The Last Unmetered Seat

The Cost Attribution Layer is the architectural component that traces the operational cost of every agentic step, task class, and job-to-be-done against the business's v0 baseline — enabling the Steward to evaluate the economic impact of each architectural optimisation decision with precision, rather than measuring compounding effect by impression or total spend. The Audit Surface Problem introduced it as the economic component of the governance digest, sitting alongside the Escalation Rate summaries that tell a Steward whether each task class is operating within its defined band. It is the most precise cost instrument this body of work has specified, and it has a boundary that has not been examined: it traces every agentic step.

The Steward is not an agentic step.

When a Steward resolves an exception, calibrates an Intervention Threshold, or works through a novel condition the Exception Architecture — the designed protocol governing what happens when an agent encounters an operational state not present in its knowledge layer — has not yet encoded, that work increasingly happens through inference. Not the system's inference — the Steward's own. A model consulted, a document reasoned over, a decision drafted and pressure-tested before it is committed. If that inference runs through a subscription seat at a model provider, it produces real cost and real emissions attached to no token count, no task class, and no traced path. The Proof of Action trail records the Steward's intervention in the same ledger as the system's autonomous decisions; that is by design and it works. What it records is the decision. The compute that produced the decision is somewhere else entirely, on an invoice with one number on it.

This is not a small gap in a large system. It is a specifically shaped gap: the only part of an autonomous business that is not traced is the only part that is human. That is the Judgment Layer — the set of decisions requiring genuine human assessment, owned by the Steward under the Stewardship Model, the operating structure in which a single named operator governs exceptions rather than a rotating team — sitting against the Execution Layer, which the Cost Attribution Layer already traces completely. The layers are supposed to be separable precisely so each can be governed on its own terms. Tracing one and not the other is not a failure of that separation. It is a failure to notice that separation cuts both ways.

Why a subscription seat cannot be reconciled

The instinct is to treat this as a data-access problem — the numbers exist somewhere, and with enough effort they can be retrieved. They cannot, and the reason is structural rather than administrative.

A subscription seat is a fixed cost with an unknown denominator. The price is published. The consumption is not. What the interface exposes is a proportion of an allowance whose absolute size is undisclosed, banded by time window and by model, resetting on a schedule that does not align to a reporting period. A Steward can observe that they are seventy per cent through a window. They cannot convert that observation into tokens, into a cost, into an emission, or into an attribution against the task class that consumed it. Cost per task is therefore not merely expensive to calculate. It is uncomputable from the data the provider exposes.

Note what this memo is not claiming. It is not claiming that providers designed this opacity to prevent customers comparing subscription and API pricing for the same work. That is a claim about intent, it is not evidenced, and this body of work does not need it. The observable effect is sufficient and it is not disputed by anyone: the data a subscription interface exposes does not support per-task reconciliation, and no amount of diligence on the customer's side changes that.

The failure mode has a name in this body of work already, one domain over. Carbon Data Is Operational State established that a figure reconstructed after the fact from invoices has no chain back to a specific event, and that a counterparty asking a specific follow-up question can tell the difference immediately. A subscription seat is that condition in its purest form: a monthly invoice with no events behind it at all. The annual sustainability report at least has estimation methodology. The seat has a percentage bar.

There is a second structural resemblance worth naming. Nominal MTTI is the condition in which a system's measured Mean Time to Intervention is long not because it has achieved genuine Architectural Certainty but because the Steward has stopped engaging with the Audit Surface — the structured governance digest derived from the Proof of Action trail and designed for Steward comprehensibility at operational tempo — producing a metric that appears to confirm autonomous operation while the system is in fact running unmonitored and unrefined. The mechanism behind it is push-based governance: the Steward must actively go and look in order to detect anything, and eventually stops. A subscription seat is push-based by construction. Nothing about consumption surfaces to the Steward; the Steward must open a settings page and interpret a bar. The predictable result is the same as the one already documented for governance — nobody looks, and the absence of alarm is mistaken for the absence of a problem.

Three consequences, in order of proximity

The carbon consequence is the one that prompted this memo, and it is the third of three. The first two are closer to daily operation and matter to a Steward who has no reporting exposure at all.

Architecture cost drift is measured against an incomplete base. The Cost Attribution Layer exists to let a Steward evaluate optimisation decisions with precision rather than impression. If a category of operational cost sits entirely outside the traced path, the precision is bounded by the size of that category, and the size is unknown. A Steward comparing a v0 baseline to current cost is comparing two figures that both exclude the same unmeasured quantity — which is survivable when the quantity is stable, and misleading the moment it is not.

Intelligence Arbitrage cannot reach the task classes the Steward handles. Routing each task class to the cheapest model capable of executing it at the required quality level requires knowing what each task class consumes. The task classes that escalate to the Steward are, by definition, the ones the Exception Architecture has not yet encoded — which is to say, the ones most likely to become encoded workflows later. They are the pipeline of future automation, and they are the only classes with no consumption data attached. The routing decision is unavailable precisely where the next architectural improvement will be made.

Revenue per Carbon has an unmetered residual. The ratio of revenue generated to kilograms of CO2-equivalent emitted by the inference and supporting infrastructure that produced it, measured per period at the level of the whole business — that is the metric, and whole business is the operative phrase. The Cleanest Asset in the Portfolio made a narrow and defensible claim: not that an autonomous business is greener, but that its footprint is provable rather than estimated. A residual that is structurally unprovable sits inside that claim. The residual is probably small. Small and unprovable is still unprovable, and the entire argument was about provability.

The third value of BYOK

BYOK as Architectural Decoupling named two values and ranked them, using the term in its portability sense — model and provider independence — distinct from the canonical Lexicon meaning of the same phrase, which is a business's structural independence from its founders. Cost control: accurate but insufficient. The structural argument, in this memo's sense: the condition that makes Intelligence Arbitrage available, the difference between owning the logic and renaming the vendor on the invoice. That memo was written entirely at the agent execution layer. The Steward does not appear in it once.

There is a third value, and it is the one this memo needs: attribution.

A key held by the business, routed per identity, produces a traced record of every call regardless of who or what made it. Agent execution and Steward execution land in the same path, distinguished by a usage-class tag rather than separated by a billing boundary. The Cost Attribution Layer gains the dimension it was missing. The Proof of Action trail can now record not only that the Steward intervened but what the intervention consumed. Revenue per Carbon's denominator closes.

The construction from the earlier memo extends directly. BYOK without a portable Context Architecture is a pricing feature — the routing changes, the dependency does not. BYOK without per-identity attribution is a portability feature — the vendor relationship is neutral, the ledger is still incomplete. Both are real gains. Neither is the whole of what the mechanism makes available.

The surface the Steward works through

The Audit Surface Problem specified two artefacts derived from the same underlying record: the Proof of Action trail, structured for auditor replay, and the governance digest, structured for a five-minute operational review. Both are reading surfaces. Neither is a surface through which work is performed.

That omission was correct for the argument being made at the time, and it is the remaining gap now. If a Steward's inference must run through org-owned keys to be attributable, it has to run through something. A subscription seat is the default because it requires no engineering decision, which is exactly why it becomes the architecture nobody chose.

This is where an apparent contradiction with earlier work has to be met rather than avoided. The Death of the Seat License named the UI Tax — the cost premium embedded in software designed for human cognition — and argued it is entirely redundant in an agentic stack where agents require no interface to operate. That remains true, and it is not in tension with the position here. The UI Tax is redundant for agents and load-bearing for the one human left. An autonomous business does not eliminate the interface requirement; it reduces it to a single operator, which is what makes building a purpose-shaped one tractable rather than a general-purpose product problem. What was a per-seat cost across a workforce becomes a one-seat design decision — and unlike a purchased seat, it is replicable across every business in the portfolio at the cost of the first build.

The requirement it implies is bounded and should be stated as such, because an argument that ends in a new department contradicts the reason this architecture is worth building. Three things, none of them new instrumentation: keys routed per identity rather than per subscription; a usage-class tag on the existing Cost Attribution Layer trace, distinguishing system execution from Steward execution; and the append-only discipline that Carbon Data Is Operational State already specified, applied to the new class of record rather than to a new system. One additional dimension on data that is already being captured, for the calls that are currently not being captured at all.

The Operator's Verdict

The argument for closing this gap now is not that the residual is large. It is almost certainly not large today, and a memo claiming otherwise would be inventing a number to justify an architecture. The argument is that a business whose entire measurement claim rests on provability cannot carry a category of consumption that is structurally unprovable, and that the cost of closing it while the business is small is a routing decision, while the cost of closing it later is a reconstruction exercise across every period already reported. That is the same distinction this body of work has drawn for the Accountability Trace, for License Grade, and for the carbon record itself: inexpensive to build in, expensive to retrofit, and the point at which it becomes required is not the point at which it becomes cheap.

There is a smaller and more uncomfortable version of the same point. An autonomous business that has traced every agentic step and left its Steward on an unmetered seat has not built an incomplete ledger by accident. It has built one that stops precisely where the human starts, which is the one boundary the architecture was supposed to make visible.

Technology meters what the system spends. Attribution determines whose spending is in the ledger.