where a decision record is thinnest: a stop can be made in a second and understood only later.

The diagnosis here is about attribution rather than about markets. In each of the three, some options answer the question "who may stop this?" with an instrument — a limit, a detector, a breaker — and others with a human step: the first record names one human supervisor who cannot watch every screen at once, and the second asks for human release. What none of the three gives is the office-holder behind those steps: who the supervisor is, who releases, and when either must act, so the time such a step takes cannot be read from the record either.

The three also state what would make their judgments be revisited. Who holds the switch asks for a revisit if an agent exceeds its envelope before any halt triggers. One anomaly three agents asks for one if dependency records show that the supposedly independent agents share the failed input. Three venues halt apart asks for one if synchronised halt rules or cross-venue close-out procedures are published. This study reads those lines as the answers the records give to what would change their judgment, and notes that each names a condition outside the setting rather than an outcome it reports.

8.4The human's place and H7

The human's place is one of the ledger's three extra axes, coded as A8, and it is carried by two rows of "The ledger's numbers".

The first row asks a different question from the second. The place row asks where the human stands — deciding, approving, or displaced by a rule or an agent. The role row asks whether the record specifies what that human is to do and who is to do it, and the answer is that the coders found no human role specified in almost every case: "no (implied) 38", "no (stated) 1", against "yes (implied) 1" and "yes (stated) 1", with 11 unsettled.

H7, the pre-registered hypothesis this chapter cannot settle but must report, says that the more a decision is left to AI or an automatic rule, the more the human's place in it — approval, exception, review — becomes empty or a name only. The verdicts are these.

Under the strict reading the counts are empty — supports 0, against 0, unclear 0 — and the hypothesis is not applicable to any of the 53 records. The lenient reading also takes what the setting implies, and there it finds support — five records, none going against — including two of this chapter's seven: Assistant proposes who commits and Supervisable automation scope, whose place cells both read "human-approves".

What it means, concretely, that most of these records name a human step without its role: in this chapter's cases the verbs are approve, release, review, expand, halt, and the noun behind them — who approves, who releases, who reviews — appears in no field. Two exceptions keep the generalisation honest. Who holds the switch names one human supervisor who cannot watch every screen at once, and Supervisable automation scope writes down conditions: verified controls and intervention procedures, a pre-established approval process, and a next check that would specify the agent's actions and the responsible party in case of failure. Those conditions may be incomplete, but they are written down. What a step without a named role leaves open is who schedules it and by when; what a step without a written condition leaves open is how anyone would check it. Neither opening shows that a rule takes the step over. This study reads the lenient verdict on H7 as fitting that opening rather than as proving it.

Two features of the reading deserve a caution. The place row is implied in almost every one of its values, so the human's place in these records is an inference from settings, not a written fact; and the strict and lenient readings differ only by how much inference is allowed, which is why this study reports both. Both describe what the records do and do not carry.

The seven cases make the caution concrete. In Five agents one budget the place cell is unsettled, in Agent order conflicts it is unsettled again, and in Who holds the switch it reads unknown. Only the two records whose options separate a proposal from an approval carry a place cell of human-approves, and one of those is the record whose Invalidation line asks for the responsible party to be specified. This study reads the human's place in these cases on a handful of implied values and two stated approval steps: enough to report the pattern the ledger shows, and not enough to say how an organisation actually assigns the step.

8.5Beside the literature

Two works from the literature read in this study bear directly on the question this chapter has been circling: where decision premises come from when the record does not say who holds them.

Herbert A. Simon's Administrative Behavior (1947) describes decisions made from premises within limits of knowledge and attention: a person deciding does not survey every option and every consequence. This study reads those premises as some chosen and some given by the organisation and the situation. The terms that later became the shorthand for this idea — "satisficing" and "bounded rationality" — are Simon's later terms, from the mid-1950s, so this study writes what Simon later called bounded rationality.

What Simon's account adds to these records is a way of asking what a decision was made from. The four agent settings each state a question and three alternatives and each state a judgment; none of them states the grounds, factual or in value terms, on which the judgment rests, or the attention that was available while it was formed. The example judgments supply some of those grounds after the fact, and this chapter has read them as premises. The records still leave open who took those premises as given, who could have refused them, and what else was competing for the same attention when the judgment was made. That is where the records bear on Simon's account, and where they cannot: they carry grounds, but not the limits within which the grounds were used.

Those limits matter in these settings because the scarce resource is named in one of them and left unnamed in the others. Who holds the switch states the constraint directly, one supervisor who cannot watch every screen at once; the four agent records name no attention budget at all. This study reads that difference as part of why the human's place in the ledger stays thin: a role with no stated attention limit can be described as approving, releasing or reviewing with no account of what else it was doing at the time. The relevance to the four agent records is direct. This study reads the four example judgments as stating the grounds their judgments proceed from: risk belongs to the combined portfolio; authority boundaries travel with the evidence; order ownership and account control are kept separate; analytical capability and operational authority are approved separately. The Options beside them are the alternatives those grounds are used to choose among, and this reading extends Simon rather than reporting him. What the records do not carry is who held those grounds or on what attention budget; the coders' 20 unsettled decider cells record disagreement about who decides across the corpus, and cannot show more than that.

Michael D. Cohen, James G. March and Johan P. Olsen's A Garbage Can Model of Organizational Choice (1972) describes a different condition: in organised anarchies, with unclear goals, unclear technology and changing participants, decisions come from the meeting of independent streams of problems, solutions, participants and choice opportunities, rather than from the pursuit of goals. Here the fit is uncertain, and this chapter says so rather than forcing it. The four agent records set out a question and its alternatives in advance, and describe their technology only in outline: the automation-scope record itself leaves order limits, failure detection, manual intervention and separation of authority unverified. A halted state belongs to the switch cases and not to these four. Sparse descriptions cannot show that the model's conditions are absent, and this chapter does not claim that they are. This study reads the record form as gathering the material of a decision into one question, three options and one judgment, and reads the model as not treating such a gathering as work left undone. Five agents one budget resembles the model in one limited respect: several proposals reach one decision point together, and the record calls them independent-looking rather than independent. The model's independent streams are streams of problems, solutions, participants and choice opportunities, not a set of proposals, so the resemblance is partial. This study reads the single question the record forms out of that meeting as a property of the record form, and does not claim that the model rules such a form out. What the model does ask is whether problems, solutions, participants and choice opportunities meet in the pursuit of goals or in a looser collision, and this study reads these settings as showing solutions already prepared — a gate, a reservation, an approval step, a halt — waiting for a question to attach them to. Whether participants change while a question is open is exactly what the records do not say. The model is therefore a comparison this chapter cannot resolve, not a verdict on the scenes.

What the records add beside these two works, and this is this chapter's own reading rather than theirs: this study reads the vocabulary of decision in these scenes as locating authority in boundaries rather than in named people. Simon's premises are the grounds a decision is made from; here those grounds are carried by gates, reservations, limits, halts and authority boundaries, and the people who set or apply those mechanisms are barely written into the settings. The human's place survives at the boundary — approving, releasing, reviewing — and thins where a rule can act instead. The ledger's rows give this reading a shape the works do not: a place cell that is implied rather than stated in almost every value, and a role cell that reads "no" for nearly every record.

Where the comparison fails is worth stating plainly. Simon analyses how organisations decide, from premises and under limits of knowledge and attention; Cohen, March and Olsen build a theoretical model of choice in organised anarchies. The two accounts are distinct, and this chapter keeps them apart. These records are written after the analysis is done: the options are already distinct, the judgment is already stated, and a condition for revisiting is already set. Nothing in the form records how long a person had, what was on the screen, which proposal was read first, or what the desk did while waiting. And because these are synthetic settings with no outcomes, none of the seven can show whether a boundary held, which is the question Chapter 7 treats as the difference between judging a decision before and after its answer.

This chapter has read authority as a limit in seven records, with the human named at the boundary and, in one case, as a supervisor who cannot watch every screen at once; and it has read each record's Invalidation line as the condition its judgment waits on. Chapter 9 takes up the same question from the other end: authority that travels by key, by copy and by instruction, where the hand that holds it can be named but not always watched.