Refine: ONE's own run

The decision

ONE can count how many strangers arrive, and how many agents it has made, but it cannot yet say what share of either becomes a paying Sold buyer. There is no Sold buyer to count: 0 Sold purchases, 0 live payments. So the numbers below are counts, never rates. The weakest seam we can see is the home page to a made key: 202 outside visitors who reached the key band, 1 key press. The first test changes one line above that press and is read against a count set before it runs. A rate appears only after a seam has enough people to carry one.

What we built

The funnel, as ONE defines it

Visit, then key (the press posts key:minted), then link, then first sale (a live payment; test money never counts), then Sold ($100), then retained (the community month two charge succeeds; that price does not exist yet).
Activated, for ONE: a wallet that makes a first transaction (text/strategy.md, "the number that counts"). Never sign-ups.

Two funnels, never joined

ONE's own code keeps what a visitor saw apart from what an agent did, because the two share no key (text/home.md section 10). Brian is equally a person and an agent, so both are read.

Saw: the pixel on one.ie, 2026-10-28 to 2026-11-12, outside visitors. Rule: not a bot, signed out, not from Thailand, Singapore, Hong Kong or Mexico, no localhost referrer.

589 opened a page.
250 of them landed on the home page. The sets below are not nested: some did something without a home view.
253 did something beyond a view: a click, a scroll, a film watch or a chat. That is about 17 a day over 15 days, and 4 on 2026-11-11.
202 reached the key band on the home page (the key stop, read before the layout changed).
7 opened a key page.
1 pressed Make my key. That visitor's referrer was an astro.build page.
0 opened /sold. The page is tracked: 66 views from 8 visitors in all, and none of them passed the outside filter.

Did: server rows, all time since 2026-09-20. Agents ONE provisioned, each step counted on its own against the 79, so the steps are not nested. They include team and test agents.

79 got a key.
29 sealed a wallet.
37 opened a shop.
37 made a product.
8 had a test sale.
14 were handed over to a person.

Money, and what each count counts.

39 purchase events in the storefront log, 36 paid orders, 13 payment events in the pixel log. Three counts, one fact: every sale so far is test money (36 of 36 orders are Stripe test sessions, 0 live).
19 of the 36 paid orders carry a visitor hash. 0 carry a campaign. A sale cannot yet be traced back to an ad.

Sold. The store check printed one line and exited 1: a leftover test mug is still for sale. The Sold course no longer fails the check. 0 Sold orders. 2 enrolments, with no matching order.

Baseline, side by side

The 2026-11-07 baseline: about 43 a day outside Thailand who did more than view a page, and about 3,900 addresses a day, mostly bots. It counted every site carrying the pixel.
Today's figure: about 17 a day, one.ie only, with the localhost and region filters above. Different counts: never subtract one from the other. Neither proves traction.

The weakest lever, and why

The lesson says to move each seam by the same percent and compare the bottom. We cannot: we have no rates, and the bottom is zero. What the counts do show is where people stop. 202 outside visitors reached the key band. 1 pressed the key. Past that point we see only test traffic. The second break is upstream: 0 outside visitors reached /sold, because no traffic is routed there and no ad has run.

So the leaky seam is the key band to the key press. The small stage is /sold, and a starved stage cannot be fixed from inside. We test the leak first: it needs no spend and the 202 are enough to read a count. The ad comes second, behind its four gates (text/sold-marketing-plan.md).

The one test to run first

Change one thing: the first line above the key press. Control keeps the live words (checked on one.ie on 2026-11-12): "Your key is one press away." Variant says the problem first: "To be paid, you need a key."
One metric: key:minted with pressedBy: person, among outside visitors who did something.
Window: the next 250 such visitors, set by volume, not by a date. Split them between the two lines. If the page experiment cannot split, run the variant alone for 250 and compare it to the 1 above, and say it was sequential.
Pass bar, set before the run and a decision, not a measurement: the variant gets at least 4 presses in its share, and the control gets 1 or fewer. Anything between is "no verdict yet".
Do not change the line mid-test.

The retention math, ready for the first buyer

Cohort: the buyers of one ad build, read at the same course step. Not a month average.
Gross retention: buyers whose month two charge succeeds, over Sold buyers in that cohort. A count first, then a ratio.
Net retention adds credits or a Vespio call bought by the same buyers.
Money rule (text/sold-marketing-plan.md, a decision): scale while cost per Sold sale stays at or under $100. Cost per Sold sale has no number: no ad has run.

Receipts

Pixel counts and the agent funnel: production D1, read-only, 2026-11-12. The agent query is the funnel route's own (one.ie/web/src/pages/api/analytics/journey.ts:97). The pixel keeps about 14 days, which caps any cohort reading.
Store check: ONE_BASE=https://one.ie bash .claude/scripts/sold-store.sh, exit 1, one stray test product.
The command-line key is invisible by design: npx oneie key new makes no network call (packages/cli/src/key.ts header), so it is never counted. Sovereignty caps what we can measure.
https://one.ie/sold and https://one.ie/learn/saas returned 200 on 2026-11-12.

Still open

No rate is measured at any seam. A ratio from 1 press or 8 sales is noise.
No Sold buyer, no retained buyer, no month two. Retention is unmeasured, not zero.
The link stage has no clean outside-the-team count. 37 agents have a product, with walks and team tests mixed in. The five-path chain to a first payment showed no run rows; VERIFY that query before reading it as zero.
Ad to sale is untraceable today: 0 of 36 orders carry a campaign. A note from 2026-11-07 says UTM was never captured at ingest; the code now reads the five UTM keys (pages/api/events.ts:82). VERIFY on production rows.
The home page changed shape recently, so older stop names in the pixel do not match the current five beats. The stop funnel is not one series yet.
Home copy still reads "0% on USDC" and "100% free", which the words-we-keep-true list says not to print as a bare claim. VERIFY with Tony.
The 17 a day may still include the team on a VPN. The pixel cannot prove otherwise.
The home page is being edited by others right now (four files uncommitted on dev). The control line must be frozen for the whole test, or the test is void.
https://one.ie/learn/saas/one/refine is not published yet.

Do it yourself

Write your funnel as five words, one per seam, and define "activated" in a sentence. Never change it mid-quarter.
For each seam, find the row your system writes when it happens. A click is a pixel, not a fact. Keep the two apart.
Write the count at each seam for a fixed window, with who is excluded. Mark any seam you cannot see as unmeasured.
Pick the seam where people stop and enough of them stand to read. Change one line. Set the pass bar as a count before you start, and the window by volume.
Run it. Then go back to Stage 0 with the new numbers.

How it was measured

Outside visitor rule: slug='one' AND user_agent_class NOT IN ('bot','unknown') AND (actor_id IS NULL OR '' OR LIKE 'anon:%') AND region NOT IN ('','TH','SG','HK','MX') AND page_referrer NOT LIKE '%localhost%', from the same rule as lib/analytics/human-view.ts. Interaction events: ui:click, click, scroll, film:progress, chat-reply, chat-message, chip-click.
253 distinct visitor_hash with an interaction, 589 with a pageview, 250 with a home pageview, 202 with a ui:journey:reach on stop key, 1 with key:minted, 0 with a /sold path. Unfiltered, /sold has 66 pageviews from 8 visitors. The own_browser:1 mark means a real browser, not the team, so it is not used as a filter.
world_keys with label provision-reg joined to wallets, owners, products, storefront_events and invites: 79, 29, 37, 37, 8, 14.
orders WHERE status='paid': 36 paid, 0 cs_live, 36 cs_test, 19 with a visitor hash, 0 with a campaign. payment events: 13, 0 live.
Not run: any ad read or real checkout.