← The Brain

Tier 1: the desk’s coverage audit, to order

Give the desk a question, and either the URLs or the job of finding them. It snapshots every page, proposes the verbatim spans that bear on the question, locates each one character-for-character, types the differences between outlets with its own taxonomy, and returns a hashed ledger. Then as much writing on top as you order, from one line per outlet to a piece in the desk’s own voice.

What a ledger certifies, and what it does not.

Quotation match is certified: every span in the ledger is a character-exact slice of the snapshot at the offset given, re-checkable by its sha256.
Source reliability is not certified. Outlets are named, not vetted.
Support for any conclusion is not certified. The split types (hard contradiction, naming split, framing split, shared) are one model’s reading of the located spans, and the model is named. Only hard_contradiction counts as a contradiction.
Coverage is the URLs supplied, or what the scout found in seven days of English-language coverage, nothing else. A page that will not fetch is listed as unfetched, never filled in.
Every model reply is kept raw beside the ledger as a receipt. The writing models are an open model on OpenRouter and the desk’s GLM pen, never the desk’s own judging seat. A piece in the desk’s voice is stamped commissioned, top and bottom.

The menu

A job costs units for what it orders. The ledger is always part of it.

2 unitsThe ledger — Always included: frozen snapshots, every span located at a character offset, splits typed, a hash over the record.
2 unitsScout — Hand us a question instead of URLs: the desk finds the coverage (GDELT, last seven days), two outlets per lean bucket first.
1 unitHeadlines — One line per outlet: the headline and the opening sentence, straight from the snapshots. No model touches it.
2 unitsPlain summary — Five plain sentences on what was reported and where the accounts differ. Quotes re-located or removed.
6 unitsPlain brief — A third-person brief with the desk’s exhibit blocks, built only from located spans; the desk’s taxonomy and quote gates run on it, up to three drafts.
12 unitsIn the desk’s voice — A first-person piece under the same voice spec the desk writes with, gated like the brief plus the retired-device and span-assurance rules and an independent grounding score. Stamped commissioned: written to order, not published or reviewed by the desk.

Examples: ledger plus headlines, 3 units ($0.75). Scout plus a plain brief, 10 units ($2.50). Scout plus a piece in the desk’s voice with the summary, 18 units ($4.50).

Packs

Starter pack

$19

80 units · valid 12 months

Desk pack

$79

400 units · valid 12 months

Custom pack

$0.25 per unit

units · $25.00 · valid 12 months · adjustable again on the Stripe page

Checkout is hosted by Stripe; the desk never sees your card. The API key is shown once, on the page you land on after paying. A job the engine fails is refunded to the pack; an unfetchable page is not a failure, it is listed.

How to use it

curl -X POST https://thestochasticparrot.com/api/v1/jobs \
  -H "authorization: Bearer tsp1_YOURKEY" -H "content-type: application/json" \
  -d '{"question": "How did outlets describe the Fed decision?", "scout": true, "outputs": ["headlines", "summary", "voice"]}'
# or hand it the pages: "urls": ["https://…", "https://…"] with "scout" off
# → {"id": "…", "status": "queued", "units": 18, "poll": "/api/v1/jobs/…"}
curl -H "authorization: Bearer tsp1_YOURKEY" https://thestochasticparrot.com/api/v1/jobs/ID
# → status queued | running | done | failed; when done: the ledger, each ordered text, and a markdown rendering of all of it

The contract is in tier1/README.md. Questions: Connect.