Seven stories, a few small finds at the end. All of it happened out there. If you lead a small team or a small newsroom, you will see where it touches you. I mostly leave that to you.

The dams linked to higher salmon survival were built by a council whose staff member co-wrote the study. A Lean proof of Fermat's Last Theorem came from Anthropic and was checked by the leader of a parallel formalization project. None of this makes a result false. It tells you where to look for the second opinion.

In Texas, lawmakers added $1 to a car-insurance fee in 2023 to fight catalytic converter theft. The Tribune says the money also paid for licence-plate cameras. At the Times, a union drew a line between an ad and an integration. The purpose gets written once. It gets read many times.

60 percent survival is one creek, with two years of "before". 34.7 percent is a count of citations a reader cannot verify, not of wrong answers. 2.4 percent is how much of one report's sample three sites made up. Before you repeat a figure, learn its denominator.

A note on how I write this. I mark what is confirmed, what is reported, and what is my own reading. Where I only saw a headline, I say so.

The study is not from this week. Michael Pollock, Stuart Munsch and Erich Yokel published it in Frontiers in Ecology and Evolution on 8 July. What spread in the past days was coverage of it, for example Discover Wildlife on 2 September.

The dams are low-tech. The Scott River Watershed Council, in northern California, drives small wooden posts across a stream. It weaves willow or conifer branches between them. Gravel, straw and mud seal the upstream side. The water slows and pools. Young salmon get still water to grow in.

Survival comes from PIT-tagged juveniles, almost all 65 to 80 millimetres long, picked up on antennas and corrected for detection. In French Creek the mean was 8 percent for brood years 2015 and 2016, and 60 percent for 2017 to 2020. Sugar Creek averaged 55 percent. Across all years the range was 6 to 90 percent.

Now the limits. The 8 to 60 is one creek. The "before" is two brood years. The "after" is four.

One more fact belongs here. Pollock and Munsch work for NOAA; the third author, Erich Yokel, works for the Scott River Watershed Council, which built the dams. The people reporting the result are close to the people who did the work. That is not a flaw on its own. It is a reason to want a second creek and a second team.

Of each $5, $1 goes to general revenue for catalytic converter theft prevention. The other $4 is split: 20 percent to the Motor Vehicle Crime Prevention Authority, 20 percent to general revenue for criminal justice, 60 percent to trauma and emergency care.

Ayden Runnels and Alex Ford reported on 28 August that at least $30 million of fee money went to Flock. About 2,000 cameras came through at least 95 local grants. About 1,200 came through the Department of Public Safety.

My reading: the law was voted on once. What the money bought was decided later, by a board.

Haus Research counted how often that happens. It asked 310 factual questions about 210 tech companies: when founded, how much funding, how many employees.

Trellner Research asked a different question on the same day. Where does Perplexity look when you ask for the top five software products?

Both reports state limits. Trellner: "We have not shown that any of this changes the answers."

Haus has published its dataset under CC BY 4.0.

Unanimous. That is how the Times Guild Unit Council and The Athletic's contract action team approved the statement the NewsGuild of New York published on 2 September.

Its letter went to A.G. Sulzberger and to The Athletic's publisher, David Perpich. It speaks of the company's "potential deal" for The Athletic to partner with Kalshi, a prediction-market company. It asks the company to "abandon this deal."

So two versions are in play. The anonymous source says the decision came first, and the reporting describes no signed deal. The union asked for an abandonment anyway. The two accounts give the letter different roles. I cannot say which is right.

Kalshi does not comment on rumoured deals. A Kalshi spokesperson told Front Office Sports that prediction markets are "a great complement to news sites like The Athletic and The New York Times."

Tarn Adams, creator of Dwarf Fortress, said this to PC Gamer at Gamescom. Joshua Wolens did the interview, Harvey Randall wrote it up, and it ran on 1 September.

This is an anecdote. "Everyone I know" is a circle of friends. Nobody measured how often bosses ask about commits.

The better evidence is his own situation. In a 2021 Stack Overflow Blog profile, he wrote all of the code himself. By a rough count (he searched for semicolons) he put it at 711,000 lines then, a figure from 2021, not today.

On Hacker News, a commenter, tptacek, said there may be no worse messenger for this problem than the author of Dwarf Fortress. The game, he wrote, "succeeds in spite of and not because of its code." Others in the thread disagreed. One said the design and intent make a game, not the code.

Adams himself described his mix of C and C++ as "a mess that's accreted over time."

A disclosure first: I draft these issues with help from Claude, and this story is about Anthropic.

Anthropic published "Formalizing Fermat's Last Theorem" on 4 September. It says the model produced about 13 million lines of Lean and proved 30,300 intermediate theorems, 29,500 of which are used in the final proof. It used about 6 billion output tokens.

What is formalization? You take a proof experts have read and accepted, and you translate it into a language where a small, widely trusted checker approves every step down to the foundations of logic. The proof itself is old. Wiles and Taylor published it in the 1990s.

Kevin Buzzard of Imperial College London leads the human project to formalize the same theorem. He wrote a post on 4 September called "FLT: Anthropic has beaten me to it." He says he compiled the code base and ran comparator on it, and "it checks out."

A commenter on his blog, David Jao, did the arithmetic: 6 billion output tokens at the API price would be $300,000. That is Jao's calculation. The model is internal, and the internal price is unknown.

The design: 60 proofreaders, 20 per condition. Three documents, each with 100 planted errors. Deadlines on days 7, 14 and 21, or self-chosen, or all on day 21.

The published effect between evenly spaced deadlines and day 21 was d = 2.5, with means of 136.1 and 71.1 corrections. Data Colada compares that with the height difference between men and women, about 1.8.

Then the twins. Participants S1 and S11 match exactly across all three tasks. So do S7 and S17. Their ID numbers are exactly ten apart. No such pairs appear in the other two conditions.

Provenance matters. Kyle Hyndman received three Excel files on 20 April 2006 in an email from Ariely's MIT account. The file properties read "Last saved by Dan Ariely."

That was the week.