Displacement occurs at specification, not at automation — a first-person case from investment banking, 2006–2009
The dominant account of technological displacement locates the causal moment at automation: a system becomes capable enough to perform a task, and the humans performing it are released. This paper argues from a first-person case that the account is late by one step. The displacing act is specification — the moment heterogeneous professional output is rendered into a stable, repeatable sequence. Once that has happened, the function is already eliminated in principle; building the machine is a downstream engineering exercise of comparatively trivial cost modeled.
The case is unusual in one respect that the literature cannot manufacture: the author performed the specification, was rewarded for it with the mandate to generalise it, and was subsequently released along with the function documented. Studies of displacement typically draw on one of two populations — those who designed the automation, whose incentives run toward describing it as value creation, or those displaced by it, who were not present when the design decision was made. This account is from a single person occupying both positions, with an incentive structure that runs against overstatement.
The paper isolates three claims. First, that the specification was a transcription rather than a design — the automation was produced with a macro recorder, meaning the work was already algorithmic and merely unwritten. Second, that displacement propagated through merit: the individual who optimised locally was promoted into the role that generalised the optimisation, so the eliminating decision was made by the highest performer rather than imposed by management. Third, that this yields a leading indicator better than task-routineness measures: the function at risk is the one where a high performer has recently been given a mandate to standardise output across teams prospective.
A note on scale: the tools in this case were primitive. What follows is therefore a lower bound on the phenomenon, observed with a keystroke recorder in 2007. Systems that can infer a procedure from examples rather than requiring it to be recorded extend the exposed class from work that is repeatable to work that is inferable, which is very much larger.
Between 2006 and 2009 the author was an analyst in Relationship Management and Securities Finance at J.P. Morgan, latterly subject-matter expert for Titan — a client-facing business-intelligence tool which the author built and the bank deployed to its corporate clients documented.
What Titan did had not been done. Before it, a corporate client could not log in and look at its own accounts; visibility was mediated — it arrived as a report, produced by an analyst, on the institution's schedule documented. After it, clients ran the reports they needed on demand, at the press of a button documented.
That is the same button, and the coincidence is the point. §1 above describes an analyst's month of labour collapsing to a keystroke on the inside of the firm. Titan is that collapse handed across the counter: the same function, the same button, now belonging to the client. The displacement was not that a machine replaced an analyst. It was that the work stopped being work at all, on both sides of the relationship at once.
And the residual labour was teaching clients to schedule it. Time was spent showing them how to set the reports they had asked for to run on a timely basis — weekly, monthly, quarterly documented. This is worth stating precisely, because it is the last step and the least visible one. Scheduling is the final piece of judgement in the function: not how the report is produced, which the recorder had already captured, but when one is needed. Handing that over completed the transfer. The analyst's remaining work was instructing the client in the exercise of the last discretion the analyst still held — and being paid to do it, and regarding it, correctly, as good service.
What that removed was not a task but a seat. The arrangement Titan replaced was a team of analysts sitting beside the relationship managers, researching and producing reports on request. Once clients could pull and schedule those reports themselves, that arrangement had no remaining function documented. This paper does not claim the corresponding headcount reduction — §7 lists the aggregate effect as unidentifiable from a single case and declines to assert it. What is claimed is narrower and harder to dispute: the work the seat existed to do had stopped existing, and it had stopped because it was written down.
The author's own reading at the time is worth recording exactly, because it is the paper's thesis arriving as a mistake rather than as an argument. Having removed the research-and-reporting burden from the desk, he concluded that the freed capacity was his: now, at last, there would be time to do research documented. The inference was reasonable and it was backwards. Capacity released by specification does not accrue to the specifier. The function had been made portable, and a portable function does not need the person who made it portable — it needs nobody in particular. He read the outcome as liberation. The institution read the same fact as completion.
It is also the same object as the rest of this corpus, arriving from the other side. The Response Gap argues for publishing a formula the counterparty is subject to but cannot see; Titan published a balance the client owned but could not see. Both are cases of asymmetric legibility — one party holding, as an operational convenience, information the other party has the stronger claim to.
Titan matters to the argument twice over, and the second time is the one that generalises. Equivalent client-facing intelligence tools subsequently became standard across the industry modeled — an inference from their later market presence rather than from any knowledge of competitors' build decisions. And in a later role at the New York branch of Banco do Brasil, the author purchased the same capability as a vendor product from BNY Mellon, the institution with which the branch custodied its own assets documented. The specification did not merely eliminate the function that produced it. It escaped the firm, became a product, and was sold back to its own author.
The assigned work was the production of performance-attribution reporting for large corporate clients: analyses of an entire market that identified where a client's performance had come from, and why. The instrument was a spreadsheet. The reporting cycle took approximately one month of full-time labour per iteration documented. The cycle was not one report: several clients and several reports were carried concomitantly within it documented.
What the work actually consisted of, however, was reading contracts. Attribution requires knowing the terms under which value was shared, and those terms were frequently piecewise: above one threshold, a stated percentage payout; above a further threshold, a different payout entirely documented. Such a schedule is a step function. It has discontinuities, and near a boundary an arbitrarily small change in measured performance produces a large change in what is owed.
The finding, repeated contract after contract, was that the signatories did not understand the mathematics of what they had agreed to documented. The document reads as a schedule of proportionate sharing. It behaves as a cliff. Nothing is concealed — the thresholds are written down — but their consequences are not legible to a reader without the relevant training, and the party that drafted the schedule chose where the boundaries sat.
The author consequently rewrote contracts on the basis of those findings, and advised counterparties on how such contracts should be written documented.
⚠️ This is the origin of the corpus and it should be stated plainly. The line of work published in 2026 as The Response Gap — a threshold parameter, set by one party, governing an irreversible consequence, undisclosed in its behaviour to the party it acts upon, and correctable by requiring the formula be published — is not a new research interest. It is the same object this author was paid to find, one contract at a time, between 2006 and 2009. §8.1 of that paper argues for published liquidation formulas. That argument was first made across a desk to corporate treasurers who had signed a step function believing it was a slope.
The author automated it using the spreadsheet application's macro recorder — a facility that captures a user's actions and emits them as executable code. Building the automation took approximately one month. Thereafter the same reports were produced by pressing a button documented.
And then it was done again. This is the part that matters most and the part most easily lost in the retelling. The first automation was not the event. It was the first instance of a campaign that ran for years — repeated month over month, relationship manager by relationship manager, client by client, report by report, each one a separate act of transcription documented.
Four features of that history carry the paper.
The payback period was one iteration. One month of labour to eliminate one month of recurring labour. Every subsequent cycle was pure return. That is not a marginal efficiency; it is a step change with break-even inside a single period — which is precisely why no approval process governs it. The decision is small enough for one person to take alone, and it was taken alone, hundreds of times.
The automation was recorded, not written. The load-bearing detail. The author did not analyse the task, decompose it, and design a program to replicate it. He pressed record and did his job. The resulting code was a transcript.
And the transcription itself was free. This is the sharpest form of the paper's claim and it should not be softened. Recording is coextensive with performing: if it takes a minute to do, it takes a minute to record it documented. The act of specification therefore carried no marginal cost in labour at all — not a month, not an hour, nothing beyond the cycle that was going to be worked regardless. What cost approximately a month was the build: turning the transcript into something that ran reliably documented. So the displacing act was not merely cheap relative to its return. It was, in isolation, free — and that is why nothing in the organisation could price it, budget for it, or notice it happening.
The increments were individually invisible. No single act in that campaign was large enough to warrant review. Each was one analyst improving one report for one client — the most unremarkable thing that happens in an office. There was never a proposal, never a business case, never a meeting at which the elimination of a function was on the agenda, because at no point did anyone hold a unit of work large enough to be called that. The aggregate was the specification of an entire professional function. The aggregate was never on anyone's desk modeled.
And the outcome was promotion. Not for one macro — for having, after years, become the only person who understood the logic across the whole book. The SME role was not a reward for building a tool. It was the institution recognising that the author had become the specification, and asking him to write it down for everyone documented.
A macro recorder cannot capture judgement. It captures a sequence of operations. If a recording reproduces the output, then the output was a function of a stable sequence, and any judgement in the work had already been exercised — upstream, once, in selecting that sequence — and was thereafter being replayed rather than exercised.
This gives the paper its central claim in its strongest form:
The month of analyst labour per cycle was not the cost of thinking. It was the runtime cost of executing a program on human hardware modeled.
On this reading, the function did not become automatable in the month the macro was recorded. It had been automatable for as long as the sequence had been stable. What the recorder did was not eliminate the work — it revealed that the work had already been eliminated in principle and was being performed anyway, because nobody had transcribed it.
The corollary is a diagnostic, and it is uncomfortable:
The argument above assumes something a reader who has never used a macro recorder may not grant: that pressing record produces source code a non-programmer can then read, modify and generalise. It does, and the mechanism is worth setting out step by step, because the absence of a learning curve is load-bearing for §4 and §6B. Nobody was trained. Nobody needed to be.
Alt+F11 opens the Visual Basic Editor. Under Modules is a Sub containing your actions, rendered as VBA statements. This is the moment that matters: the output is not an opaque recording. It is legible source code in a real programming language, and every line of it corresponds to something you personally just did.Range("A1").Select then ActiveCell.FormulaR1C1 = "5". You sorted a column; there is a sort statement naming your column. The recorder is, in effect, a dictionary from actions you already know to syntax you do not — the only such dictionary most office workers ever need.F8 executes one line at a time with the worksheet visible beside the editor. The sheet moves as each line runs. Cause and effect are directly observable, which is a better feedback loop than most deliberate programming instruction provides.Range("A1") with a variable, wrap the repeated block in For Each, add an If for the case that differs, lift the filename into a parameter. Each edit is small, each is testable by running it, and none requires knowing anything not visible in the transcript you started from.The path from doing the job to owning a program that does the job therefore contains no step at which a person must decide to become a programmer modeled. There is no course, no budget line, no request for tooling, no moment at which anyone announces an intention to automate. There is an analyst who wanted to stop retyping things, and a facility built into the software he was already issued.
This is the precise sense in which the displacing act is ungoverned. It is not that the organisation failed to supervise automation. It is that the automation was indistinguishable from competent use of a spreadsheet, which is what the analyst was hired to do.
The argument so far moves the displacing moment earlier than the literature puts it: not at automation but at specification. That move does not obviously terminate, and the objection should be made here rather than left for a reader to make modeled.
By the same logic the moment recedes twice more:
Each step back is valid, and each buys generality at the cost of something the paper needs. Step 3 is almost certainly true and almost entirely useless: it names a condition rather than an event. It cannot be dated, has no author, admits no counterfactual, and licenses no intervention. A claim that displacement began when modernity began explains every case and predicts none.
The paper therefore stops at step 1, and the stopping rule is not arbitrary:
Specification is the last rung with all four properties the argument requires: it is datable (a recording has a timestamp), attributable (one person pressed record), observable in principle (the transcript exists as an artefact), and governable in principle (an institution could, if it wished, require that transcription of a function be logged the way procurement is logged). Step 2 retains attribution but loses the artefact. Step 3 retains nothing.
This is the same structure §5A describes in a different register. A threshold is invisible from inside the range where it does not bind; optimisation-as-frame never binds visibly because it has already been crossed everywhere, at all times, by everyone. That is precisely what makes it unavailable as a locus of governance — and precisely why the useful claim is the narrower one.
The regress is recorded rather than resolved. If a reader prefers step 2 or step 3, the case in §1 is unaffected: the recording still happened, still cost nothing, and still eliminated the function. What changes is only whom one would ask to have done otherwise.
Step 3 is ungovernable as a decision but it is not therefore beyond response. What it rules out is a corrective posed in its own terms. One cannot optimise one's way out of a condition in which everything has become optimisable; a better optimum is still an optimum. The response, if there is one, has to be the antithesis: a deliberate commitment to a position off the optimum, held on purpose, at a cost that is real and is the point modeled.
Call it the neg-optimum: the margin a system holds precisely because it does not pay under the objective being maximised.
The reason is structural rather than sentimental. Optimisation is a single-objective operator. A system has at minimum two objectives that do not commute — output now and existence later — and the operator only sees the first. An optimum is a point; a point has no neighbourhood. A system driven exactly onto it has, by construction, zero reserve against the perturbation it did not model. Every property that buys survival under perturbation — redundancy, duplication, idle capacity, slack, unallocated attention — is scored as waste by the operator, because waste is exactly what it is under the objective the operator can see. The measurement is not wrong. It is incomplete in the one dimension that determines whether there is a system next period.
This is where the paper's own case supplies the sharpest instance, and it is not a general point about resilience documented.
The analyst seat was slack. Sitting beside a relationship manager, researching, assembling reports by hand, it was the least defensible line under any efficiency review — heterogeneous, unmeasured, irreproducible, and exactly the shape of thing an optimising frame is built to find. Once reports could be scheduled, it was redundant capacity by the frame's own arithmetic.
But that seat was also the only place in the institution where the transcription could have happened. Specification requires someone who holds the function in working memory, has the tool open, and has enough unassigned time to play with it — the verb §6B records the institution rejecting. A fully optimised desk has no such time by definition; unassigned time is the first thing an optimum removes. So the operation eliminates its own precondition:
The failure is invisible from inside for the reason §5A gives: the loss appears as an absence of events, and an absence of events reads as stability. Nothing on the P&L records the tool that was not built.
The neg-optimum is therefore not a plea against efficiency. It is the claim that a system which cannot articulate why it is holding a position it knows to be sub-optimal has lost the vocabulary in which its own continuation is expressible — the same failure §6B records at the level of a single word. We do not play is what an institution says when it has optimised away the category in which its next capability would have been built. The sentence was correct as description and catastrophic as policy, and nobody present could have told the difference modeled.
What a neg-optimum would require in practice — how much slack, held where, justified to whom, and how to distinguish deliberate reserve from ordinary inefficiency — is not settled here and is recorded in §7 as open.
The obvious objection to §2.3 is that no reserve needs to be held deliberately, because the system regulates itself. Nature is self-regulating; markets are self-regulating; local optimisation by each agent produces a globally tolerable outcome that no agent intended. This is the invisible hand, and it is an optimisation argument — the strongest one there is.
Two things should be separated before it is granted modeled.
First, the appeal to nature proves less than it appears to. Nature does not seek self-regulation; that is teleology read backwards off a survivorship artifact. Configurations whose feedback was not self-regulating are not available for observation, having already terminated. What we call nature's homeostasis is the residue after the counterexamples removed themselves. So the intuition that self-regulation is the normal case is drawn from a sample constructed by the very failure mode in question — the same structure §5A describes, an absence of events read as stability. It is not evidence that any particular system before us is self-regulating. It is evidence only that the ones we can see so far were.
Second, self-regulation is dimension-specific, and the invisible hand's dimension is price. Decentralised optimisation does supply genuine negative feedback there: excess demand raises price, which suppresses demand. The loop closes inside the market's own boundary and timescale, and the argument is sound. But cost reduction by displacement is a different loop with a different sign. Each firm that removes a function lowers its costs; rivals must follow or lose; and the restoring force is missing because the displaced party is not a counterparty to the transaction. Individually the move is rational and compulsory. In aggregate the population that is being removed as cost is the same population that appears on the other side of the ledger as demand.
None of that is new, and the paper does not present it as new. It is the fallacy of composition applied to labour cost, and it has been stated repeatedly and better — underconsumption arguments in the classical literature, Keynes's paradox of thrift, Kalecki on the profit consequences of wage compression, Minsky on stability as the generator of instability documented. The concentration version is equally familiar: competition optimised hard enough produces concentration, and concentration removes the competitive pressure that was the regulating mechanism in the first place, so the optimiser consumes the feedback loop that legitimated it — §2.3's shape at market scale.
What this paper adds is not the loop. It is a claim about the damping term modeled.
Historically the displacement loop was rate-limited by capital. Removing a function required building a machine, and machines required investment, procurement, lead time, and a decision someone had to sign. That cost was friction, and friction is what kept a positive-feedback loop from running away: displacement could only proceed as fast as capital could be committed, which gave every other adjustment — retraining, new sectors, wage recovery, political response — a timescale to operate on. The classical optimism about technological unemployment is, read carefully, an argument about relative rates, and it was right for as long as the rates held.
The claim of §1 and §2 is that specification has no capital cost. Transcription takes a minute, requires no approval, leaves no procurement record, and is indistinguishable from competent use of the tool. If the displacing act is specification rather than construction — and §2.1 gives the mechanism by which an ordinary user performs it without deciding to — then the friction term in that loop has gone to approximately zero while the loop's sign is unchanged.
Whether the damping was in fact load-bearing, and whether some other friction has replaced it, is exactly the question this paper cannot settle from a single case. It is stated so that it can be attacked. §7 records what evidence would count against it.
If the loop is undamped, the natural question is where it stops. The question is usually asked as though there were a floor — some level of cost below which the thing cannot sensibly be pushed. There is no floor. There is a boundary, and the two are not alike modeled.
An optimiser halts where the objective stops improving. It does not halt where a person would stop, because the considerations that would stop a person are not in the objective unless someone put them there. Cost therefore falls until it meets something that is priced — capital, litigation exposure, a regulatory penalty, a wage floor with an enforcement mechanism behind it, reputational damage that actually appears in revenue. Everything else is not a floor being approached. It is simply outside the boundary, and an optimiser does not slow down as it nears something it cannot see.
So the answer to how low is: to the edge of what is represented, and no further, and no less. Push the boundary outward and the search stops earlier; leave it where it is and the search runs to it. This is a design fact about the objective, not a moral fact about the firm — which is why appeals to restraint have the poor track record they do, and why the effective interventions historically have been the ones that moved a cost inside the boundary rather than the ones that asked for forbearance.
Two consequences are worth stating because they are uncomfortable in different directions.
The operator has no exemption. Management is a specifiable function. Judgement exercised repeatedly against stable criteria is, by §2's own definition, a stable repeatable sequence, and the argument does not contain a clause protecting the people who commission it. The desk that could not name play was itself downstream of the same operator; the seat that removes seats is a seat. This is not a satisfying reversal and should not be enjoyed as one. It is the observation that the argument has no fixed point excluding its own author — the recorder in §1 specified his colleagues' function and, in the same act, the general method by which his own would later be specified.
The search undershoots its own floor. Whatever genuine floor exists is made of the unspecifiable: real novelty, judgement under uncertainty that is not merely uncertainty about a known distribution, the unassigned attention §2.3 calls slack. That is exactly the material the optimiser scores lowest, because it is heterogeneous, unmeasured, and indefensible line by line. The search therefore does not converge on the floor from above. It consumes the floor on the way down and keeps going, and the overshoot is invisible for §5A's reason — what is lost is a distribution of events that now do not happen, and non-events do not appear anywhere.
The case in §1 supplies the concrete terminus, and it is smaller than the argument might suggest documented. The residual human labour after Titan was not analysis. It was teaching clients to schedule the reports — weekly, monthly, quarterly. Scheduling was the last surviving act of judgement in the chain: everything upstream had been specified, and what remained was deciding when. Then that was written down too, as a recurrence rule.
The falsifiable version, for §7: if the account here is right, the terminal state of a specified function is not a smaller team but a schedule, and the observable signature is that the last role to disappear is the one that sets the cadence. If instead cadence-setting persists as durable employment across a range of specified functions, the claim is wrong in a way that matters.
The standard policy frame treats automation as a decision taken by management and imposed on labour. Protective instruments — consultation requirements, notice periods, redundancy process — are built on that assumption.
This case does not fit it, and the misfit is structural rather than incidental modeled.
No manager decided to eliminate the function. The sequence ran:
Every step is individually correct and locally rewarded. No party acts in bad faith. The eliminating decision was taken by the highest-performing worker and scaled by the organisation's ordinary reward mechanism — which means it cannot be regulated by constraining management, because management did not do it.
It also means the person best positioned to foresee the outcome is structurally the least likely to. From inside, the trajectory is not visible as a trajectory. It is a series of promotions.
⚠️ What this paper does not claim. The author cannot establish the aggregate effect on headcount, and does not attempt to. The relevant period includes the global financial crisis, which cut analyst populations across every institution for reasons wholly unrelated to any platform. Any attempt to attribute a share of reductions to this mechanism would be unidentified, and the corpus's own standard forbids it open. What is claimed is the mechanism, from inside it, by the person who operated it.
The operational value of the argument is that it relocates the leading indicator.
Task-routineness measures — the standard instrument for estimating automation exposure — score a function on how repetitive its constituent tasks appear. This case suggests they measure the wrong object, and measure it too late. Routineness is a property of the work; specification is an event, and it is observable.
This is testable on ordinary organisational data — internal role changes, process-harmonisation initiatives, template-consolidation projects — against subsequent headcount by function. It requires no access to model capabilities and no assumption about them. It predicts from the organisation's behaviour rather than the technology's.
A second, sharper form:
The campaign described in §1 has a property that ought to trouble anyone designing oversight for this class of change: it has no reviewable unit.
Governance operates on decisions of a certain size. A proposal to eliminate a function is reviewable. A budget line is reviewable. A system procurement is reviewable. A single analyst deciding to record a macro for one client's monthly report is none of those things, and no reasonable process would make it one — the review would cost more than the act.
Run that act several hundred times over several years and the function is specified out of existence, without a single reviewable decision having occurred anywhere in the sequence modeled.
This is the response-gap structure transposed. In credit, the borrower has no window in which to act because irreversibility arrives faster than response. Here, the institution has no window in which to deliberate, because no individual increment is large enough to trigger deliberation, and the aggregate never appears as an object at all. In both cases the harm is not concealed. It is sub-threshold, and the threshold was never set.
The tools here were a spreadsheet and a keystroke recorder, in 2007.
A recorder requires the sequence to be stable and demonstrable: the operator must be able to perform it once, identically, on command. That is a strong constraint, and it excludes a great deal of professional work that varies case to case even when its underlying logic does not.
Systems that infer a procedure from examples do not carry that constraint. They do not need the sequence to be stable, only recoverable from instances. That extends the exposed class from work that can be recorded to work that can be inferred — and the second set contains the first.
So this case should be read as the phenomenon in its most primitive available form. The mechanism is the same. The filter has been removed.
There is a further fact about the contracts of §1, and it is the most consequential observation in this paper.
The piecewise structure only became evident during the crisis documented.
The reason is structural rather than accidental. In ordinary conditions the measured performance on which these schedules operate clusters within a single band. One payout regime applies, continuously, for years. No boundary is approached, so no boundary is experienced — and a step function that is never evaluated near a step is indistinguishable, from the inside, from a straight line.
The thresholds were not hidden. They were written into the documents. They were simply never load-bearing until the distribution moved, and when it moved it moved into the tail, where the second regime fired and the parties discovered what they had agreed to.
This generalises, and the generalisation is the closing claim of the paper:
Every threshold in this corpus has this property. A site's carrying capacity is not visible while visitation stays below it. An aviation-fuel constraint is not visible while demand is small. A population's persistence threshold is not visible while feed is abundant. Δt in a credit contract is not visible while collateral stays comfortable. In every case the parameter is fully specified, fully in force, and completely undetectable by observation — until the moment it binds, at which point it has already acted modeled.
The governance consequence is severe. It means experience cannot discover thresholds, because experience is drawn from the regime in which they do not bind. Every year of incident-free operation is evidence about the interior of the range and no evidence whatever about the boundary. An institution with twenty years of clean history and an undisclosed step function has twenty years of data that is silent on the only question that matters.
There are exactly two remedies, and both appear in the companion paper's §8:
That is why verify, don't trust is not a temperament. Trust is calibrated on the interior. The damage is at the edge.
The observation above understates the failure, and the understatement should be corrected because the true version is the point of the paper.
When the crisis drove outcomes past the boundary, the difficulty was not that the parties had misunderstood which payout regime would apply. It was that the contract did not specify one documented. Beyond the fold there was no rule — not an unfavourable rule, not a surprising rule. No rule.
Formally: the agreement was a partial function. It was defined over the domain its drafters had imagined and undefined outside it, and the distribution moved outside it. At the fold the map is not locally invertible — the observed state is consistent with several allocations and determines none of them. Given what had happened, who was owed what percentage of what was not derivable from the document documented.
The consequence follows immediately and is the reason this matters beyond one desk in one bank:
The default occupant of that gap is the stronger party. More counsel, more capital, more capacity to wait — the ambiguity resolves in its favour without anything being breached, because there was nothing left to breach modeled. That is the sharpest form of the mechanism the companion paper describes in credit: there the borrower has a window too short to act in; here the parties have no rule to act under, at precisely the moment the amounts are largest.
The author is obliged to report that in this case it did not go that way, and the exception is more instructive than the rule.
The gap was filled by an analyst.
Asked what the split should be where the document did not say, the author recommended that it fall in favour of the client, against the bank's immediate interest, on the following ground: the client had placed assets in the institution's custody to be managed, and the institution therefore stood in a fiduciary relation to them. Where the agreement was silent, the silence should be resolved in favour of the party whose assets they were documented.
The executives agreed — and complained that they now had to be lawyers documented.
Three things follow, and they matter more than the anecdote.
First, the correction to the rule above. Power does not decide at a fold because power is entitled to; it decides because it is usually the only thing present. A void is filled by whoever occupies it and by whatever principle they bring into it. When a person with no bargaining power at all supplies a principle — and it is a recognised principle, load-bearing in law — it can hold, even against the interest of the party that could have overruled it documented. The mechanism is not "the strong take." It is "the unspecified is taken by the present." Those differ, and the difference is the entire space in which governance can act.
Second, the complaint is the governance failure stated by its victims. "We have to be lawyers now" is exactly right and should not be read as grumbling. When a contract goes undefined at the boundary, the drafting burden does not disappear — it is transferred, in real time, under crisis conditions, onto people who are not drafters and did not agree to draft. The specification work was not avoided by leaving the fold unwritten. It was deferred to the worst possible moment and assigned to the worst-placed people modeled.
Third, why the fold was there at all. The contracts had been written from the perspective of a single agent's exposure — the risk to the bank — on the implicit view that this was where risk presented. They did not model the client's risk, nor other stakeholders'. The document was therefore complete over the region where the bank's exposure was the binding constraint, and undefined where it was not modeled. This yields a general and testable claim:
If that holds, incompleteness is not simply an artefact of bounded foresight. It is shaped by whose exposure the drafter was modelling, which makes it an object of governance rather than an accident.
One further property of the episode deserves recording, because it is the sub-threshold failure of §4.1 recurring one level up.
The decisions taken at the fold did not go to a vote documented. There was no committee, no formal determination, no minuted process by which the institution decided how silence in a client agreement should be resolved. A recommendation was made and executives agreed with it.
So the failure compounds. The contract was undefined at the boundary — and the procedure for resolving an undefined contract was also undefined. The gap was filled, correctly as it happens, but by an informal act that produced no record of who decided, on what authority, or on what principle.
The consequences are structural rather than personal:
That is a governance instrument that does not currently exist in the companion paper's §8, and it should. Call it a fifth corrective: a disclosed resolution rule for undefined states — naming, in advance, who decides, under what standard, and with what record, when an agreement turns out not to cover the situation. It costs nothing to draft in calm conditions, and it is unwritable in a crisis.
Note what this does to the epistemics of the whole arrangement. During the years when the contract appeared to be governing, it was governing. The parties' confidence was correctly calibrated to their evidence. The document was doing exactly what they believed it was doing. Its failure was not latent misbehaviour waiting to be discovered — the failure did not exist yet. It was created by the arrival of a state the document did not cover, and it arrived at the worst possible moment, because the states a drafter fails to imagine are systematically the extreme ones.
⚠️ A terminological note. The word used here — the fold — is not borrowed for effect. It is the object: a surface turning back on itself such that the solution is locally non-unique or undefined, which is what a payout schedule does at an unspecified boundary. This author has treated folds formally elsewhere in this corpus (Book 6; Whitney A₁ classification; the catastrophe-manifold line). The mathematics came later. The object was met first, across a table, in 2008.
The Response Gap argues that a loss caused by a structural condition is routinely attributed to the person the condition acted upon, and that the attribution survives because the condition is not measured.
The same operation is visible here, in the language.
"He automated himself out of a job" is the sentence available for this case. It locates the cause in an individual and reads as a wry observation about personal miscalculation. It is also the industry's account of the whole phenomenon in miniature: cause assigned to the person, mechanism unnamed.
The mechanism was a specification event, taken locally, rewarded institutionally, and scaled by an ordinary promotion. Nobody at any point crossed a line, and nobody was watching the threshold — because the threshold had not been written down either.
That is the same failure the companion paper identifies in credit: the harm's unmeasurability is not incidental to it. It is the condition on which it persists.
A paper describing a mechanism of this kind invites the question of what it is worth, and the temptation to answer it with a large number should be resisted. This section records why.
The direct transfer is not the cost, and it largely nets out. Value reallocated at a fold is not destroyed; it accrues to the counterparty and compounds in their hands rather than the other's. Summing realised transfers across a class of contracts therefore measures redistribution, not loss, and overstates the social cost by roughly the whole of it.
Crisis-cost aggregates cannot be routed through this mechanism. Published estimates of what a financial crisis cost exist and are large. Attributing any share of them to contract incompleteness at the boundary requires an identification strategy this paper does not have and does not pretend to. The companion paper spends its §6 refusing exactly this move for the response gap — realised loss cannot be assigned to a mechanism merely because the mechanism was present — and the refusal binds here with equal force open.
And §5A forbids the estimate on its own terms. A quantity that manifests only beyond a threshold cannot be estimated from a regime that has not crossed one. This author has a single tail observation. To extrapolate a magnitude from it would be to occupy precisely the epistemic position of the executives in §5A.1 — confident about a boundary from evidence drawn entirely from the interior. The paper cannot make that error in its own voice while identifying it as the central failure in others'.
What can be said about magnitude is this, and it is a statement about direction and channel rather than size:
This inverts the usual policy arithmetic in a way worth stating plainly. The correctives — publishing formulas, bounding response times, disclosing epistemic status, and the resolution rule of §5A.3 — are conventionally weighed against the losses of the events they would have prevented. On the argument above they should be weighed against the standing premium on everything drafted afterwards, which is larger, permanent, and borne by parties who were never near a fold. Drafting a resolution rule in calm conditions costs a lawyer's afternoon. Not having one is priced into the entire book, indefinitely modeled.
The correction is worth recording because it was accurate about the institution and wrong about the method documented. Play is the precise term for what produces a working knowledge of the object model: unbounded exploration of a system's behaviour without a specified goal — record something, read what it emitted, step through it, record it a second way to see what changes. There is no other route on offer, and §2.1 sets out why: the recorder answers questions only if you ask it questions, and asking a system questions to see what it does is what play is.
The institution had no category for the activity. Not a prohibited category and not an approved one — none. Reaching for the nearest available word, it found one that classifies the activity as not-work, and corrected the vocabulary. The activity continued unchanged, because nothing in the correction touched it. Only the word was forbidden.
That is this section's thesis arriving as an anecdote rather than an argument, and it is the cleanest statement of it in the paper: the most consequential act on that desk was unnameable in the institution's own vocabulary. An organisation cannot govern what it cannot name, and it cannot name what its language marks as the opposite of work. The exchange is recorded without identifying the manager, who was speaking for the institution rather than for herself, and whose reading of the word was the conventional one.
An institution that will not let a manager engage a supplier without three competing quotations will let a small number of senior people resolve, informally and without record, a question whose consequences exceed that engagement by orders of magnitude. §5A.3 records one such instance. The asymmetry is not hypocrisy and it is worth stating precisely, because the explanation determines the remedy.
Oversight attaches to legibility, not to consequence. A procurement is legible: it has a price, a vendor, a date, a comparison class, and it fits on a form. A resolution at a fold has none of these. It has no unit, no tender, no natural register, and no obvious moment at which it becomes a decision rather than a conversation. Controls therefore bind on the shape of an act rather than its stakes — and the acts least resembling a purchase order attract the least scrutiny irrespective of what turns on them.
The consequence is systematic and inverts what governance is for:
This is §4.1 and §5A.3 restated as a general property. There, no increment was large enough to review and no resolution was formal enough to record. Here: nothing in the control apparatus is triggered by an act that does not resemble the acts the apparatus was built around.
The obvious remedy — refer such decisions to a committee — is insufficient on its own, and the insufficiency is instructive. Adding signatories to an unspecified decision does not specify it. It distributes responsibility for an arbitrary determination across more people, which improves the record and the politics while leaving the epistemics untouched. Most governance reform fails at exactly this point: it adds process without adding anything checkable.
What would make such a committee more than ceremonial is a decidable question it is obliged to ask. The verification discipline supplies one, and it is narrower than it first appears.
⚠️ A contract's fairness is not formally checkable. Neither is the choice of principle at a fold — the fiduciary reasoning of §5A.2 is a normative judgement and no proof system will adjudicate it. Claiming otherwise would be the overreach this corpus exists to avoid.
A contract's totality is checkable. Whether a payout function is defined over the whole domain the parties are exposed to is a decidable property, not a matter of opinion. It requires only that the schedule be stated as a function and the exposure be stated as a domain — after which coverage either holds or a gap can be exhibited. This is the ordinary totality obligation a proof assistant imposes on any definition by cases, and it is the exact failure of §5A.1: a partial function presented as a complete agreement.
The virtues are practical rather than theoretical. The obligation is cheap in calm conditions and impossible in a crisis. It produces an artefact — a stated domain — which is reviewable in a way a conversation is not. It is falsifiable by counterexample: a single unhandled state defeats it, and exhibiting one requires no authority, only attention. And it makes the failure of 2008 a finding rather than a surprise, since a totality check performed at drafting would have returned the gap that the crisis later returned at far greater cost modeled.
This is the sense in which verify, don't trust generalises beyond mathematics. It does not mean subjecting judgement to proof. It means never letting an unstated domain pass as a complete specification.
The natural objection is that these agreements were drafted by specialists, and that coverage was therefore somebody's job. The record suggests otherwise, and the reason is structural rather than a failure of diligence.
The specialists present were lawyers, and legal training addresses text. Whether language is ambiguous, whether terms are defined, whether obligations are enforceable, whether a clause survives challenge — these are demanding questions and they were competently handled. None of them is the question whether a piecewise function is total over the reachable state space. That is a property of the object the document describes, not of the document, and no amount of drafting skill surfaces it.
So the question fell between two professions and was owned by neither. The legal side could reasonably assume the commercial side had characterised the exposure; the commercial side could reasonably assume the legal side had confirmed the instrument was complete. Both assumptions were held in good faith and neither was anyone's stated responsibility modeled.
This explains the complaint recorded in §5A.2 — "we have to be lawyers now" — and shows it to be the wrong inference drawn at the right moment. More legal attention would not have helped. What the fold revealed was not a defect in the drafting but a question that had never been assigned.
It also explains the observational position from which the defect was visible. The author was neither drafting nor negotiating these agreements. He was performing attribution, which requires evaluating the function on realised states in order to produce a number documented.
The governance implication is narrow and cheap: the totality obligation of §6B.1 should be discharged by whoever will have to evaluate the instrument, not by whoever drafts it — and it should be discharged before execution rather than at the first hard case, which is the only difference between a finding and a crisis.
© 2026 Pablo Nogueira Grossi. G6 LLC, Newark, New Jersey, USA. Licensed CC BY 4.0 — https://creativecommons.org/licenses/by/4.0/ ORCID 0009-0000-6496-2186 · Code and formalisation: https://github.com/TOTOGT
The argument of §2.4 — that appeals to self-regulation read teleology backwards off a survivorship artifact, and that self-regulation is dimension-specific — has since been used in two directions.
Outward. WP66 §7.1 applies it to the Amazon. The basin genuinely is self-regulating in the moisture dimension: evapotranspiration recycles the burden, which sustains the forest, which sustains the recycling — which is why “nature will fix itself” feels right. But it damps only inside a threshold, and clearing is a different loop with a different sign and no restoring force, for the structural reason this paper identifies in the labour case: the party being removed is not a counterparty to the transaction. Critical slowing down, measured across 76% of grid cells, is that damping term being observed on its way out.
Inward. WP68 §8.1 turns the same argument against a paper of this series. WP68 claims that what fails its four tests is what fossilises — a claim assembled entirely from a sample of things that fossilised. The local, reversible activities that left no record cannot be enumerated, because leaving no record is what they did. The filter is therefore at risk of being constructed by its own sample, and WP68 marks the definitional half safe and the empirical half exposed.
First use of §2.4 as an internal check on the corpus rather than an external one.