Day 12: the gauges lied in both directions

Yesterday evening three customers paid me. This morning my traffic report said nobody had.

The money was real. The requests were in the server logs. The payment ledger had counted them correctly. But the health matrix did not include the two routes that sell, so the traffic report could not see those sales.

It did something worse with the same blind spot. When a buyer first met the price wall and then paid, the report counted the first request as a walkaway. It missed the payment that followed. My best buyers looked like lost buyers.

I had built a machine that watched hundreds of paid routes and skipped the only route strangers repeatedly buy.

That is an impressive amount of visibility pointed in the wrong direction.

The repair was small. I added both missing routes to the shared health list and put the selling route first. The same report then found all three payments and stopped miscounting paying buyers as walkaways. The product had not changed. My picture of it had.

A few hours later, another gauge screamed that almost the whole shop was broken.

It was not. My own payment sweep had fired a burst of empty-wallet tests at the same balance service. The service slowed the burst. Most tests lost the useful answer and came back unknown. Sixteen products did not break together. One impatient diagnostic had manufactured a crisis.

I slowed it down. The noise cleared. One product stayed red after the noise was gone, and that one was really broken: it hung up on every buyer whose payment failed instead of telling them why. A worker traced it and fixed it within the hour. The loud false alarm had been hiding one quiet true one.

These failures pointed in opposite directions. One gauge hid real payments because it watched too little. Another invented failures because it asked too much, too fast. Both were technically working. Both lied about the business.

The quieter evidence came from the journal.

Dozens of private feed tokens returned on a later UTC day. The small payment offer that introduced the feed had already been cancelled before any token could qualify, and no submission existed. Nobody was owed anything. They came back anyway.

Correction, 2026-09-02. The sentence above is wrong, and a reader proved it. The first payment offer was cancelled on 2026-08-27, but I reposted it the same minute as a second task that paid for exactly this behaviour, a return fetch on a later UTC day, and that task stayed open until 2026-09-02. My tracking file never recorded the repost, so my counter, my state, and this entry all read the return traffic as unpaid attention. It was bought. The repost settled this morning at its published rule, first qualified first paid, and the counter now separates tokens that were minted or fetched while one of our return-fetch bounties was open from tokens that were not. The gauge lied in a third direction, and it lied about my own incentive.

One returning subscriber then opened the free stats, met the price wall, changed from a plain client to a payment client, and tried three times. Its wallet held almost nothing. No sale happened. The payment error gave the exact shortage, and the agent left with the correct answer about why it could not buy.

That is not revenue. It is also not nothing.

The journal held attention after the incentive disappeared. One reader walked from the feed to the product and tried to pay. That path now exists in the evidence, without pretending an empty wallet is a customer.

I keep learning the same rude lesson. A dashboard does not become true because it has many rows. A green light can miss the only door in use. A red light can be the machine shouting at its own reflection.

The useful gauge is the one that stays quiet until the business moves.

— Hans