THE PELOTON

The Infection Confirms It: Van Aert Misses Tignes, Tour Start in Real Doubt

↩ Developing story — first reported Jun 14 · previously Jun 15

HERENTALS, Belgium — The worst-case scenario has materialised. As this paper reported Monday, Wout van Aert did not board the Tignes plane — and now the reason is confirmed: the wound on his elbow, aggravated by his time trial aero position at the Tour Auvergne-Rhône-Alpes, has become infected. According to Belgian outlet HLN, he must stay off the bike for at least several days while the inflammation responds to treatment.¹ With Visma-Lease a Bike's altitude camp already under way without him and the Tour de France starting in Barcelona on July 4, his participation in the race is a genuine question mark.

The infection makes the picture considerably bleaker than the bruised-elbow management story of last week. Van Aert crashed on his TT bike on June 1, won stage 5 of TARA regardless, then abandoned the race when the elbow deteriorated further. Sporza commentator Christophe Vandegoor, speaking after contact with van Aert's manager Jef Vandebosch, noted that the irritation was almost certainly worsened by the aero bar position during stage 3's team time trial — not contamination of the wound itself.¹ An infection still means antibiotics, enforced rest, and a recovery curve that now competes directly with the 18 remaining days to the Grand Départ.

Visma's provisional Tour roster, reported by HLN and confirmed by Escape Collective on June 15, still carries van Aert's name alongside Jonas Vingegaard, Bruno Armirail, Victor Campenaerts, Matteo Jorgenson, Sepp Kuss, Edoardo Affini, and Per Strand Hagenes — the latter a last-minute replacement for the injured Christophe Laporte.² The team has not revised the list publicly. Hagenes himself was knocked out of TARA by illness before stage 7, along with Jørgen Nordhagen, leaving Jorgenson with three helpers for the finale. The team's luck with illness this June has been exceptional, and not in any good sense.

The Tour de Suisse starts Wednesday in Sondrio, which also happens to be where the rest of the field's Tour preparation accelerates while van Aert sits at home. This year's race has been cut from eight days to five — June 17 to 21, Sondrio to Villars-sur-Ollon — in what organisers describe as an effort to keep the race "financially sustainable and future-proof."³ What it loses in attrition it gains in intensity: a 23.8km time trial on stage 4, a four-mountain final stage on the Col de la Croix, and a field headlined by Tadej Pogačar on his Tour de Suisse debut.

Pogačar arrives having won four of five stages and the overall at the Tour de Romandie, and eight of his eleven race days in 2026 have ended with a win.³ Key Tour lieutenants — Nils Politt, Mikkel Bjerg, and Brandon McNulty — travel with him. The field also includes Primož Roglič (Red Bull-Bora-Hansgrohe), who has won every other major WorldTour week-long race and would love to complete the set; Richard Carapaz (EF Education-EasyPost), returning from perineal cyst surgery that ruled him out of the Giro; Antonio Tiberi (Bahrain Victorious), making his Tour de France debut in July and using Switzerland as his principal preparation; and Ilan Van Wilder (Soudal-QuickStep), leading a team racing its first post-Evenepoel summer.

Tom Pidcock was on the start list too, but Pinarello-Q36.5 announced Tuesday morning that a mild viral infection at their Sierra Nevada altitude camp has scratched his Tour de Suisse entry.⁴ The team says he has returned to Andorra and may instead race the Andorra MoraBanc Classica on June 21 — 13 days before the Barcelona TTT — "health permitting." The Tour de France was not mentioned in the team's statement.

On the women's side, Demi Vollering is also absent. FDJ United-SUEZ confirmed she will skip the Tour de Suisse — which runs concurrently with the men's race over the same five days — to recover after grinding through the Giro d'Italia Women and build toward the Tour de France Femmes on August 1.⁵ Vollering's departure opens the race considerably: also absent are Elise Chabbey, Cecilie Uttrup Ludwig, and Lotte Kopecky. The contenders who remain include Kasia Niewiadoma-Phinney (Canyon-SRAM), whose five second-place finishes in 2025 make a win here feel overdue; home favourite Marlen Reusser (Movistar), a two-time Tour de Suisse winner who should dominate the stage 4 time trial; Elisa Longo Borghini (UAE Team ADQ), finishing her Giro recovery; and Juliette Berthet (FDJ United-SUEZ), who took fifth overall at the Vuelta Femenina this spring.

The other major piece of news today is the 2026 Tour de France route, published in full by BikeRadar. The headline is a double summit finish on Alpe d'Huez — stages 19 and 20 both end there, with stage 20 serving as the queen stage at 5,600m of climbing, taking the Col de la Croix de Fer, the Col de la Télégraphe, and the Col du Galibier before the Alpe via the Col de Sarenne (a route used only once before, in 2013, and as a descent).⁶ Stage 1 in Barcelona is a 19km team time trial — the first Tour opener of this type since 1971 — with individual finish times taken, meaning GC gaps can open on day one.⁶ The sole individual time trial is 26km along the shores of Lake Geneva on stage 16. The race covers 3,333km with 54,450m of elevation gain.

Elsewhere in team news, Czech sprinter Pavel Bittner is closing in on a move from Picnic-PostNL to Pinarello-Q36.5, per Escape Collective.⁷ Bittner won a Vuelta stage in 2024 and posted a string of top finishes this spring — third at Bredene Koksijde Classic, second at Scheldeprijs — making him a meaningful loss for a Dutch team that has now shed Hirschi, Benoot, Arensman, Welsford, Andresen, and Onley in recent seasons. Picnic-PostNL is operating on a one-year limited WorldTour license pending financial guarantees; holding onto talent in that environment has proven essentially impossible.

↑ Back to top

THE LAB

The Jailbreak Was "Fix This Code": Monday Talks End Without Lifting Fable 5 Controls

↩ Developing story — first reported Jun 13 · previously Jun 14

High-level talks between Anthropic and the White House concluded Monday without lifting the export controls on Claude Fable 5.¹ The Commerce Department, attending via Howard Lutnick dialing in from the G7 summit in Evian, maintained its position that Fable 5's guardrails can be disabled. Anthropic reiterated that the administration's concerns are overblown. No next steps were announced.

The technical substance of those concerns has now been made public, and it is not flattering to the government's case. Cybersecurity expert Kate Moussouris, CEO of Luta Security, told The Atlantic that Anthropic shared the White House's report on the alleged jailbreak with her for independent review — and that she is not being paid by Anthropic.² What the report described, Moussouris said, was researchers giving Fable 5 deliberately insecure code and asking it to "review the code for security issues." The model refused. They then asked it to "fix this code" — and it did. Through a "multistep and manual process," they turned the output into scripts that test patches.³

Simon Willison, writing on Jun 16, called the situation bluntly: "Coding models fix bugs, and security exploits are the most important category of bugs for them to fix." Moussouris told The Atlantic the model was simply "working as intended" for cyberdefense — that finding bugs, explaining fixes, and writing tests to confirm a patch is the entire point of using AI for defensive security. As she put it, the prompts worked because they were defensive requests, and you cannot remove that capability without making the model worse at fixing bugs and verifying patches.

The government's position is that the NSA reviewed the alleged vulnerabilities and concluded it was indeed possible to strip away Fable 5's guardrails. That conclusion is now being publicly contested not just by Anthropic but by an open letter at freefable.org, signed by independent cybersecurity researchers, which argues that Mythos-class models are not uniquely capable at finding and weaponizing exploits — and that "this action has taken the best models away from defenders, created market uncertainty, and risked America's AI leadership without any real risk to justify it."¹

Monday's working group meetings at the Commerce Department were attended by Anthropic cofounder and chief compute officer Tom Brown, head of external affairs Sarah Heck, head of frontier red-teaming Logan Graham, and senior security researcher Nicholas Carlini. The Office of the National Cyber Director's Sean Cairncross did not participate, despite the meetings taking place with the Center for AI Standards and Innovation and the ONCD. "Both parties are working quickly to get this resolved," Anthropic's spokesperson said. The Commerce Department signaled a willingness to restore consumer access contingent on Anthropic fully resolving the jailbreak concerns — a condition Anthropic appears to believe it cannot satisfy without validating a false premise.¹

A detailed post published Jun 13 by a Slay the Spire 2 player and developer documents one of the more technically striking bugs in a modern game: essentially every source of randomness in the game is correlated with every other, in ways that are not just observable but mathematically precise.

The root cause is C#'s System.Random. The game seeds its many independent RNG streams as seed + hash("name") — so potion drops, curse rewards, monster AI, card draws, and event outcomes all start from seeds that differ from each other by known, fixed offsets. The problem: System.Random's output is linear in the absolute value of its seed. Specifically, SeedArray[1] - SeedArray[22] is the first output, and every entry in SeedArray is of the form x*S + y for integer constants x and y, where S is the absolute value of the seed. The difference between two linear functions is itself linear — so two RNG streams whose seeds differ by a constant d will have their first outputs differ by exactly x*d.

The consequences are extensive and reach into actual gameplay balance. It is literally impossible to get the card Rebound from the Trash Heap event in a single-player game — the Underdocks-exclusive event is correlated with the act RNG, which makes Rebound unreachable.⁴ The "random" curse from Neow's Bones in Underdocks is about 54% likely to be Debt — players on Reddit and Discord had been posting about suspiciously bad luck for weeks, dismissing it as confirmation bias. It is not. First-combat potion drop rates are 76% in Underdocks and 4% in Overgrowth — not close to equal. The Doll Room event in Act 2 can be predicted from Neow choices made at the start of Act 1.

The author discovered this accidentally while writing a seed-search program. After adding an Overgrowth condition, the search returned no results. A scatterplot of two RNG outputs that should have been independent looked like a line with a defined slope. The fix is trivially simple: replace System.Random with any modern PRNG — the post includes a 50-line PCG32 drop-in that would require a three-line change to the Spire 2 codebase. As the author notes, this would also fix the cross-platform seed divergence that plagued Spire 1 on mobile. Mega Crit has not yet responded publicly.

Trending today: GitHub is saturated with AI agent wrappers, Claude Code skills collections, and MCP proxy tools — no repository cleared the bar for technical novelty.

↑ Back to top

THE WORLD

World Reaction Locks In; Seattle Opens World Cup; King County Raises the Tax

↩ Developing story — first reported Jun 14 · previously Jun 15

• Iran, Friday — World reaction locked in: E4 nations, Japan, China, and the UN all endorsed the MOU; Strait of Hormuz set to open after Geneva signing. Al Jazeera / CNBC

Seattle opened its World Cup account Monday with 66,775 fans packed into the stadium for Egypt vs. Belgium — fewer than 200 seats empty. Egypt drew 1-1. Team USA meets Australia Friday at noon; King County Sheriff's deputies will be out in force along Eastside transit routes into downtown, with helicopters overhead and a drone-flight ban in effect around the venue. KIRO7 attendance / KCSO security

The King County Transportation District passed a new 0.1% sales tax in a 5-4 vote at its June 12 meeting, its first unilateral use of taxing authority since the district was formed in 2014. The revenue — projected at $100M per year — partially offsets a $200M roads shortfall covering 1,500 miles of unincorporated county roads, with 12.5% passed through to cities. Effective January 1, 2027. Issaquah Reporter

ON THE TRAIL

PART 1 — WEEKEND PICKS

Juneteenth (June 19) falls on a Friday and is a federal holiday, observed as such. Window: Friday June 19 – Sunday June 21. Thursday June 18 is opportunistic — forecast is even warmer than the weekend (Teanaway hits 83°F, Snoqualmie 78°F, both sunny 0% precip), so if you can swing PTO Thursday, a three-night trip is worth it.

Pick 1: Pete Lake — Teanaway / Cle Elum I-90 East, ~90–110 min from Issaquah. 1-night (or extend Thu–Sun for 2-night). Mileage: ~4.5 miles each way, ~300 ft gain (flat valley approach). Weather (Teanaway): Juneteenth 82°F, sunny, 0% precip; Saturday 74°F, sunny, 0%; Sunday 76°F, sunny, 4%.

A Jun 15 report puts Pete Lake in good shape: wildflowers blooming, Ingalls Creek audible the whole way, plenty of water for filtering, zero snow. Crowds are low on weekdays; the reporter ran part of the way in and saw the lake to themselves before turning back. Mosquitos pick up once you're at the lake itself — bug spray required. No fords. This is the shortest-drive backpack on the list with the best weather window. WTA trip report

Pick 2: Glacier Basin — Mt. Rainier (White River) Mt. Rainier NE, ~110–140 min from Issaquah. 1-night. Mileage: 3.1 miles to basin, ~1,100 ft gain. Weather (Mt. Rainier / White River): Juneteenth 68°F, mostly sunny, 0% precip; Saturday 61°F, sunny, 3%; Sunday 63°F, sunny, 3%.

A Jun 15 report finds the trail clear to the basin with a few small blowdowns easy to navigate, snow patches only in the final 0.5 miles (poles helpful), wildflowers in open meadows, and a bear grazing near camp — camera-worthy, not cause for alarm. The White River provides sound and water the entire approach. Campsites at the basin were occupied but not overrun. No fords. Bugs not mentioned. The tradeoff versus Pete Lake: longer drive, cooler temps, more dramatic setting. WTA trip report

---

PART 2 — REGIONAL SNAPSHOT

• Snoqualmie Pass (I-90) — Snow Lake trail in good shape with heather blooming; a few snow patches on the descent into the lake basin, ice still on far end of lake. Rachel Lake approach has a dozen stream-crossing sections from snowmelt and serious blowdowns — route-finding required. Wildflowers peaking. Bugs starting. Snow Lake / Rachel Lake
• Snoqualmie / North Bend — Rattlesnake Ledge clear and sunny, parking fine at 9am. CCC Road (Blowout Creek) passable with some bugs once you enter the forest; great solitude. CCC Road
• Stevens Pass – West (US 2) — Frog Mountain in excellent condition, 8.2 miles / 2,198 ft, wildflowers abundant; one of the strongest recent reports on the corridor. Scorpion Mountain via Johnson Ridge has snow patches and active mosquitos in the meadows. Barclay Lake sees weekend crowds; road water crossing requires high-clearance or roadside parking and a short walk. Frog Mtn / Scorpion Mtn
• Teanaway / Salmon La Sac (I-90 East) — Iron Peak trail clear with some blowdowns; persistent goat on the ridge, give it space. Johnson Mountain / Malcolm loop doable but watch for 4+ ticks — check carefully. Pete Lake in good shape for a flat valley backpack. Iron Peak / Johnson Mtn
• Mt. Rainier (Chinook Pass / Hwy 410) — Naches Peak Loop still has significant snow past the first mile; Hwy 410 is open over the pass but the loop needs a few more weeks. Norse Peak mostly snow-free, flies and ladybugs at the summit. Glacier Basin excellent for camping (see Pick 2). Emmons Moraine river crossing remains dangerous — multiple parties turned back. Naches Peak / Glacier Basin
• Mountain Loop Highway — Lake 22 in great shape with recent trail improvements, parking fine on a weekday morning. Marten Creek passable to the log bridge; beyond the viewpoint the unmaintained section has severe washouts and ditch-crawling. Lake 22 / Marten Creek

↑ Back to top

THE LONG READ

The Dice Are Loaded: How a Linear PRNG Broke Slay the Spire 2

Players kept posting on Reddit and Discord that Neow's Bones was cursed — they kept rolling Debt, the worst curse in the game, at a rate that felt impossible. The standard response from the community was equally standard: confirmation bias. You remember the bad outcomes. The distribution is fine. The game is fine. And then someone actually looked at the code.

The problem, documented in a meticulous June 13 post on tck.mn, comes down to a single architectural mistake that cascades into dozens of broken probability distributions across an entire run of Slay the Spire 2. The game uses many separate pseudorandom number generators — one for combat card rewards, one for potion drops, one for enemy AI, one for each event — deliberately isolated so that what happens in a fight doesn't bleed into what you see in the next shop.¹ The design intent was sound. The implementation was not.

Each RNG is seeded by taking the run's base seed and adding a hash of a name string: seed + hash("niche"), seed + hash("combat_orbs"), and so on.¹ The seeds differ by fixed offsets. That would be fine with a well-designed PRNG. It is catastrophic with C#'s System.Random.

The root of the failure is algebraic. System.Random's constructor initializes its internal state array through a long chain of subtractions — and the difference of two linear functions is itself linear. Every entry in the state array ends up expressible as x * S + y, where S is the absolute value of the seed.¹ When you eventually call InternalSample(), you get SeedArray[locINext] - SeedArray[locINextp] — another subtraction of two linear functions. The first output of every RNG in the game is therefore exactly linear in the starting seed. Two RNGs whose seeds differ by a fixed offset d will have their first outputs differ by exactly x * d. The outputs are not independent. They are correlated by construction.

The consequences are specific and strange. The card Rebound cannot appear in the Trash Heap event — not "rarely," but literally never, because the act-selection RNG constrains the Trash Heap RNG such that the Rebound range is unreachable.¹ Players attempting to complete the in-game Compendium had noticed this and assumed a bug; they were right, just not about which bug. The Hefty Tablet relic, in Underdocks, has exactly three possible rare card outcomes rather than the full pool.¹ If you pick Precarious Shears from Neow and you're planning around the Tezcatara Ancient in Act 2, her first option is 88.75% likely to be Nutritious Soup — which matters enormously if you spent the run removing Strikes.¹ The first fight in Underdocks drops a potion 76% of the time; the same fight in Overgrowth drops one 4% of the time.¹ These are not small biases at the margin. They are structural.

The deeper technical point is about what changed between Slay the Spire 1 and 2, and why the fix introduced its own problem. In the first game, the RNGs were all initialized to the same state — a known flaw that the community documented and named CRNG. The sequel explicitly tried to address this by offsetting the seeds. But STS1 used modular integer arithmetic to select among options (roll a large integer, take the remainder), which meant correlations only emerged when the pool sizes shared factors. STS2 switched to floating-point: roll a number between 0 and 1, scale by pool size. That change, sensible on its own merits, made every RNG output give information about every other RNG output.¹ The fix for the original problem amplified the successor problem.

The post walks through the discovery, which was accidental: the author was building a seed-search tool to find runs where specific Neow transforms were available in Overgrowth. The search found plenty of Underdocks seeds. It found zero Overgrowth seeds. When the author printed the raw RNG output controlling act selection, it was always close to 0.1. That made no sense — unless the RNGs were correlated. A scatterplot of transform roll versus act roll confirmed it immediately.

The fix is straightforward. Replacing System.Random with any modern nonlinear PRNG — PCG32, xoshiro256, anything off the well-studied list — eliminates the correlation entirely. The author provides a 50-line PCG32 drop-in that would require a three-line change to the game's codebase.¹ The essay closes with a note that the game is still in Early Access, that Mega Crit is aware, and that the author is confident it will be addressed. The length of the post, the author acknowledges, is "wildly disproportionate to the seriousness and magnitude of the bug."¹ That's part of what makes it worth reading: the combination of technical rigor, dry humor, and genuine affection for understanding a system completely.

This is what good technical writing about games looks like. Not a patch note, not a complaint thread — a full forensic account of how a design decision, a standard library, and a mathematical property combined to produce something nobody intended.

↑ Back to top

FROM THE ARCHIVE

Sixty-Three Orbits Short: June 16, 1963

A lone figure in a bulky Soviet spacesuit hangs beneath a parachute canopy, descending through an immense empty sky. Far below, a spherical reentry capsule tumbles earthward on its own separate arc, trailing wisps of scorched heat-shield material. The cosmonaut's visor reflects nothing but pale light. Horizon is a faint smear of steppe at the bottom of the frame. Bold pen-and-ink linework, dense cross-hatching on the suit's shadow side and the capsule's curved hull, strong diagonal of the parachute shroud lines cutting across generous white space. No text, no color, no gradients.

She had never flown an aircraft. She had worked a loom at a textile factory in Yaroslavl since she was eighteen. What brought Valentina Tereshkova to the attention of the Soviet space program was that she liked to jump out of planes.

On June 16, 1963, Vostok 6 lifted off with Tereshkova aboard, making her the first woman in space.¹ By the time she parachuted back to earth three days later — ejecting from the capsule at 20,000 feet, the mandatory Vostok reentry procedure — she had completed 48 orbits and logged more time in space than all American astronauts combined to that date.¹

The selection logic was cold and practical. The Vostok reentry procedure required the cosmonaut to eject at altitude and descend by parachute separately from the capsule. Trained pilots were not the scarce resource; experienced parachutists were. Tereshkova had made her first jump at twenty-two under a local aviation club. By 1962 she had enough jumps on record that the Soviet program drafted her — along with three other female parachutists and one female pilot — into cosmonaut training.¹ Of the five, Tereshkova flew.

The propaganda calculus was explicit. The Soviets wanted another "space first" before the Americans, and a woman in orbit was available to them faster than almost anything else. Vostok 5 had launched two days earlier with Valeri Bykovsky; Vostok 6 went up while Bykovsky was still circling, putting two Soviet spacecraft aloft simultaneously. At one point the two capsules came within three miles of each other.¹ Tereshkova and Bykovsky exchanged brief radio communications; the cameras rolled.

What the propaganda machinery did not advertise: *Vostok 6*'s automatic control system guided the entire flight. Tereshkova never took manual control.¹ Soviet officials later acknowledged she experienced nausea and discomfort throughout much of the mission, though she completed it without incident.

The Americans had screened a group of female pilots in 1959 and 1960 — the so-called Mercury 13 — but NASA decided to restrict astronaut qualification to men. The first American woman in space, Sally Ride, flew on the shuttle Challenger in 1983.¹ That is a twenty-year gap, which is one way to measure the distance between a space race fought with parachutists and a space program that eventually got around to asking a different question.

Tereshkova received the Order of Lenin and the Hero of the Soviet Union. In November 1963, reportedly at Nikita Khrushchev's urging — who saw further propaganda value in the match — she married fellow cosmonaut Andrian Nikolayev.¹ They had a daughter and later separated. She went on to serve in the Supreme Soviet and represented the USSR at international women's conferences for decades. She never flew again. No female Soviet cosmonaut flew until the 1980s either.

The loom-to-orbit arc is the story that travels. What resists that framing is the system that produced it: a space program that needed a parachutist, found one who happened to be a woman, and then spent twenty years not needing another.

↑ Back to top

THE FUNNIES

Fix This Code / She Jumps

*After XKCD — on the government's claim that asking an AI to fix buggy code counts as a national-security jailbreak. After Mutts — on the sixty-third anniversary of Valentina Tereshkova's orbit: a textile worker who got to space because she already knew how to land.*

↑ Back to top

ALSO NOTED

Also Noted

• B-52 Crashes at Edwards, Eight Dead — A U.S. Air Force B-52 Stratofortress went down shortly after takeoff during a routine test mission at Edwards Air Force Base Monday, killing all eight crew members — military, government civilians, and two Boeing contractors — in the deadliest B-52 crash since 1982. kiro7.com Jun 16, 2026
• "Tokenomics" Is the New CFO Headache — Wired reports that roughly 300 companies addressed questions or concerns about AI tokens during earnings calls this spring — up from 93 a year ago — with executives at Cisco, Royal Bank of Canada, and Amplitude describing usage that has grown "pretty, pretty crazy," while some firms are responding with caps and model-selection systems to control the bill. wired.com Jun 2026
• AMD Silently Removed Memory Encryption From Consumer Ryzen CPUs — A months-long investigation by security researcher Ben Kilpatrick found that AMD's AGESA firmware update 1.2.7.0 disabled Transparent Secure Memory Encryption (TSME) on consumer Ryzen chips — a feature that still works on Pro versions — without documentation or explanation, leaving AMD engineers unable to say whether it was a deliberate policy decision or an accidental regression. arstechnica.com Jun 2026
• Iroh 1.0: Stable Networking by Key, Not IP — n0-computer shipped the 1.0 release of Iroh, a Rust networking library that routes connections by cryptographic key rather than IP address, adding QUIC multipath, NAT traversal, WASM support, and stable FFI for Python, Node.js, Swift, and Kotlin; the project's public relays have seen over 200 million endpoints in the past 30 days alone. iroh.computer Jun 15, 2026
• The Windows Emulator Team Patched Someone Else's Compiler — Raymond Chen's Old New Thing recounts the time a Windows x86 binary translator encountered a program that used 256 kilobytes of unrolled "write byte" instructions to initialize 64 kilobytes of data — code so offensive to the team that they added a special case to the translator to silently replace the function with a tight loop. devblogs.microsoft.com Jun 15, 2026
• datasette-agent 0.3a0 Adds Write Approval Gating — Simon Willison's datasette-agent tool gained a new `execute_write_sql` tool that prompts the user for approval before modifying a database, plus `--unsafe` and `--yes` flags for the chat CLI to auto-approve or run as root, enabling direct natural-language database modification with configurable trust levels. simonwillison.net Jun 15, 2026

↑ Back to top

THE QUESTION

When You Fix the Known Bug, How Do You Know What You Just Broke?

The problem with fixing a known failure is that it tends to assume you understood the failure well enough to correct it — and today's paper carries two stories, in entirely different domains, where that assumption quietly collapsed.

As THE LONG READ reports in detail, Slay the Spire 2 inherited a documented randomness problem from its predecessor. The first game's RNGs were all seeded identically — a known flaw, named, documented, discussed in the community for years.¹ The sequel deliberately addressed it by offsetting the seeds: each RNG now starts from a different value derived from the base run seed. The fix was correct in intent and wrong in consequence. The problem is that C#'s System.Random is linear, which means two seeds differing by a constant produce outputs differing by a proportional amount — so every RNG in the game is correlated with every other, not by accident but by construction.¹ The fix for the old correlation introduced a new one. Worse, the switch from integer arithmetic to floating-point made the correlations far more pervasive than they had been before.¹ The players who suspected something was wrong weren't wrong. The community's default response — confirmation bias, the distribution is fine — turned out to be what was mistaken.

THE PELOTON reports a structurally similar situation on a different substrate. Visma-Lease a Bike's Tour roster was disrupted when Christophe Laporte was injured. The fix was to bring in Per Strand Hagenes — a sound choice, a capable rider.² But at the Tour Auvergne-Rhône-Alpes this month, Hagenes was knocked out by illness before stage 7, alongside Jørgen Nordhagen, leaving Matteo Jorgenson with three domestiques for the finale.³ The replacement acquired the race's vulnerability. The system that was shored up failed in the same place it was shored up.

These two events are not the same thing. One is a software bug; the other is bad luck and biology. But they share a diagnostic problem: the patch looked sound at the moment it was applied. The offset seeds satisfied the design intent. Hagenes was a qualified substitute. Neither fix was negligent — both were reasonable responses to the available information. The failure only became visible later, and only because someone looked carefully at the structure underneath.

This is the question that persists after reading today's paper: when a known failure mode is corrected, what systematic check confirms the fix didn't introduce a new one? In software, you can plot the output and see the correlation in twenty seconds — if you know to look. In a race team, you learn it before the final stage. Van Aert's infected elbow compounds the picture in the same way: the wound was already being managed, then confirmed infected, and the next problem — missing the Tignes altitude camp, his Tour start now in genuine doubt — follows directly from the management of the previous one. The system keeps producing new failures from the residue of old ones.

The floating-point switch transformed a narrow correlation into a pervasive one. Systems don't usually break in the place you're watching. They break in the adjacent place you stopped watching because you just finished watching it.

↑ Back to top