Silencing the Span

At a glance

Every figure below is read from the client's own per-request log. The method, the caveats and the things this page got wrong are underneath.

Which models did the work

Where the money went

What this investigation cost

From tokens to public infrastructure

Why a study about bridge noise carries a page about its own consumption.

Every infrastructure system becomes governable at the moment it becomes measurable. Electricity became governable through the kilowatt-hour. Telecommunications became governable through the minute and the megabit. Cloud computing became governable through the CPU-hour and the gigabyte of egress. The unit is never only a technical detail — it is the language in which allocation, pricing and accountability can be argued at all.

Machine inference is entering that stage now, and the unit is published. GitHub Docs states it directly:

Each token is priced based on the model used, and the total is converted into AI credits, where 1 AI credit = $0.01 USD. GitHub. (2026). Usage-based billing for organizations and enterprises. GitHub Docs. Retrieved 4 August 2026. VERIFIED 5/5

That sentence is what makes this page possible. It is also why the page sits in a research repository about a bridge rather than in a billing tool. This programme's central finding about DUMBO is that the noise is unmeasured and therefore ungoverned — New York City's 311 system has no category that can accept a rail-noise complaint, so the exposure is never counted, so it is never budgeted against. Applying that same argument to this project's own production is what produced this directory. A programme that argues measurement precedes governance, and then declines to measure itself, is making an exception for the one party it controls.

The public term is AI credits. The field this dashboard actually reads is nano_aiu, a telemetry subunit that appears in the local session store and on no public billing page. It is used here because it is what the data contains; everything quotable is converted to credits and to dollars. Anyone citing this work should cite credits.

The AI age will not only be defined by who has access to models. It will also be defined by who can afford inference, who can measure it, who can govern it, and who benefits from the systems built on top of it.

The repository this page lives in, —, was produced for of list-price inference by one person who did not have to ask anyone for the budget. Both halves of that sentence are the finding. The ledger below is the first half. The methods note carries the second, with the full citations and the ratings.

The ledger

Every request that was billed, by model and by channel. Read from token_details_json, which reconciles to the cost the client recorded on — not from the token columns beside it, which do not.

Where it actually went

Requests are grouped by what caused them. This is the panel that changed how the work reads.

By initiator

By reasoning effort

The cache is doing most of the work

Same traffic, same models, every cached token repriced at its model's full input rate.

What this is not. It is not a saving anyone banked. Without a cache the agent would have been driven differently — smaller context, fewer re-reads, more summarising — so the uncached column prices a workload that would never have been run in that shape. It bounds the cache's contribution; it does not measure a decision.

Three honest answers to “how many hours”

They differ by an order of magnitude, and each one is measuring something real. The question does not have a single answer.

Active time, by how long a pause has to be before it stops counting

Day by day

model generating person — residual, not measured

The rest of the project

Cost per turn

One bar per thing that was asked for. The label is the opening of the request as it was typed.

Models, and sub-agents

What was offered against what was used

Sub-agents

Energy

The other half of the ledger

Volume is not value. Every figure above is a count of output, and this project’s own record argues against reading counts as worth: several of its most useful moments were withdrawals, which reduce the word count and improve the work. A ratio of dollars to words prices typing. Nothing here prices whether the study is any good.

What is not measured

Named, so that the gaps are as visible as the numbers.

What would have to exist next time

The point of a first case study is to specify the instrument. These are the fields that would close the gaps above, in the order they would pay off.

    Where this dashboard is likely to be wrong

    Process note: a claim this page made and withdrew

    Kept at the bottom because it is a note about method, not a number anyone needs before reading the ledger. It is here at all because deleting it would make the page look more certain than the work was.

    That is withdrawn.

    Why it is still on the page. The claim was three minutes from being written up as the finding of the case study — that agentic work is unmeasurable because nobody records it. It was wrong, and it was wrong in the flattering direction: an absence of evidence is a much easier thing to publish than a ledger that can be checked. The correction stays visible for the same reason every other withdrawal in this project does.