Cost

The question this answers: where the money and the time actually go, and which of the two decides the schedule.

This page states the shape of the cost and links to the documents that fix the figures. It quotes no number that a ruling owns, because a number repeated in a guide page is a second source of truth that no gate watches — and this project already has one recorded case of an unchecked claim being quoted back as though it were measured.

Three buckets, and only one of them matters

Bucket	What is in it	Governed by
Human Curation	Reading each SDS, selecting spans, verifying extracted text against the rendered page, reconciling restrictions, writing Curated Translations	ADR-0028, ADR-0034
Provider tokens	The Query Intent tier, which reads a question into structured intent	ADR-0033
Infrastructure	A Bun server, a Postgres database, static hosting for a precached shell	ADR-0020

Curation dominates, by orders of magnitude, and it is not close. ADR-0028 measures the effort per document and ADR-0034 fixes what multiplies it. Its conclusion — corpus size is the schedule — is the single most load-bearing cost statement in the project, and it is a statement about person-days rather than about money. Curation works through why the expensive part is review rather than extraction.

The other two buckets are rounding errors against it. That is not an argument for ignoring them: the provider bucket is the only one that can run away on its own, which is why it has a cap.

Why the provider bucket has a cap and not a forecast

The tier sends a question and receives structured intent — a short input and a small structured output, per question asked, on a surface that is not on any safety path. ADR-0033 records the model, the tier, the prices it read and the date it read them, and prices it records as rising on 2027-01-01.

Two things about that bucket are not settled and are worth knowing before quoting anything:

• The cap is a number the owner sets (D-77) and it is not set. The endpoint is unbuilt, and which fact that waits on is itself a ruling: ADR-0055 puts it after the Study Area's Corpus, so that the price, the cap and the per-address limit are taken from that day's facts rather than from prices read months earlier.
• The per-address limit is a fairness guard, not a spend control (ADR-0043), and D-103 requires it to be sized for a shared egress, because The Plant's devices sit behind one NAT. A limit tuned as though one address meant one person locks out the site.

Seam 5 exercises this tier with recorded intent rather than a live provider, so the test suite costs nothing to run and the fixture does not depend on the provider having a good day.

The bucket that turned out to be misfiled

Five Curation artifacts were each introduced as costing something once per Corpus — outside the per-document figure and outside corpus size. The sizing note added them up and found that three of the five scale with corpus size after all, so ADR-0028's ruling covers more of the cost than the label suggested, not less.

That note deliberately states no hours. No spike has been run for any of the five artifacts, and a figure with no measurement behind it gets quoted later as though it had one. What a measurement would take is written down there instead.

Two of the five are also gated on people this project does not employ — a site authority for the Curated Escalation, and someone who knows how Handlers phrase an emergency in Thai for the Escalation Trigger — so they belong with the blockers, not with a schedule.

One cost of the same shape is not Curation at all and is easy to file wrongly. The sixteen GHS Section headings are not curated per document — ADR-0060 says why — so they cost nothing per document, and the Thai half of them is at present uncited. Closing that is a /research pass against a primary source, once, for the whole project. What it is and why it is open is in Sources & data.

What the platform decision cost

ADR-0020 chose the web stack over a native Android build and carried the two costs its superseded predecessor predicted into its own text rather than discarding them: what the platform costs camera Identification, and what it costs offline storage. Both are still live, and both appear on the blocker list rather than in a budget.