The Arbitrage Has a Twin

The Slowest Interface in the Market closed the market-computation chapter by noticing that five memos had made every participant in a market fast except the Steward, and that nobody had specified the Steward's channel. Four memos have since specified it. The Human and the Agent Are Fast in Opposite Directions named the mismatch. The Steward Speaks. The Steward Sees. The System Reads. named the rule. The Steward Was Not Slow. The Steward Was Blind. named the failure and its path to a false autonomy score. Heard Is Not Meant named the commit rule at the irreversible gate. Each is a snapshot: the parties as they are, the route as it should be, the failure as it presents.

This memo is about the trend. The gap #124 named is not a fixed distance. It moves, in one direction, for a reason the corpus already owns — and it moves for the same reason Arco's economic thesis works. That is the uncomfortable part of the synthesis, and the reason the chapter needs one.

The quantity

Two rates describe the relationship between an agentic system and its Steward. The first is the rate at which the system changes its own state: commitments made, allocations recomputed, escalations raised, actions taken in the physical world. The second is the rate at which the Steward can take in that state and put a decision back — the perception-and-response cycle that runs across the verification and command routes of the Transcoding Layer. The distance between them has a name.

The distance between the rate at which an agentic system changes its own state and the rate at which its Steward can perceive that state and act on it — a gap that widens as execution tempo rises, because the agent's side of Bandwidth Inversion moves in release cycles while the Steward's moves in generations, and whose size at any moment sets how much of the system remains governable without rerouting.

That is the Governance Tempo Gap. It is a property of the pair, not of either party, and it has a size at any given moment: decision-relevant state changes per Steward perception-and-response cycle on the verification route. Decision-relevant is the load-bearing qualifier — a mutation the Steward would never act on does not count, or the ratio explodes on every write. A system that changes state once a minute and a Steward whose cycle is thirty seconds has a gap below one, and the Steward sees the system as it is. A system that recomputes an allocation on every relevant change across every participant, and the same Steward, has a gap in the hundreds, and the Steward sees a system that has already moved. Nothing about the Steward changed between the two cases. The system did.

Why it moves one way

The direction comes from the asymmetry #125 established and this chapter has leaned on since. Bandwidth Inversion is asymmetric in rate: the agent's efficiency profile moves in release cycles; the Steward's moves in generations. Every model release that lets agents change state faster widens the numerator of the gap. The denominator does not compound on that cycle. Left alone, the ratio rises with each release, and a Steward surface that was adequate in the spring is producing Verification Blindness by the autumn without a single line of it having changed.

This corpus has one other quantity with exactly that shape.

The cost and output delta between a human-staffed operation and an equivalent agentic operation, widening over time as AI costs fall and human costs rise.

Operational Arbitrage is the founding economic claim of the studio: the margin an autonomous business captures grows because one side of the delta improves on a technology curve and the other is a human cost that does not. The Governance Tempo Gap is the same curve read from the other end. The arbitrage has two halves — the agent's cost falls, and the agent's output rises — and the gap rides the second. A release that only makes agents cheaper widens the margin and leaves the gap where it was. A release that lets agents act faster widens both, because a higher rate of action is a higher rate of state change against a Steward whose perception rate is fixed. Arco's thesis has always been that the output half is where the arbitrage compounds, and that is exactly the half that moves the gap. The arbitrage has a twin, and the twin is a liability.

Naming that is the point of the synthesis. A studio that tracks the arbitrage and not the gap is measuring the upside of a curve while ignoring that the same curve is moving its Stewards toward blindness. The Stewardship Model captures the arbitrage by keeping a human above the Execution Layer at a fraction of the cost of the workforce the agents replace. It only keeps capturing it if that human can still see the system.

What the gap does at each size

The gap is a ratio, and the chapter's terms sit along its range.

At a low ratio the Steward's picture is the system's state. Interventions land on the condition they were aimed at, the routes work as designed, and the Stewardship Model does what its definition says. As the ratio rises the picture lags further behind the system, continuously, with no step in it.

Verification Blindness is the point on that slope where the lag is material to the decision the Steward is making. #127 traced what follows: interventions arrive late or wrong, the Steward correctly learns the surface is unreliable, they stop consulting it, and Nominal MTTI records their withdrawal as autonomy. The gap is the variable that moves a Steward from the first state to the second without anyone redesigning anything.

At the Physical Intervention Threshold the gap becomes a safety quantity. The threshold escalates irreversible actions to a human who must see current state and answer; a wide gap means the state they see has moved by the time they answer. Round-Trip Commit renders the change before it is made, and #128 was explicit that the rule depends on the verification route being current — which is to say, on the gap being narrow enough for the rendering to be true when the Steward confirms it.

And at the top of the scale sit the two conditions the market chapter projected. Market Computation is an allocation rule recomputing across every participant on every relevant change. The Executable Economy is signals translating directly into action with no planning cycle between. When the Market Decides What Gets Made placed the Steward above the allocation rule — a second stewardship surface at market scale — and this is where that placement's cost becomes visible. At that altitude the Steward is no longer governing one operation's state; they are governing the consequences of a rule that fires across the whole participant set, and the aggregation is what makes the numerator largest. The Steward there sits at the widest point of the gap the whole system produces. That was the residual #124 named. This chapter's answer is not that the gap can be made small at that altitude. It is that the gap cannot be closed by making the Steward faster, and is made governable by changing what reaches the Steward.

The half of this that was said in 1983

Lisanne Bainbridge's Ironies of Automation ran to five pages in Automatica. Barry Strauch's 2018 review of its legacy counted roughly 1,800 citations by late 2016, still accelerating, and titled itself Still Unresolved After All These Years. The paper's argument is the static form of this memo's. Automating most of a process does not eliminate the problem of the human operator; it expands it. The operator is left with the tasks that could not be automated — monitoring, and the rare crucial intervention — and is therefore required to be more skilled and less loaded than before, at the moment the design treats them as the unreliable component to be minimised. The plant runs itself until it does not, and then it needs the one person the design forgot to equip.

Bainbridge's operator faced a plant that changed at plant speed, and plant speed did not change across the operator's career. The ironies were structural but they were not a trend. What the agentic system adds is the tempo term: the process the Steward monitors gets faster every release, so the irony that was static in 1983 has a slope in 2026. The residual human task does not merely stay hard as automation deepens; the object of that task accelerates away from the human's fixed perception rate, on a schedule set by model releases. Bainbridge described the gap. This chapter's claim is narrower than hers and adds one thing: the gap widens, the widening is driven by the same variable as the arbitrage, and the correction has a location.

Closed from one side

The Steward cannot be made to perceive faster than a human perceives. Adding Stewards increases capacity by headcount, adds Coordination Tax on the split, and reintroduces the human cost base whose removal created the arbitrage — while the agent-side tempo keeps compounding with each release. Slowing the system to the Steward's tempo forfeits the arbitrage directly. None of the three moves on the Steward's side of the gap is available to a studio that intends to keep the margin.

The move that is available changes what the Steward has to perceive rather than how fast they perceive it. That is the Transcoding Layer, read as a trend rather than a snapshot. Route state to vision, so the Steward's fastest input is the one in use. Compress state to what the governance task requires, so the fixed perception rate is spent on the changes that carry a decision and not on the trail. Route intent through speech, so the return half of the cycle runs at the Steward's fastest output. Render the irreversible change before it is made, so the one confirmation that cannot be undone is made against current state. Each of these lowers the number of things the Steward must take in per unit of system change. None of them touches the numerator. Together they are the lever Arco can pull without surrendering the agent-side tempo that creates the arbitrage.

The consequence for how Arco works is procedural and it is the chapter's operational residue. Under a snapshot view, Steward-facing routes are specified at Full-System Design and are done. Under the trend view, every model release that widens the gap on the agent's side is a release that reopens the Steward's routes for review. The Transcoding Layer is not a design-time deliverable. It is a maintenance obligation that recurs on the model release cycle, because that is the cycle the gap moves on.

What the chapter adds to the Stewardship Model

The Stewardship Model was defined by who does what: agents execute, the Steward governs and handles exceptions. Five memos have added how the two reach each other, and what happens when they do not. Bandwidth Inversion says the parties have opposite fast channels, so the interface has to translate. The Transcoding Layer says how: route by receiver, system carries the translation. Verification Blindness says what the failure is and why it becomes a false score. Round-Trip Commit says what must happen at the one gate where a wrong action cannot be undone. The Governance Tempo Gap says the whole problem gets harder on a schedule, and which side it is fixed from.

One residual from #124 remains open and this chapter does not touch it: Arco still has no term for the economic position of owning the allocation rule an entire market runs through. That is a market-structure question, not an interface one, and it waits for its own chapter.

The Operator's Verdict

An operator running the Stewardship Model has two curves to watch and has been watching one. The arbitrage is the margin the agents produce against the workforce they replaced. The gap is the number of times the system changes state for every cycle in which the Steward can see it and act. They rise together, off the same releases.

The check is the ratio, taken on the verification route, per release. If it rises and the Steward's surface has not been revisited, the surface is failing on a schedule the operator can read off the model provider's changelog, and the first symptom will be an autonomy metric that improves. The correction is not a faster Steward or a slower system. It is a rerouting, and it is due every time the numerator moves.

Technology changes what is possible. Tempo determines what remains governable.