THE LAB

Bevy 0.19 Rewires the Renderer: GPU-Driven Pipelines, a New Scene Language, and the RenderGraph Finally Dies

↩ Developing story — first reported Jun 19

261 contributors and 1,185 pull requests later, Bevy 0.19 landed yesterday with changes deep enough to shift how you think about both game architecture and GPU workloads.¹ The headline feature is BSN — Bevy Scene Notation — a years-in-the-making replacement for the bundle/spawn pattern that has been the engine's primary friction point since day one. But the renderer work underneath is arguably just as consequential.

Start with the benchmarks. On a laptop with a mobile RTX 4090, the many_cubes test with 1.6 million entities and 116,000 in view went from 49.47ms per frame to 18.77ms — 21 FPS to 53 FPS. With culling off and all 1.6 million cubes rendered: 93.1ms to 41.2ms. The city scene benchmark (55,000 rendered entities) dropped from 19.3ms to 11.8ms static, 22.1ms to 16.2ms with motion.¹ These are not incremental gains.

The work behind those numbers spans the entire pipeline. Batch-set unpacking for multi-draw-indirect moves from CPU to GPU, saving nearly 1.5ms at a million cubes. GPU clustering for lights improved the many_lights benchmark roughly 20x. Parallel mesh collection saves around 4ms at 200,000 moving meshes. Sparse mesh uniform buffer uploads — only changed uniforms go to the GPU — help scenes with many static objects disproportionately.

The partial bindless work deserves attention separately. WGPU's Metal backend supports texture binding arrays but not buffer binding arrays. Bevy previously required both before enabling bindless at all, which meant Apple hardware — no Vulkan, no fallback — never got the fast path. The fix is per-material: if your material only needs texture arrays (which includes StandardMaterial), it uses bindless on Metal. Benchmarked on the Bistro Exterior scene with 698 materials: M2 Max gets a 15% frame time reduction and drops 57 MB of RAM; an NVIDIA 5060 Ti improves 46%.¹

The RenderGraph is gone. Render passes are now regular ECS systems running in schedules. What previously required implementing a Node trait, deriving a RenderLabel, and wiring edges through a targeted API is now add_systems(Core3d, my_render_fn.after(Core3dSystems::MainPass)). It's a clean deletion of an abstraction that outlived its justification — when the ECS matured enough to express the same ordering constraints directly, the graph became redundant scaffolding.

BSN itself is the other structural change. The old pattern — spawn a bundle, manually thread in asset handles, hope callers know what dependencies a component needs — gets replaced by a macro-based scene syntax where asset paths are first-class, components compose as patches, and spawning a Player guarantees its full scene (hands, weapons, whatever the component declares) is present at spawn time. The SceneComponent trait enforces this: if Player derives SceneComponent, you can query for Player and assume the declared children exist. That invariant was not achievable before without external conventions.

Other changes in the release worth noting: contact shadows via screen-space raycast fill the short-range shadow gap that shadow maps can't solve without prohibitive texture cost; resources are now stored as components on singleton entities, unifying ECS internals and making lifecycle hooks, observers, and relations available on resources; contiguous_iter exposes raw table slices to the SIMD path, delivering a 3.5x speedup over normal iteration with AVX2 on bulk transform updates; render error recovery gives games a typed handler for device loss, OOM, and validation errors instead of a silent hang.

As this paper reported Friday, Tim Sweeney announced at Unreal Fest that UE6 will replace Git with a new system called LORE. That repo is now public. Epic shipped EpicGames/lore on GitHub under MIT, with docs, a quickstart, and SDKs for JavaScript, Python, C#, and Go.

The architecture is content-addressed storage over Merkle trees — closer to Git's object model than to Perforce's lock-and-trunk approach, but with chunked storage for large binary assets, on-demand sparse workspace hydration, and centralized service backing for team scale. The design doc describes it as optimized for binary-first deduplication: files are stored as reusable chunks with indexed lookup, so a 4 GB texture asset updated by one pixel doesn't re-upload the whole thing.

One caveat in the README is worth flagging: LORE is the built-in VCS for UEFN (Unreal Editor for Fortnite), but today's open-source build can't yet talk to the UEFN server because the production build uses a proprietary compression format.² Epic says they are migrating UEFN to the same open format used by the public repo. Until that gap closes, the UEFN and open-source deployments are separate. LORE is pre-1.0 and marked under active development; on-disk formats may change.

On the LLM-as-SRE front: Lorin Hochstein published a pointed essay arguing that LLM-written incident reports are more dangerous than LLM-generated code.³ The reasoning is structural. Incorrect code fails tests. An incorrect operational diagnosis either resolves the incident or it doesn't. But an incident report has no immediate correctness signal — a plausible-sounding narrative that invents couplings between systems, or misses a critical interaction, gets filed, read, nodded at, and archived. Nobody pushed back because nobody did the synthesis work that would have surfaced the contradiction. The problem compounds: the same organizations offloading report generation to LLMs will likely use LLMs to summarize them too. Hochstein draws on Dick Guindon's observation that writing is how you discover you didn't understand something — outsourcing that step doesn't skip the cognitive work, it just skips the discovery that the work wasn't done.

Today's THE LONG READ covers the reverse engineering of Qualcomm's NPU compiler — including the discovery that VTCM tensor placement is solved as a Mixed Integer Linear Program using HiGHS, that the compiler silently downcasts precision during placement to relieve memory pressure, and that there's a hidden analytical simulator called Hextimate inside the toolchain.⁴ Worth a full read if you do edge ML work.

Trending today: GitHub is saturated with AI agent wrappers, CLAUDE.md skills collections, and LLM scaffolding repos — EpicGames/lore is the one technically substantive repo on the list, and it's covered above.

↑ Back to top

THE PELOTON

Reusser Takes Yellow in Aarburg; SafeR Under Fire After Žigart's Fractured Jaw

↩ Developing story — first reported Jun 18 · previously Jun 19

AARBURG, Switzerland — Marlen Reusser clocked the fastest time of the day through the 23.7-kilometre test around Aarburg and pulled on the yellow jersey of the Tour de Suisse Women with one stage remaining, displacing Elisa Longo Borghini by a margin of just ten seconds.

The world ITT champion, riding for Movistar, stopped the clock at 29:36 — 11 seconds clear of Zoe Bäckstedt, who had herself been on course for a second consecutive stage victory after winning the sprint in Bad Ragaz on Friday.¹ Bäckstedt led into the intermediate check by 25 seconds over the field, but Reusser was five seconds faster at that split and kept the gap growing to the line.¹ Loes Adegeest (Lidl-Trek) took third on the stage, 54 seconds down. Longo Borghini rode in fifth, a minute and four seconds behind Reusser — enough to surrender the race lead.

Reusser described going to the limit from the gun: "It was very hard, really very, very hard. I went really hard from the start and I was suffering a lot. You go fast, but you don't really know if the others go faster."¹ She remained cautious about Sunday's mountain finale to Villars-sur-Ollon, which includes two ascents of the Col de la Croix. "I think one weak moment can change the whole GC, not only getting me out of the yellow, but for everybody. Tomorrow is going to be super tough."¹

The ten-second margin is thin enough that the GC is not settled. Kerbaol (EF Education-Oatly) sits third at 1:20, with Sarah Van Dam (Visma | Lease a Bike) fourth at 1:35. The final stage promises to reshuffle the order.

On the men's side, the Stage 4 ITT in Aarburg started with Pogačar last off the ramp in yellow, carrying a 2:50 cushion over Richard Carapaz.² Carapaz started a minute ahead of the race leader; third-placed Andrea Bagioli at 3:07 and Mathias Vacek at 4:16 rounded out the top tier of GC threats. Roglic, slipped to 16th overall at 6:25, started at 16:48 — outside the yellow jersey equation but with time to recover against rivals below him.

The crash that hospitalised Urška Žigart during Stage 2 is now carrying institutional weight.³ The CPA Women — the UCI-affiliated riders' union — called for a "substantial review" of SafeR, the multi-stakeholder road safety project established in 2023, after Žigart suffered a fractured jaw when she went down over an uneven road surface in the final kilometre of Thursday's stage. Several riders crashed at the same spot going at high speed; Žigart's injury was the most severe.

Managing director Alessandra Cappellotto stopped short of abandoning the project but questioned whether its current committee structure was catching course hazards before they injured riders: "Serious incidents continue to occur, and it is our responsibility to ask whether the current system is providing all the answers that our sport requires."³ SafeR operates three committees — a quarterly supervisory board, a monthly technical commission, and a weekly case management committee that reviews incidents and audits upcoming races.³ The CPA Women's position is that the system's framework needs sharpening rather than scrapping.

The criticism arrives at a bad moment for SafeR. The project has been under fire since it emerged the UCI used SafeR funds to finance its legal battle against SRAM over gear ratio restrictions — a case the UCI subsequently lost on appeal.³ Three years into SafeR's existence, the UCI has not published data demonstrating a reduction in crashes or accidents.

The Tour de France sprinter selection at Soudal-QuickStep remains deliberately unresolved. CEO Jurgen Foré, speaking after the Giro d'Italia, reiterated that Tim Merlier remains the team's designated Tour sprinter while 22-year-old Paul Magnier — who won three Giro stages and the points jersey — continues building toward the biggest race of the year. "That is a luxury problem that we will have to solve later," Foré said of any future switch.⁴ Merlier holds five wins this season after a knee injury delayed his spring start; the team sees no reason to unsettle the plan now.⁴ As this paper reported last week, Magnier raced Circuit Franco-Belge in Belgium the day before Tour selection sharpened, finishing third — a reminder of how close the margin between them has become.

↑ Back to top

THE WORLD

USA Advances at Home; Seattle Rattles the Seismometers

• USA advances — Without Pulisic, the U.S. beat Australia 2-0 at Seattle Stadium, punching their knockout-round ticket. Next match: Turkey, Los Angeles, Thursday. Source
• Seismic confirmation — PNSN installed seven strong-motion seismometers around Seattle Stadium; fan noise registers as measurable seismic signals during each of the six Seattle matches. Source

Mayor Katie Wilson rallied with transit advocates Thursday to defend her proposal to renew the Seattle Transit Measure with an expanded 0.15% sales tax increase. Council debate continues over the coming weeks. Source

In the year since Tacoma adopted new zoning standards last February, housing permits entering the pipeline jumped 62% compared to the five-year average — an early signal that the 2025 overhaul is moving supply in the right direction. Source

King County has issued a burn ban starting June 22. Fireworks remain illegal in unincorporated areas.

ON THE TRAIL

Weekend Picks — Sat–Sun Jun 21–22

Weather is excellent region-wide: sunny and 73–79°F in the lowlands, 60–65°F at Rainier, 62–69°F at Stevens Pass. No rain through the weekend.

• Moonshine Lake via Cathedral Pass (Teanaway, 90–110 min) — A Jun 19 overnight report describes a solitary camp at Moonshine Lake, base of Goat Mountain-North Peak. The Cathedral Pass and Lake Michael trails have blowdowns (some significant) but all passable. No snow until the summit push, where gaiters and spikes were used on the upper gullies — the lake and approach are manageable in trail runners. 15.5 mi RT, 4,186 ft gain. Scatter Creek crossing was flowing but passable. Road (Cle Elum/Tucquala) is in fair shape with washboard and potholes; small trailers made it through. Quiet enough to have the lake entirely to yourself.
• Sheep Lake to Sourdough Gap (Chinook Pass/Hwy 410, 110–140 min) — Multiple Jun 19 reports confirm the trail is snow-free to Sheep Lake and Sourdough Gap; a small snow section begins about 0.5 mi past the gap toward Crystal Lakes, with mosquitoes getting bad in that stretch. Stay at or before the gap for a bug-minimal, water-accessible overnight. Wildflowers blooming. Rainier area forecast: 60°F Saturday, 65°F Sunday, sunny both days.
• DeRoux Peak (Teanaway, 90–110 min) — No snow on the trail or peak (a Jun 18 report notes only a couple of small patches visible on Esmerelda and Hawkins from a distance). Trail in good shape to the pass; the peak route east is straightforward though piecemeal past the saddle. Panoramic views of Stuart, Rainier, Lemah, and Goat Rocks. Low crowds — two groups spotted. Note: motorbikers are appearing on western Teanaway trails due to Labor fire closures rerouting traffic. Saturday forecast: 72°F sunny.

Regional Snapshot

• Issaquah Alps — Cougar Mountain trails quiet and in good shape; salmonberries finishing, some trails slightly overgrown through grasses. Zero bug complaints. 10–25 min from home. - Snoqualmie / North Bend — Active bear sightings on the Snoqualmie Lake trail. Avoid the PCT Section J between Snoqualmie Pass and Stevens Pass: a massive rockslide just south of Lake Susan Jane (multiple chair-sized boulders, unstable) has rendered ~100 feet of trail dangerous — a trail steward is recommending closure until cleared. - Snoqualmie Pass — Kendall Katwalk's steep snowfield before the Katwalk has melted off; trail now on dirt. Gem Lake fully snow-free. Both trails seeing very high weekend traffic. - Teanaway — Roads in notably better condition than prior years (fire crew grading credited). Snow-free on DeRoux and Moonshine Lake approach. Lake Ingalls still has a snowy talus section requiring poles and solid footwear. - Rainier Area (Chinook Pass/Hwy 410) — Sheep Lake and Sourdough Gap accessible and snow-free. Huckleberry Creek trail has 14 large blowdowns below 3,700 ft; expect slow going. Owyhigh Lakes: ~10 trees down, mosquitoes at lakeside. - Mountain Loop Highway — Del Campo Peak has two snow gullies on the summit requiring spikes for descent; basin approach is snow-free. Heather Lake trail wet with roots and rocks; manageable but bring poles.

↑ Back to top

THE LONG READ

What Qualcomm Didn't Document: A Solo Engineer Opens the Black Box

The documentation wasn't bad. It didn't exist. Qualcomm publishes essentially nothing about the internal behavior of its Hexagon NPU compiler — not how it decides what fits on-chip, not how it handles memory pressure, not what tradeoffs it makes when your model is too large. A solo engineer working on edge ML deployment, frustrated enough to consider quitting, instead pointed Ghidra at Qualcomm's shared libraries and started reading machine code.

What came out is one of the more quietly remarkable pieces of reverse engineering published this year. The engineer — writing under the handle datavorous — targeted libHtpPrepare.so from QAIRT SDK version 2.46.0, armed with decompiled output and Claude Code for symbol interpretation.¹ The findings weren't incremental clarifications of vague documentation. They were things nobody outside Qualcomm appears to have known at all.

The most striking: Qualcomm's compiler solves tensor placement as a Mixed Integer Linear Program and hands it to HiGHS, a well-known open-source optimization solver. Not heuristics. Not a greedy allocator. A formal MILP, dumped to .mps format for debugging. The objective function minimizes total bytes moved — spills to DDR, fills back from DDR, and inter-core transfers on multi-core chips. Two live tensors can't share the same on-chip bytes. Everything else follows from those constraints, and when the constraints can't be satisfied cleanly, HiGHS finds the least-bad answer. That a proprietary NPU compiler was outsourcing its core placement problem to an open-source academic solver was, per the writeup, "completely unknown publicly."¹

The memory architecture the compiler is navigating is genuinely constrained. The Hexagon chip has a small pool of on-chip scratch memory — VTCM, vector tightly coupled memory — and anything that doesn't fit there has to be shuttled to DDR, which is slow and energy-expensive. The difference between fitting and not fitting is dramatic: running Qwen 0.8B on an SM8350 produced 5.46 MB of spills and 33.9 MB of fills, for 37.9 MB of total DDR traffic. The same model on an SM8650 produced 1.15 MB. A 33x difference in DDR read traffic from choosing a different target chip. And yet both chips reported the same VTCM size in the compiler output — "a field that just says 4. Now I don't know 4 'what', or whether it's some code."¹

The compiler works in stages. First a scheduler finds the operation ordering that minimizes peak working set size, using Priority BFS with a depth-first topological sort underneath and a stack of tiebreaker heuristics. Then the placer takes the fixed ordering and assigns tensors to VTCM addresses — not 1D addresses, but coordinates in a three-dimensional tile space, with a recursive backtracking allocator that retries when placement fails. The fallback path is that backtracker. When the compiler's full optimizer runs instead, that's when the MILP takes over.

One finding is less "surprising" than it is quietly alarming for anyone doing edge deployment: the compiler inserts relaxed_precision_cast operations during placement, converting tensors between float32, FP16, and BF16 to relieve memory pressure — without informing the developer. Your float32 weights can silently become FP16 if that's what makes the bytes fit. The engineer confirmed these casts are inserted during placement; whether they're variables inside the solver itself or a separate pass, the binary didn't make clear.¹

Underneath the scheduler and placer sits a hidden analytical simulator called Hextimate — Hexagon estimate — that prices every tradeoff before the chip exists. The compiler runs on a Linux box; the phone hardware isn't there. Hextimate models the entire chip analytically, running each workload twice: once assuming perfect overlap, once assuming none, and returning a range. The memory cost formula, extracted directly from machine code, is the textbook roofline model: bandwidth = channels × width × efficiency × frequency, time = bytes / bandwidth. It also carries dedicated recognizers for FlashAttention, MoEs, KV cache, and rotary embeddings — the cost model is explicitly biased toward LLM workloads at the coefficient level.¹

The writeup is careful about its own limits. One SDK version, one binary, one build ID. The engineer's prior RE experience was "easy/medium crackmes, not proprietary compilers."¹ Every claim about the MILP and Hextimate is inferred from decompiled code and unmangled symbol names that survived stripping — not from source. Some SIMD-heavy sections were left unread. The legality of the approach is still being sorted out before the methodology gets published.

But the findings that matter for practitioners are concrete and reproducible with Qualcomm's own tools as oracle. The spillFillBufferSize metadata field in the compiled binary, when zero, means everything fits on-chip — a static oracle anyone can query.¹ The MILP's objective is pure byte traffic, which means a custom scheduler that mirrors that objective will predict compiler behavior. Precision is a knob the solver pulls without asking, so inference quality on real hardware may differ from what you measured. These aren't implementation curiosities. They're the kind of things you need to know before shipping a model to a phone.

Full writeup: datavorous.github.io/writing/qairt/

↑ Back to top

FROM THE ARCHIVE

The Line to Moscow: June 20, 1963

Geneva conference room, 1963. Two figures in dark suits bent over a sparse table, pen touching paper on a three-page document. Harsh overhead light carves deep shadows across their faces and jacket lapels. In the foreground, a teletype machine dominates — its keys, paper spool, and mechanical housing rendered with dense cross-hatching. A single wire runs off the edge of the frame toward an implied horizon. Generous white space above; the room's tall windows suggest distance and cold clarity. Black-and-white pen-and-ink with bold architectural lines and controlled hatching. No text, no flags, no color.

Eight months after the Cuban Missile Crisis nearly ended the world, American and Soviet diplomats signed an agreement establishing a direct, 24-hour communications link between Washington and Moscow. The date was June 20, 1963. They called it the "hot line" — and it was, at the time, the most consequential piece of telecommunications infrastructure in the world.

The Cuban Missile Crisis had exposed something terrifying: the two nuclear superpowers had no reliable way to talk to each other in real time. During the critical days of October 1962, as Kennedy's naval quarantine surrounded Cuba and Khrushchev decided whether to back down, messages between Washington and Moscow took hours to arrive. The crisis was resolved not because the communications system worked but almost in spite of it. Kennedy declared: "This age of fast-moving events requires quick, dependable communication in time of emergency."¹

The agreement signed on this date was built on that lesson. It was framed explicitly as "a first step to help reduce the risk of war occurring by accident or miscalculation" — an acknowledgment that accidental catastrophe was at least as dangerous as intentional conflict.¹ The system was operational within months.¹ It appeared in Fail Safe as dramatic infrastructure and in Dr. Strangelove as dark comedy.¹ In practice — thankfully — it never had to prevent a nuclear war. What it represented was something new: the architecture of de-escalation, the idea that the line itself, simply being open, changes the calculation.

↑ Back to top

THE FUNNIES

The Black Box Speaks / Seismic Activity

*After XKCD — on the solo engineer who cracked open a proprietary chip compiler with Ghidra and found a real open-source math solver doing the heavy lifting inside. After Krazy Kat — on the fans at Seattle Stadium whose goal-celebration roars registered as genuine seismic events on the Pacific Northwest Seismic Network's instruments.*

↑ Back to top

ALSO NOTED

Also Noted

• SMPTE Opens Its Full Standards Library — The Society of Motion Picture and Television Engineers made its entire 110-year catalog of standards, recommended practices, and engineering guidelines freely available to the public this week, removing a long-standing access barrier for developers, integrators, and educators building on specifications for digital cinema, IP workflows, and content provenance. smpte.org Jun 17, 2026
• Waymo's Fourth Recall in 28 Months — Waymo filed a recall covering its entire US fleet of 3,871 vehicles after robotaxis in Phoenix and the Bay Area drove into closed freeway construction zones; the company has restricted all freeway operations while a software fix remains under development. wired.com Jun 18, 2026
• XLibre XServer 25.2 Due on the Summer Solstice — The X11Libre fork has tagged version 25.2.0 for release on June 21 with 976 commits since 25.1, including Nvidia support for DRI3/glamor (v470+), a major kdrive overhaul adding multi-screen and VT switching, and fixes for long-standing crashes on Radeon, amdgpu, and suspend/resume with Nvidia. github.com Jun 19, 2026
• Project Valhalla Finally Lands in JDK 28 — After twelve years and five prototype iterations, JEP 401 (Value Classes and Objects) has been integrated into OpenJDK mainline targeting JDK 28 preview — adding 197,000 lines across 1,816 files — enabling flat, allocation-free value types that "code like a class, work like an int" for high-performance Java without giving up abstraction. jvm-weekly.com Jun 18, 2026
• Mixtape Is the Highest-Rated Game of 2026 — Annapurna Interactive's narrative teenage adventure scored a 94 on OpenCritic at launch in May, the first game this year to match Hades 2's 2024 peak score and comfortably clear the 90-rated Pokémon Pokopia for the year's top spot. opencritic.com May 8, 2026
• Ultrasonic Espresso Uses 75% Less Energy — A team at the University of New South Wales demonstrated a room-temperature brewing process that uses high-frequency sound waves and acoustic cavitation to extract espresso-equivalent coffee in three minutes, with sensory tests showing no preference between it and conventional espresso. wired.com Jun 20, 2026
• MCP's Real Value Might Just Be Auth Isolation — Simon Willison flagged a Hacker News comment from Sean Lynch arguing that MCP's practical advantage over plain CLI tools is moving auth flows outside the agent's context window entirely — and that an idealized MCP could be little more than an auth gateway for the API. simonwillison.net Jun 19, 2026
• QUOD: A Full 3D Shooter in 64 KB — Solo developer Daivuk shipped a Quake-style boomer shooter with three levels, a boss fight, four enemy types, four weapons, and all assets — textures, models, animations, sounds, music — packed into a 64 KB Windows executable using demoscene compression techniques and TrenchBroom-authored levels. lucasgraphic.com Feb 2026

↑ Back to top

THE QUESTION

When Does Process Documentation Replace the Understanding It Was Supposed to Produce?

Three years is long enough to know if something is working. SafeR — the UCI's multi-stakeholder road safety project, established in 2023 — operates a quarterly supervisory board, a monthly technical commission, and a weekly case management committee that reviews incidents and audits upcoming race routes.¹ As THE PELOTON reports today, the UCI has not published data demonstrating a reduction in crashes or accidents across three years of that operation.¹ The CPA Women responded to Urška Žigart's fractured jaw at Tour de Suisse by calling for a "substantial review," not because the structure obviously failed but because the structure hasn't proven it worked.¹

The same structural failure turns up in a different domain today. THE LAB covers Lorin Hochstein's essay on LLM-written incident reports, which argues that a plausible-sounding operational post-mortem — one that names causes, traces couplings, describes corrective actions — can be filed, read, nodded at, and archived without anyone having done the synthesis work that would surface a wrong answer. The correctness signal for a software incident report isn't "did the incident recur?" — that might take months, or it might not recur for unrelated reasons. The report itself becomes the artifact of record. Hochstein draws on Dick Guindon's observation that writing is how you discover you didn't understand something.² Outsourcing the writing doesn't skip the report; it skips the discovery.

Both cases share the same architecture: an output that looks like understanding gets produced at regular intervals, and the underlying question — did this actually help? — never gets its correctness signal published. SafeR's committees meet. Incident reports get filed. The process is running. Whether it is working is a different question, and it's the one neither organization has forced itself to answer in public.

The hot line, which this paper's archive section traces to June 20, 1963, was built on exactly the opposite premise: that the value of a communication channel was not the volume of traffic through it but whether the channel changed the calculation in a crisis. That premise demanded a concrete test — would the line be there when needed? — not a schedule of meetings. The lesson from the Cuban Missile Crisis wasn't "establish a committee"; it was "build the thing that makes the catastrophic outcome less likely, and verify it's ready."

The question worth carrying through the day: at what point does the process documentation become the institution's evidence that it cares, rather than evidence that the problem is being solved?

↑ Back to top