Another Ledgerline

Thursday, Finn shows Dana where the assistant actually lives. The tour is one diagram long, and the reveal is that there is no reveal. It lives nowhere in the building.

Every answer is a call to an AI APIConcept · lights on your mapAI APIA provider’s endpoint that answers prompts. The backend sends the assembled prompt as an HTTPS POST, JSON in, JSON out, API key in the header, status codes, timeouts, retries. The oldest trace in this story, with a model at the far end., and you have read this exact diagram before. It is Case’s registration POST from Act I with a different address on the envelope: request out, response back, headers, status codes, a timeout, a retry policy. The magic box is a vendor with an endpoint.

The newest costume

Which means the relationship is one Traversal already knows by heart. AI as a third-party serviceConcept · lights on your mapAI as a third-party serviceThe model provider is the third-party service in its newest costume: rented capability, a dependency with a contract, versioned models, pricing changes, outages that become your outages. Every dynamic from the API economy, unchanged. is the thesis of the whole week. Payments are Ledgerline. Notifications are Semaphore. Intelligence, it turns out, is another line on the same vendor page: a contract, model versions that deprecate like any API’s, prices that change, and an outage that becomes Traversal’s outage at the worst possible minute. The oldest rule in Act I’s vendor catalogue applies unchanged. Every third-party service is a component you depend on and do not control.

Physics and the meter

Two properties make this vendor stranger than the others. First, AI latencyConcept · lights on your mapAI latencyInference in a model that talks takes seconds, not milliseconds, which is why answers stream in as they generate. When the real number will not budge, perceived latency becomes the UX lever.: the mat’s whole half-second budget is spent forty times over on one assistant answer, and no amount of engineering makes a large language model answer in 38 milliseconds. Streaming is the armistice: show fragments as they come, and the wait feels like typing instead of silence. Second, AI costConcept · lights on your mapAI costEvery token is priced, in and out, on every request: a per-request meter that scales with usage. Prompt size becomes a cost decision, not only a quality one.: Monday’s scaffolding, re-sent on every tap, is a meter running in both directions, and the RAG passages joined the bill on Wednesday. Somebody at Traversal now reviews prompt length the way somebody reviews the cloud invoice, and it is the same somebody.

And one property makes the audit’s ears prick up retroactively: AI privacy considerationsConcept · lights on your mapAI privacy considerationsThe data-protection questions pointed at a new flow: what leaves your walls inside prompts, what the provider may retain or train on, and which regulations follow the data. “What is in the prompt?” is the new “what is in the export?”. Case’s question left the building inside a prompt. So does whatever a panicked runner types at midnight, medical details and all. Friday’s report already asked what Traversal keeps; the same pen now asks what the model vendor keeps, for how long, and under whose law. The prompt is an export, and it gets the export’s scrutiny.

And that is the last box. Payments, messages, timing hardware, intelligence: rented, contracted, understood. On Friday, Finn uncaps a marker, and this story draws itself for the last time.