On 24 October 2026 we shipped a small change to the NeuraGrowth site: 9 files, 41 lines added, 41 removed. It was called a truth audit. Its job was to make what the site says match what we actually sell.
The checking was done against the files and the catalogue, not against memory. The most visible correction was a number. The site said about 75 guides. The real figure is about 85, made up of 83 active guides on Gumroad plus bundles.
This post walks through that audit, what else shipped in the same ten days, and two worker alerts from the Orqestra pipeline that the record does not explain.
What the truth audit corrected
The audit was one commit in the NeuraGrowthHTML repo, shipped on 2026-10-24. Every line it changed was a claim the site made about the catalogue. Each one was checked against the files and the catalogue itself.
The guide count was the headline. The site said 75 and the right answer was about 85. That is 83 active guides listed on Gumroad, plus bundles. Our log has a separate short entry on the Gumroad count, titled "Gumroad lists 83 active guides, so our site said 85": https://neuragrowth.co/log/gumroad-lists-83-active-guides-so-our-site-said-85/
The other corrections were about scope. The products page holds guides and flashcards. It does not hold every product, because the books and games live elsewhere. Before the audit, the page implied it was the whole shop.
Smaller claims that were wrong or loose
The FAQ got the most care. The books are paperbacks printed by Amazon, so not every product is a download. The FAQ now says that.
Card stock is stated as 160-200 gsm, which is what the free flashcard sample's own instructions say. The PDFs are A4. The templates render A4 only, so there is no US Letter version, and the FAQ no longer leaves room to assume one.
The free flashcard sample got the same treatment. It is 20 cards on 7 A4 pages, cut along dashed lines, not folded. Each card shows the name in English and Polish, with the fact in English only. There is no Polish fact. The commit message is cut off after that, so we will not guess at the rest of the list.
Why a guide count belongs in an AI workflow post
The record does not say when it went stale, so we will not guess.
What the facts do show is the method. The corrections were checked against Gumroad's active listing and the files, not against what anyone remembered.
If you run a store and let an automated pipeline or an assistant write your site copy, the numbers in that copy are claims about a live system. The Gumroad listing is the source here, and the copy is a view of it.
The same ten days, in numbers
Across 2026-10-23 to 2026-11-01 we shipped 238 changes on 10 days with commits. That touched 3408 files, with 36997 lines added and 4852 removed, across four repos.
By repo: skat-game had 100 commits, new-Orqestra 57, MythMarket-Quantum 48 and NeuraGrowthHTML 33. The truth audit above was one of the 33 in NeuraGrowthHTML.
A second change in that window came from skat-game on 2026-10-31, an audit of server-side purchase checks: 13 files, 420 lines added, 73 removed. Its commit message is in Polish. In short, the purchase webhook and the daily check now judge the specific purchase rather than any premium the customer has, through a shared function called assessPurchase. We wrote about that change separately: https://neuragrowth.co/log/webhook-checks-now-judge-the-one-purchase-not-any-premium/
Two stale worker alerts and one slow run
Orqestra raised a warn-level alert titled Worker heartbeat stale on 2026-11-01 at 19:20. The body said the Celery worker was alive but the system was not draining work as expected, and that the last AgentRun was 11861 seconds ago, against a warning threshold of 300 seconds.
A second alert with the same title fired on 2026-11-02 at 06:20, with the last AgentRun 1179 seconds ago. The alert text points to checking the orqestra-worker container logs on the VPS, and to restarting the worker if it is looping. The facts we have do not say whether either alert led to a restart, or what caused the gap. We will not guess at a cause.
Separately, one kids_episode run that started on 2026-10-26 at 01:01:49 took 106.4 minutes and finished with status success. The median across 21 runs of that operation is 21.0 minutes, so this run was 5.1 times slower. The record does not say why it was slow. An earlier alert of the same type is written up here: https://neuragrowth.co/log/celery-worker-heartbeat-stale-last-agentrun-1175s-ago-warn-at-300s/
If an automated system writes or maintains your store copy, keep a periodic audit that checks every count and scope claim against the live listing, as we did against Gumroad on 24 October 2026, and do not trust a number just because it was once true. Treat a success status as separate from a normal duration: our 106.4 minute kids_episode run succeeded and was still 5.1 times the 21.0 minute median, so track duration next to status.
Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.
Top comments (0)