What we changed, and what it cost to find out.

2026-09-02

We shipped a page that told readers to run a check that fails

The Genesis retirement page published a way for anyone to verify the outcome themselves: compare the token contract’s own balance against total supply. If they match, every token that exists is at an address it cannot leave.

They do not match, and never will. UniswapV2 permanently locks a minimum amount of liquidity when a pair is created, so 0.0000001 tokens are stranded in the pair forever. The contract holds 999,999,999.9999999 against a supply of 1,000,000,000.

A reader running our published check would have got a mismatch — which the same page told them to read as something is wrong. We would have handed people an instrument that reports a failure that did not happen.

The corrected check adds the stranded dust and holds exactly. It is a better check than the original, because it accounts for every token in two named places rather than asserting that one number equals another.

The uncomfortable part: we found that dust ourselves, hours earlier, and wrote the phrase “drained, not deleted” specifically because of it. Knowing a fact and carrying it into every place it bears on are different acts, and the second one is where this failed.

2026-09-02

Four numbers, all wrong, all wrong in the same direction

While working out how fast supply could become movable, four separate figures were produced. Every one of them was wrong, and every one understated the risk. One aggregate omitted the largest bucket, so it bounded 65% of supply while reading as a bound on all of it. One divided total supply by only the locked ceilings. One assumed a 360-day year. One treated five schedules as a shared pool when they drain independently — understating the true duration by eight months.

Nobody was trying to flatter anything. All four came from collapsing a five-part structure into a single number, and collapsing drops whichever constraint made the real picture more restrictive.

One of the four was published. 21.6% per year went live at 22:14:19 and was pulled at 22:15:01 — 42 seconds. It is in this repository’s history in both directions: 59a713a put it up, e26759d took it down. The other three were caught before publication.

We are stating that rather than the version we first drafted for this entry, which said none of them reached the site. That draft was false, and it was false in the direction that flattered us — on a page whose entire subject is claims that flatter the person making them. It was caught by checking the sentence against the commit log instead of against memory.

No rate figure is published today. When a summary cannot be derived honestly from its parts, the absence and the reason carry more information than a smaller number, because the reason names a failure a reader can go looking for elsewhere.

2026-09-02

A correct sentence can be defeated by the thing it sits in

Three times in one night, prose that was accurate was contradicted by its own container.

A page path named after a placeholder asserted that the placeholder was the name. A row in a table of locks asserted “this is a lock” while the paragraph beside it said the opposite. A heading reading Allocation asserted recipients that do not exist on chain — all five schedules release to one address.

In every case the words were right and the frame said otherwise, and the frame wins, because a reader does not experience a layout as a claim to be checked.

Which is also why it survives review: prose gets proofread, frames do not. Nobody reads a table asking what its existence asserts.

2026-09-02

Our own site had a one-click path into $25 of liquidity

A page on this site carried a pre-filled swap link and an add-to-wallet button for the Genesis token, with no risk language of any kind. At the time, a buyer spending $100 could recover $0.95 if the majority holder sold one tenth of one percent of their position.

None of it existed in our repository. The published file was 1,695 bytes larger than the source, meaning it had been deployed from an uncommitted working directory. The live site was serving something nobody could review.

It was removed by reconciling the published file rather than overwriting it with our older source — which would also have removed the buy link, but silently, as an invisible side effect of an unrelated change. Removing a buy funnel should be a decision with a commit message, not an accident. It has one: 8ac74e7.