How cost accounting works

The problem: an LLM platform’s unit economics live or die on one number — what does one published article actually cost? — and that number is surprisingly hard to get. Flat per-call estimates drift from reality (here, a flat constant once under-counted the platform’s dominant cost driver by ~6×); provider dashboards aggregate by account, not by article.

Layer 1 — the immutable usage event

Every run emits one immutable usage event — broken down by model, with per-step token deltas kept alongside the run — carrying the full attribution vector: tenant, contract, record, pipeline, step, model, token counts, timestamp. Not a log line — a queryable ledger row. Nothing about cost is derived at call time; events store facts (tokens, model), and money is computed later.

Layer 2 — the versioned pricebook

Prices change; history must not. Cost is computed by joining events against a versioned pricebook — the price valid at event time — so last month’s numbers stay true after a provider reprices. The same join reconciles the computed total against the provider’s actual invoice: if the ledger and the bill diverge, something is unmetered, and that is a defect, not a rounding note.

Layer 3 — the cost of one article, amortized

“Cost per article” includes more than its Generate run: Collect research that fed it is amortized across the articles it enabled. This is the number the platform actually steers by — it exposed, for example, that a period of post-gate rejects multiplied the effective cost per published article several-fold even though per-call prices never moved. Without per-article attribution that regression is invisible inside a monthly total.

Layer 4 — budgets (designed, not yet armed)

Metering is meant to feed enforcement: per-run budgets with a watcher that would stop a breaching tenant’s runs through the run control channel. The enforcement stage is not enabled yet: no budgets are configured, and the control channel the watcher would use is the one the run engine documents as not interrupting a run mid-flight. There is deliberately no automatic provider-level kill switch either — an automatic billing cut-off takes production down with the runaway it was meant to stop — so the working control today is a daily spend-anomaly signal in the nightly report: it notifies the owner, and a human decides. This page will change when budgets are configured and the watcher is measured.

Trade-offs

Metering every call costs a write per call and a discipline: any new LLM call path must emit the event, and the reconciliation check exists precisely to catch paths that forget. Amortization rules are a modeling choice — the platform records the raw vectors so the model can be changed retroactively without losing data. Absolute figures — per article, per night, per month — are deliberately not published on this portal: the mechanism is public, the ledger is not.

See it in two minutes

The metering domain in list_domains exposes the read surface; per-run costs appear in run reports.