What is true about this product, right now.

I build The Collective in public. Below: the claims I stand behind and their proof, the real AI bill, the build log, what broke and what it taught me, and the standing list of what I won’t build. When a number here is bad, it stays up.

Who wrote this

A model drafts and I decide. It proposes the week’s readings, writes three candidate one-lines and the sixty-second script, and corrects a word or a sentence when I click that word or that sentence — never without the click, never into the field on its own. Nothing here publishes itself. So the honest way to put it is: the words you read are mine, and some of them started as its. This is how many.

1 OF 1Drafts I published without changing a word

That is the number to watch, and not in the direction you would expect. A model that learned to write what I never edit is indistinguishable, from the inside, from a model that got good. If it climbs toward 8 of 8 the loop has stopped helping me and started flattering me — and you will read it here before I have found a kind way to say it.

One is a count, not a rate, and I will not turn it into a percentage until it can survive being one. There is no median change here for the same reason: at three pairs it swings twenty points on a single Monday. It appears at four — which is also, roughly, where the drafts stop reading like a house style and start reading like me.

Draft→final pairs1median change0%

Teaching windowlast 8 lineslast 8 pairs

What it is allowed to do — five operations, all logged

propose_themes: Two or three readings of the week, from the ranked pool.
draft_one_line: Three candidate sentences from the reading I chose and the five stories I kept.
draft_script: One sixty-second script, for me to read to camera or not.
improve_one_line: Corrects one word — the word I clicked, and only when I click it.
improve_script: Corrects one sentence, the same way. Neither one rewrites the whole.

Drafting$ 0.35across 95 calls — failures included71 failed

AI spend · OCT$ 0.02Cap$ 50.00Revenue$ 0

As of [OCT 05 · 04:34 UTC] · From the product’s own ledger

Claims register

Nothing is for sale.: Fact · Price list: none · Revenue: $ 0
Every AI call is logged, and the bill is public.: Mechanic · Cap $ 50.00/mo, hard-coded
A model drafts; nothing publishes without a person.: Mechanic · 5 operations, all logged · no write without a click
When its words go out untouched, the count says so.: Kept verbatim 1 of 1 · median change 0%
Members read it, and keep what’s useful.: № 2 · Read 25 · Kept 5 · Weekly since № 1

The AI bill

What the AI in this product costs — monthly, by feature, with honest denominators. If it spikes, you’ll see it spike.

$0.02Spent this month

$50The cap — hard-coded

Feature	USD
EVERYTHING ELSE103 calls · 2 failed	$ 0.02

Every call logged — failed calls billed and included

Build log

No entries yet — the first posts after the first Friday ritual, [FRI OCT 09]. Each one will say what shipped, what it cost, what broke, and what the eval said. The format is fixed before the first entry exists, so there’s nothing to soften later.

Lessons ledger

12 of 96Still at "documented" — this number should fall

Every mistake either becomes a rule the tooling enforces, or it happens again. The ladder: documented → rule → impossible. Most recent:

A `## Files` glob over a Next.js route folder matched nothingL096 · CI-GATE
Workers wrote result files outside their worktree, twice in one dayL095 · RULE

Refusals — standing

Deliberately small, and not up for A/B testing:

R1No vendor money — no sponsorships, no affiliates, no placements.
R2No engagement optimization — no streaks, no feeds, no time-on-site metric anywhere in the code.
R3No voice cloning — nothing here speaks as a person.
R4No autonomous outreach — nothing sends on my behalf while I sleep.
R5No dark-pool data — nothing bought, nothing scraped from behind logins.