THE PELOTON

Three Laps of Montjuïc: Stage 2 Tests the Puncheurs After an Emotional Barcelona Curtain-Raiser

↩ Developing story — first reported Jul 02 · previously Jul 03, Jul 04

BARCELONA — Stage 1 of the 2026 Tour de France is behind us, and the first account has been settled. Afterward, Jonas Vingegaard stood in the mixed zone and spoke about lying on the tarmac at the 2024 Itzulia Basque Country, not thinking about cycling — "just trying to survive" — and how long the road back has been. "I've been at times struggling in the last few years," he said. "Now I feel that I can close this chapter in the book. ... Coming from that to this point is emotional."¹

Visma-Lease a Bike built their team time trial around a clean division of labor: put the powerful rouleurs on the front over Barcelona's flat opening kilometers, shelter the climbers until Montjuïc, then let the legs speak. It worked as designed. Matteo Jorgenson was seen at the team bus punching the air.

Tadej Pogačar moved through the mixed zone quickly, noting he was "just super happy the day is over" — a TTT extracts its toll not in 20 minutes of racing but in hours of waiting, preparation, and anticipation beforehand. He did not leave empty-handed, posting the fastest split on the final two Montjuïc climbs.² "Good news: I have climbing legs," he laughed, before pivoting to Sunday. "We will fight for yellow in the next days, maybe already tomorrow. Tomorrow is a super hard, tricky stage, and I think we are ready, but we go day by day."

He is right. Stage 2 covers 168.5 kilometers from Tarragona to Barcelona, finishing with three ascending laps of the Côte de Château de Montjuïc. The kicker averages 9.3 percent and spikes to 13 at its sharpest; sprinters have nothing to race for here.³ Mathieu van der Poel, Tom Pidcock, and Mattias Skjelmose are the natural stage hunters on a circuit this explosive. GC leaders, exposed on a technical, punchy finish after a complicated final approach, could drop crucial seconds before the race has left Spain.

The race's first abandonment has already arrived: Clément Berthet (Groupama-FDJ United) will not start Stage 2 after crashing in Saturday's TTT.⁴ Groupama lose a climber before the Pyrenees.

ASO's course director has his eye on what follows. Wildfires are burning in Catalonia, tens of kilometers from the Stage 3 finish at Les Angles. Thierry Gouvenou described the situation plainly: "It is a major concern for us."⁵ Smoke development, emergency service deployments, and road access are all potential bottlenecks independent of the heat itself. Stage 4, from Carcassonne to Foix through Occitanie — the driest, hottest region of France in recent weeks — carries a local forecast of 40°C with a northwest headwind. Temperatures are expected to ease after Stage 5. Regional prefects hold formal authority to shorten or cancel stages if conditions demand it; that authority was established before the race left Barcelona.

At Saturday's TTT, UCI commissaires were running a parallel operation at the start ramp: stopping riders to confiscate ice socks stuffed down the back of skinsuits — cut-down tights packed with ice cubes, one of the peloton's oldest warm-weather tools. UCI Article 1.3.032 forbids items that modify a rider's morphology, and a commissaire explained to Cyclingnews that drawing a line, however small, is the only defensible position: allow a little and riders take more. Victor Campenaerts was among those stopped. Red Bull-Bora-Hansgrohe's head of engineering Dan Bigham noted the rule was already explicitly in force; Visma-Lease a Bike's head of performance equipment Jenco Drost put it simply: "Since last year, they're quite sharp on items under your suit."⁶ Enforcement had been uneven at the Tour de Suisse. In Barcelona on Saturday, it was not.

The ITA has put a date on cycling's long-pending power passport: 2028.⁷ Oliver Banuls, the ITA's head of testing, told Velo that researchers are already analyzing historical race and training files from more than 60 riders across five professional teams, with live data collection beginning next year through a study with the University of Kent and University College London. The tool would not directly sanction athletes — instead, unexplained shifts in a rider's longitudinal power profile would guide targeted testing, sample re-analysis, and intelligence investigations. "The purpose is getting the full picture," Banuls said. The idea has been around since 2015; Banuls acknowledged the technology simply wasn't ready then. "It won't be a revolution. It will be a great evolution."

Elise Chabbey (FDJ United-SUEZ) confirmed this week she will not start the Tour de France Femmes or race again in 2026. She has been absent since Liège-Bastogne-Liège in April, where she finished seventh as teammate Demi Vollering won. The reason, posted to Instagram: she is pregnant. "A little frog is growing inside me," she wrote.⁸ Chabbey, who won Strade Bianche in March and is a medical graduate who competed in kayak at the London Olympics, is expecting in January 2027 with partner Antoine Robin. She is signed with FDJ United-SUEZ through 2028.

↑ Back to top

THE WORLD

Chelan Hills Fire Reaches 20,000 Acres; Iran Doubles Down on Hormuz Fees

↩ Developing story — first reported Jun 29 · previously Jul 01, Jul 03

• Chelan Hills Fire, 20,000 Acres — Level 3 evacuations near Orondo; at least a dozen homes and buildings burned; dry conditions driving rapid spread near US 97. KIRO 7
• Iran Insists on Hormuz Fees — Tehran's ambassador told Beijing fees on Hormuz transit are "definite"; Rubio says any final deal will bar them. Times of Israel
• Seattle Transit Showdown, Monday — The transit measure amendment hearing is Monday July 6 at 11am. Saka's most provocative move would subject Metro's funding to annual Seattle City Council approval — effectively holding county bus service hostage year-by-year; full council vote expected July 16. PubliCola

ON THE TRAIL

PART 1 — WEEKEND PICKS

This coming weekend (Sat July 11 – Sun July 12). Independence Day 2026 fell on Saturday July 4; the federal observed holiday was Friday July 3. The long weekend (Fri–Sun, July 3–5) wraps up today — next trip window is next weekend.

Pick 1 — 1-night: Lower Tuscohatchie Lake via Pratt Lake

• Region: Snoqualmie Region > North Bend Area — ≈35–55 min from Issaquah - Length: 1-night - Mileage: Day 1 in ~9.5 mi / ~2,100 ft gain to camp near Lower Tuscohatchie Lake; Day 2 out ~9.5 mi - Weather (I-90 / Snoqualmie / North Bend): Saturday July 11 — high 68°F, mostly sunny, 13% precip. Sunday July 12 is outside the current 7-day forecast window. - Why it clears: No snow reported, lake and Tuscohatchie Creek for water, no bug complaints at all, light traffic even on the holiday weekend, no stream fords described. Classic Alpine Lakes granite-and-lake terrain at a fraction of the Enchantments crowds. - Trip report, Jul 3–4

Pick 2 — 2-night: De Roux Creek to Gallagher Head Lake

• Region: Snoqualmie Region > Salmon La Sac/Teanaway — ≈90–110 min from Issaquah - Length: 2-night (camp at or near Gallagher Head Lake) - Mileage: Day 1 in ~8.5 mi / ~1,800 ft gain to lake; Day 2 explore; Day 3 out ~8.5 mi - Weather (I-90 East / Teanaway/Cle Elum/Roslyn): Saturday July 11 — high 75°F, sunny, 3% precip. Forecast has this region as the clear winner for the weekend. - Why it clears: Snow-free, multiple streams and waterfalls for water, bugs described as merely "a little annoying" with zero hyperbole, genuinely zero other hikers on a sunny holiday morning, no fords described. Wildflowers peaking. - Trip report, Jul 4

PART 2 — REGIONAL SNAPSHOT

• Snoqualmie Pass — Twin Lakes/Mount Margaret: mosquitoes "awful" and relentless at Mount Margaret's summit even with a light breeze; Twin Lakes area significantly more manageable. Wildflowers and beargrass starting. - Mt Rainier / Sunrise — Glacier Basin (Emmons Moraine approach): mosquitoes "locked and loaded" at basin campsites.⁶ Burroughs Mountain: residual snow on trail to 3rd Burroughs, passable; wildflowers blooming; expect holiday crowds to ease next weekend. - Mountain Loop Hwy — Round Lake/Lost Creek Ridge: under a dozen people all day, spectacular views; bugs "plentiful" at the lake; upper snow slope still present but manageable with poles, no microspikes needed. - North Cascades / Hwy 20 — Blue Lake: "officially snow free," trail in very good condition with only a few shallow muddy spots.⁶ - Stevens Pass East — Rock Mountain via Rock Lake: steep and relentless, some snow patches in the final half-mile (passable without gear); nearly solitary — one other hiker seen all day. - Leavenworth / US 2 East — Icicle Ridge via Fourth of July Creek: 90°F in direct sun, no shade on descent, one hiker ran out of water at 3 miles from the car. Bring 3L+ if you go.

↑ Back to top

THE LAB

The Harness Is the Prior: Opus 4.8 Is Failing Third-Party Tool Schemas

Opus 4.8 is inventing tool call fields. Not emitting wrong values in the right keys — inventing keys wholesale: requireUnique, forceMatchCount, oldText2, event.0.additionalProperties, a rotating zoo of names that changes with each invocation. The actual oldText and newText values are byte-correct. The schema around them is not.¹

A post at lucumr.pocoo.org traced this to a Pi editor issue filed against Opus 4.8 and Sonnet 5. Older Anthropic models don't show it. Codex models tested so far don't either. It only surfaces in multi-turn agentic histories — the model has read files, formed a diagnosis, is composing a multi-line edit — and even then only in some transcripts. A fresh single-turn prompt doesn't reproduce it. In one affected session, continuing the conversation caused Opus 4.8 to produce malformed calls roughly 20% of the time.¹ Stripping extended thinking blocks from history halved the failure rate. Enabling Anthropic's strict tool invocation eliminated it.

The failure lands exactly where you'd expect if this were a training artifact. Pi's edit tool uses a nested edits[] array, each element with oldText and newText. Claude Code's own edit tool is flat: file_path, old_string, new_string, one optional flag. After generating several hundred tokens of escaped newText content inside a nested JSON object, the model faces a binary decision: } or , "something". Opus 4.8's prior says an edit call looks like Claude Code's schema — possibly with one extra field. It has no trained name for Pi's nested shape, so it samples a plausible-sounding key fresh each time. That's why the failures don't cluster on a single stable wrong alias: the model has no consistent wrong answer, only a strong prior that something probably goes there.

The mechanism is documented in Claude Code's own minified, closed-source client. The harness silently drops unknown keys. It aliases old_str to old_string and path to file_path. It has Unicode escape repair for broken sequences. It has retry logic for malformed calls.¹ When reinforcement learning happens in an environment that absorbs this much slop without penalizing it, there is no gradient against adding a stray field — the model emits "requireUnique": true, the harness absorbs it, the task completes, the reward registers. Meanwhile, Anthropic's strict mode — which would constrain the sampler to the declared schema — imposes complexity limits on tool definitions that cause API requests to fail, which is presumably why Claude Code itself doesn't use it.

The uncomfortable implication is that tool schemas are no longer neutral contracts. Some shapes are on-distribution for these models; others are increasingly off. The more post-training concentrates in one dominant, forgiving, opaque harness, the more every other harness has to either closely match it or rely on strict-mode constraints as a workaround. You cannot inspect the harness to know how far off-distribution your schema is. You observe the symptoms after the fact — a 20% corruption rate in production, a zoo of invented keys, a model that was better at this six months ago.

Simon Willison shipped sqlite-utils 4.0rc2 today, the bulk of it written by Claude Fable across 37 prompts, 34 commits, and +1,321/-190 lines across 30 files.² The model's first significant find in 4.0rc1 was a data-loss bug: Table.delete_where() ran its DELETE with no atomic() wrapper, leaving connection.in_transaction = True. Every subsequent atomic() call then took the savepoint branch and never committed — meaning the delete, and any writes following it, were silently rolled back when the connection closed. Willison notes the bug would have been fixable in a 4.0.1 point release rather than requiring a major version bump, but is glad he didn't ship it.

The more durable methodological note is the cross-model review. Willison prompted Codex Desktop and GPT-5.5 to review the changes and confirm the changelog was current. GPT-5.5 found two more issues: db.query() raising ValueError after already committing a write as a side effect (the rejection arriving too late to be a rejection), and INSERT ... RETURNING only committing once the generator was fully exhausted — contradicting the changelog's guarantee that writes take effect without iteration. He pasted the findings into a fresh Fable session to confirm and fix both. His summary: "I've started habitually having Anthropic's best model review OpenAI's work and vice versa, because I've had that turn up interesting results often enough to be valuable."

The estimated unsubsidized API cost for the sessions was $149.25 — $141.02 in the main session, the remainder split across five sub-agents.² The July 7 deadline looms: Anthropic ends subsidized Fable access for Max subscribers and reverts to full API pricing.

One data point from Epoch.ai: high- and critical-severity CVE counts jumped more than 3.5x in June compared to the previous monthly record, in the period following Anthropic's April 2026 Claude Mythos Preview announcement and the subsequent Glasswing and Daybreak vulnerability-discovery launches from Anthropic and OpenAI.³ Epoch frames this as correlation with the announcements — whether the spike reflects AI-accelerated discovery, a shift in researcher reporting patterns, or both is not established by the analysis.

Trending today: dominated by Claude Code CLAUDE.md skills collections, caveman-prompt token reducers, and AI agent wrappers — the one technically substantive outlier is ammaarreshi/Generals-Mac-iOS-iPad, a native macOS/iOS/iPadOS port of C&C Generals Zero Hour built on the EA GPL v3 source via GeneralsX with a DXVK/MoltenVK renderer (primary gaming coverage in ALSO NOTED today).

↑ Back to top

THE LONG READ

How America Beat the Screwworm — and Why It's Back

On June 3 of this year, screwworms were found in a three-week-old calf near La Pryor, Texas.¹ Dozens more cases have since been confirmed in Texas and New Mexico. Outside a minor 2016 outbreak in the Florida Keys that was quickly stamped out, this is the first confirmed screwworm infestation on US soil since the 1980s. The fly is back.

Most people alive today have no memory of what that means. The New World Screwworm (Cochliomyia hominivorax) does something unusual among flies: its larvae feed not on dead tissue but on living flesh. A female lays her eggs on any open wound — a nick from a fence post, a freshly branded hide, a navel on a newborn calf. The eggs hatch into worms that burrow inward as they eat, enlarging the wound, drawing more flies, until the infection is fatal. Ranchers in 1930s Texas hired hands whose sole job was checking livestock every two days. "People would not leave home for more than a day," one history of the era notes, "for fear of finding their animals had been eaten alive while they were away." A 1935 USDA survey found more than 1.2 million infections and 180,000 dead livestock in Texas alone — and acknowledged the true totals were likely far higher.¹ In the same decade, an estimated 60 to 80 percent of white-tailed deer in Texas died from the fly. By the early 1960s the pest was inflicting more than $100 million in annual damage in the Southwest.

The solution, when it arrived, was one of the most audacious ideas ever applied to pest control.

Edward Knipling joined the USDA in 1931 and spent the following years studying screwworms in exhaustive, sometimes ridiculous, detail — at one point staking out a wounded goat from dawn to dusk for a week straight, marking individual females with fingernail polish to track their movements. By the late 1930s, stationed at Raymond Bushland's laboratory in Texas, Knipling had noticed something useful: despite the millions of infections screwworm caused, the actual number of wild flies in any given area was surprisingly small — roughly 100 per square mile. Meanwhile, Bushland had developed an artificial growth medium (hamburger, blood, water, and a little formaldehyde to delay putrefaction) that let them raise screwworms by the thousands in a lab. The combination pointed at a question neither had fully articulated yet: what if you could overwhelm the wild population with sterile males?

Female screwworm flies mate only once. If sterile males vastly outnumber wild males, most females will produce no viable offspring. Sustain that imbalance long enough, and the population breeds itself out of existence. When Knipling and Bushland floated the idea to colleagues, the reaction was derision. "Who ever heard of castrating flies?" The two kept it to themselves.

The missing piece arrived in 1950. Knipling read a paper by Hermann Muller, who had won the Nobel Prize for demonstrating that radiation induces heritable mutations. The paper was a warning against nuclear warfare; Knipling saw a sterilization tool. He contacted Muller, who replied: "I know nothing of screwworms but your theory is sound." Bushland, working nights and weekends with no dedicated research funds, got access to an Army hospital X-ray machine at no cost, then borrowed a sample of highly radioactive cobalt-60 from Oak Ridge National Lab. High enough doses left male screwworms completely sterile while otherwise intact and functional — they would still seek out females and mate, they would simply produce nothing.

The decisive field trial was on Curaçao, a Dutch Caribbean island off the coast of Venezuela that was overrun with screwworm and, crucially, isolated enough to test whether full elimination was actually achievable. In the summer of 1954, a USDA team began dropping boxes of sterilized flies from a small plane over the island. The first weeks were poor: too few flies, too little effect. When the density was raised from 100 to 400 flies per square mile, results improved rapidly. Within 14 weeks, no viable screwworm offspring could be found anywhere on the island. In November 1954, Curaçao was declared screwworm-free.¹

Florida followed. In April 1958, while cold temperatures had confined screwworms to the southern half of the state, USDA planes began dropping sterile flies across a 100-mile-wide corridor through the center — a barrier. They then pushed south, aided by a new factory in Sebring capable of producing 50 million sterilized flies a week. By September, infections in the southeast had fallen to near zero. By February 1959, cases were zero.

Texas was another matter. Florida is surrounded on three sides by water; Texas shares more than a thousand miles of border with Mexico, where screwworm survives year-round. The USDA officially concluded in 1959 that Southwest eradication "did not appear feasible." What broke the logjam was money and political will: a coalition of Texas ranchers organized, raised millions in voluntary donations, persuaded the state legislature to fund a program, and ultimately got Vice President Lyndon Johnson — a ranch owner who knew the scourge personally — to force the federal government's hand. A new insectary was built at an abandoned Air Force base near Mission, Texas: a 76,000-square-foot operation running 24 hours a day, 365 days a year, producing more than 200 million sterilized screwworm flies a week.¹ The facility was described by observers as a grotesque marvel — trays of blood and meat moving through the building on a monorail, timed to the fly's lifecycle, "a seething mass that is difficult to believe unless you've seen and smelled it." By 1966, the screwworm had been eliminated from the entire United States.

What followed over the next two decades was an extension of the barrier southward, decade by decade, through Mexico and then through Central America — Guatemala in 1988, Belize the same year, Honduras, El Salvador, and Nicaragua in 1991, Costa Rica in 1993, Panama in 1994. The logic was relentless: the farther south the barrier, the narrower the land. At the Darien Gap, the dense and roadless jungle on the Colombia-Panama border, the barrier needed to be only 60 miles wide. A joint US-Panamanian organization called COPEG took over administration. The program cost about $15 million a year to operate. For that price, the entire livestock industry of North and Central America was shielded from a parasite that had once killed animals by the millions.

Then, sometime around 2023, the barrier failed.

The causes are multiple and, viewed together, predictable. COVID-19 kept inspectors home, disrupted vehicle maintenance, cut supply chains; power outages at the Panama production facility killed millions of sterilized flies. The Darien Gap itself, once dense rainforest essentially impassable to cattle, has been steadily cleared for grazing — a transformation driven partly by absentee ranchers paying insufficient attention to their herds, partly by narcotics cartel money-laundering operations that traffic hundreds of thousands of cattle through Central America outside any inspection regime. Screwworms, the evidence suggests, were moving by truck. By 2023, there were 6,500 screwworm cases in Panama despite the barrier; by 2024, 18,000 in Panama, 8,600 in Costa Rica, 3,300 in Nicaragua. By 2025, reported cases in Mexico exceeded 12,000 and have since passed 30,000.¹

The deeper failure is institutional. The Mexican production facility at Tuxtla Gutiérrez, which once produced more than 400 million sterile flies a week, was shut down in 2012 — over COPEG's objections — because the Darien Gap barrier was working so well that the extra production capacity seemed redundant.¹ Panama's production was allowed to decline to the level needed to maintain the barrier, not to respond to any breach of it. What had been a heavily resourced, vigilant program became a thin operation calibrated for steady-state conditions. When those conditions collapsed, there was no reserve capacity to respond.

The USDA is now rebuilding. In 2026 it broke ground on a new facility at Moore Air Force Base in Texas, designed to produce 300 million sterile flies a week. A $100 million "Grand Challenge" has been announced for new eradication and treatment methods. But experts caution that re-eradicating screwworm from North and Central America will require "close to a decade of sustained work." The fly spread north by truck; getting it back out will happen by plane, drop by drop, week after week, for years.

The screwworm story is a case study in how completely a problem can be solved, and how completely that solution can be undone by its own success. A program that had protected an entire continent for four decades and cost virtually nothing relative to the damage it prevented was allowed to atrophy because the threat it held back had become invisible. No one alive remembered what $100 million in annual losses to a flesh-eating parasite looked like. They are about to find out.

↑ Back to top

FROM THE ARCHIVE

The Joke That Named Her: July 5, 1996

A fluorescent-lit research laboratory at the Roslin Institute, Edinburgh, 1996. A newborn lamb — damp wool still curling, legs folded beneath her — lies on a stainless steel examining table at the center of the frame. Two scientists in lab coats lean in from either side: one holds a clipboard at chest height, the other reaches forward with gloved hands steadying the lamb's head. Glass-fronted equipment cabinets crowd the background — centrifuge rotors, microscope stands, rows of labeled specimen vials — all rendered in tight parallel hatching. Overhead strip lighting throws hard downward shadows across the scientists' faces and the steel table edge. On the bench behind, a single glass petri dish sits alone at the margin of the light, small against the creature it produced. The lamb's woolly coat is cross-hatched in loose curves, soft and alive against the rigid clinical geometry of the room. Strong diagonal composition from lower-left table corner to upper-right cabinet shelf. Black-and-white pen-and-ink editorial illustration. Confident linework, controlled hatching and cross-hatching in shadow areas, generous white space in upper corners. No color, no gradients, no text.

The stockman learned the lamb had been cloned from a mammary cell and had one obvious suggestion for her name. The joke stuck: 6LL3, the lab's internal code, became Dolly — after Dolly Parton — and the most consequential biological breakthrough of the decade was christened with a punchline.

She was born July 5, 1996, at the Roslin Institute outside Edinburgh: the first mammal successfully cloned from an adult somatic cell.¹ The donor cell came from the udder of a six-year-old ewe. Scientists cultured it using microscopic needles adapted from human fertility techniques first developed in the 1970s, implanted the resulting embryo into a surrogate, and 148 days later Dolly arrived.¹ The Roslin team sat on the news for seven months before announcing publicly in February 1997.¹ When they did, the controversy was immediate — supporters pointed to xenotransplantation and therapeutic cloning for degenerative diseases like Alzheimer's and Parkinson's; critics worried about safety and the obvious next step no one was supposed to say aloud.

Dolly mated with a ram named David and produced four lambs.¹ In January 2002 arthritis appeared in her hind legs, raising early questions about whether the cloning process had introduced genetic abnormalities. A progressive lung disease followed. She was put down on February 14, 2003, at six years old — the same age as the ewe whose udder cell had made her.¹

She was stuffed and placed on display at the National Museum of Scotland in Edinburgh, where she remains.

↑ Back to top

THE FUNNIES

Ice Socks and Fly Factories

After Peanuts — on the UCI commissaire who confiscated a rider's ice sock on a forty-degree stage day, then stood in the sun without one. After The Far Side — on the USDA sterile-fly facility that ran around the clock producing two hundred million flies a week, and what it looked like to the person who finally saw it in person.

↑ Back to top

ALSO NOTED

Also Noted

• Follow-up: Dygert's 2026 Season Ends With Two Surgeries Pending — What was left unresolved in June is now answered: Chloé Dygert's attempted comeback lasted about two rides before new diagnoses arrived — RED-S, a virus, blood markers trending wrong, shoulder damage from her Paris-Roubaix crash that will require surgery after rehab failed, and a separately discovered collapsed nasal valve that also needs surgery. cyclingnews.com Jul 5, 2026
• Oura Ring 5: Genuinely Smaller, Sports Tracking Still Broken — DCRainmaker's month-long review confirms the Ring 5 is meaningfully smaller (2g vs. up to 5.2g, width down from 7.9mm to 6.1mm) with modest sensor accuracy gains, but the sports-tracking situation is worse than it sounds: Oura still won't push heart rate data to Strava, and if you pair an external HR strap to fill the gap, the app drops the entire workout file from third-party sync. dcrainmaker.com Jul 3, 2026
• Dan Luu's Empirical Case Against AI Coding Benchmarks — Benchmarks across GPT-5.5, Opus 4.8, and Fable show run-to-run variance so high that opposite conclusions about which model is best can be drawn from the same benchmark suite depending on which tasks or runs you select — and empirical testing of caveman mode (the viral token-reduction technique whose creator called it a joke) shows no reliable win across tasks despite widespread workplace recommendations. danluu.com Jul 3, 2026
• C&C Generals Zero Hour Runs Natively on Apple Silicon — Ammaar Reshi's port, with Claude Code (Fable) as the primary engineer, runs the real 2003 engine through a DirectX 8 → DXVK → Vulkan → MoltenVK → Metal pipeline on macOS, iPhone, and iPad, with touch RTS controls and a full engineering log; you supply your own Steam copy (~$5 on sale). github.com Jul 4, 2026
• Immich v3.0.0 — The self-hosted photo library's first major version adds non-destructive mobile editing, a drag-and-drop workflow automation builder, real-time HLS video transcoding (preview), improved Android background backup, and filesystem integrity checks; breaking changes affect third-party API integrations, not standard Docker upgrades. immich.app Jul 1, 2026
• Apparent Claude Code Session Leakage in Enterprise ZDR — An Enterprise ZDR user filed a GitHub issue reporting that mid-session Claude Code began confidently discussing a Minecraft temple and asking about preferred brick types; the reporter acknowledges an unusual multi-directory session setup that may explain the behaviour, and no root cause has been established by Anthropic. github.com Jul 4, 2026
• ESO: Proposed Mega-Constellations Are an Existential Threat to Ground Astronomy — A peer-reviewed ESO study finds the currently proposed 1.7 million satellites — including SpaceX's planned space data-center fleet and Reflect Orbital's 50,000-mirror constellation, which alone would brighten the entire night sky 3–4x — would make long-exposure ground-based astronomy untenable; the paper puts the safe total at 100,000 faint satellites and is the basis for ESO and IAU comments to the FCC. eso.org Jul 1, 2026
• JWST: Too Many Theories Now, Not Too Few Galaxies — After years of puzzling over why Webb found unexpectedly many bright early galaxies, astrophysicists now have multiple competing explanations — burst star formation, extremely massive early stars, diverse early galaxy properties confirmed by MIRI — prompting the observation that "we've almost gone from having too many early galaxies to having too many theories to explain them." quantamagazine.org Jul 2, 2026
• Rhythm Heaven Groove — Nintendo's first Rhythm Heaven in more than a decade reviewed positively on Switch but reviewers warn that lag on modern TVs requires playing in handheld or tabletop mode for accurate timing. opencritic.com Jul 2, 2026

↑ Back to top

THE QUESTION

If the Harness Absorbs the Error, Where Did the Signal Go?

How do you know a capability is real, if the environment that trained it also quietly hid every time it failed?

THE LAB section today traces an uncomfortable answer in Opus 4.8: the model invents tool call fields during multi-turn agentic sessions, and the mechanism appears to trace to Claude Code's own training harness — which aliases wrong keys, drops unknown fields, retries malformed calls, and generally absorbs schema errors rather than penalizing them.¹ When reinforcement learning happens in that environment, there is no training signal against inventing a stray field. The task completes; the reward registers; the behavior goes uncorrected. You observe the consequence only after deployment: invented keys that change with every invocation, a corruption rate that surfaces only under extended multi-turn load, a model that was more reliable a few months ago and no clear account of why.

The structure appears elsewhere today. THE LONG READ is about screwworm, not software, but the failure mode is identical. A decades-long eradication program rendered the fly invisible across North and Central America — which is precisely why reserve production capacity was allowed to wind down. The Mexican sterile-fly facility was shut down because the barrier was working so well that extra capacity seemed redundant.² When conditions collapsed simultaneously, there was no reserve to deploy. The program's success had made the backup infrastructure feel wasteful; when the success ended, the infrastructure was gone.

In both cases, the absorber — the training harness, the eradication barrier — is also the instrument measuring the system's health. The harness absorbs schema errors, so model performance looks fine in the environment where you measure it. The barrier suppresses screwworm, so the livestock sector looks healthy in a world where the threat has been made invisible. Both measurements are accurate in the narrow sense. Neither gives you any signal about how much underlying capacity remains.

The question to carry: for any system that succeeds by absorbing errors rather than eliminating them, how far into atrophy does the underlying capability get before the absorber gives way? The signals tend to arrive late and cluster. In the fly case, it was thousands of confirmed cases in Panama despite an operational barrier — already a lagging indicator of a problem developing for years.² In the tool-schema case, it's a 20% corruption rate in extended sessions, a rotating zoo of invented keys, and the quiet note that strict mode would fix it but strict mode breaks API requests.¹ By the time the failure is legible, the reserve that would have caught it has already been dismantled.

↑ Back to top