cost declaration on atproto (first-principles sketch)

The bigger idea behind the cost-tracking system: a standard atproto protocol for declaring infrastructure costs — private or public — so any project/actor can say "running this costs $X, here's how I counted." The working system at infra-cost-optimization is a v0 reference implementation under a personal NSID; this note is the design sketch for generalizing it.

why this is interesting

atproto is full of public-good infra people run out of pocket (relays, appviews, labelers, feed gens, PDSes). There's no standard way to declare what that costs. A shared protocol enables: transparent grant solicitation, ecosystem-level "what does public atproto cost to run?" aggregation, and honest internal tracking — all on records anyone can read/verify.

the crux: how do you count?

Cost is not in application code — it lives in infrastructure + provider billing. Fidelity ladder, and every figure must carry its provenance (basis):

1. billed — from a provider billing/invoice API. Authoritative but rare: the APIs are partial and gated (Cloudflare 403'd us, Fly has no real billing API, Hetzner has no invoice endpoint).
2. derived-from-IaC — the interesting one. If infra is declarative (fly.toml, k8s manifests, wrangler.jsonc, Terraform/Pulumi), the billable units are IN the repo: machine types, counts, plan tiers. Cost = declared resources × a published price book. Reproducible, version-controlled, lives next to the thing it bills for. (Our Hetzner connector is basically this: server type → list price.)
3. estimated — inventory × rates when you can't get the declaration (our Fly connector).

The strong version of a public cost claim lets a reader recompute: publish the inputs (declared resources + price book), not just the number. "Show your work" is what makes a figure trustworthy for grants. Our estimated:bool is a crude v0 of this.

what a real lexicon would add (vs io.zzstoatzz.cost.snapshot)

• Neutral NSID (not io.zzstoatzz.*) — a shared lexicon any actor publishes under (e.g. community.lexicon.*-style).
• Subject as AT-URI — a cost references a repo/service/DID, so costs compose: service → project → org → ecosystem. An appview could sum across many DIDs.
• basis as a first-class enum (billed | derived | estimated) + optional evidence links + recompute inputs.
• Visibility / private vs public — records are public by default. Public path = the grant-transparency story. Private tracking → private PDS / encryption / a visibility field.

Rough shape:

• cost.declaration — one line item: {subject (AT-URI), period, amount{minor,currency}, basis, source{provider, how}, evidence?, inputs? (for recompute)}.
• cost.report — the aggregate/snapshot over a set of declarations.

migration path (when we build it)

Define the generalized lexicon, refactor the connector hub (my-prefect-server/packages/mps/src/mps/costs/) to emit it, and keep io.zzstoatzz.cost.snapshot as a thin alias so the dashboards don't break. Turns "my dashboard" into "a thing anyone can adopt."