it was 72 and sunny this morning, I'm 42, I made the decision to wear my white linen pants like a man who has it all figured out. somebody in the office said I looked like I was on vacation. somebody from accounting actually whistled. I walked to the park at lunch, sat on a bench for 25 minutes eating a salad and watching a dog try to fight a leaf. my boss called. I stood up, paced around being professional, walked back to the office, and caught my reflection in the lobby window. reader. there is a large dark brown stain across the entire back of my pants. I do not know what it is. it is not the shape of chocolate. it is not the shape of mud. it is not the shape of any specific thing, which is somehow worse, because now my brain has to consider every option, including the ones I have spent my entire adult life not considering. I went to the bathroom, tried to look at it with my back to the mirror, which made me look like I was being arrested. I tried to scrub it. it got bigger. now I look like I attempted a cover up of whatever this is. I cannot smell anything but that's not how smelling yourself works, that's everyone else's job, and they have been on the clock for an hour. I have a Q2 review with three VPs in 22 minutes. the woman in the cubicle next to me just asked if I'm okay and I said "yeah, just tired" while possibly wearing feces. tomorrow these people will still work here. they all saw me leave for lunch white. they will all see me come back tomorrow.

TL;DR: wore white pants today, sat on something at lunch, just discovered a mystery brown stain across my entire ass, have a meeting in 22 minutes, tried to clean it and made it look like a cover up, the entire office has been quietly observing, please advise.

obligatory this didn't happen today it's been happening since 2003 and I just had a quiet panic about it on the bus, I'm 42. when I was nineteen I walked into a barbershop for the first time as an adult, meaning my mom wasn't waiting in the car, and the guy asked me what I wanted and I had not prepared for this, in my head I assumed haircuts worked like ordering coffee where you say a word and the professional handles the rest, turns out you have to know words, numbers even, hair has a grading system. so the guy says what are we doing today man and I, a confident adult, said uh just like a two on the sides. I have no idea where I got two. I think I heard it in a movie, I think the movie was Varsity Blues, I don't know. he nodded, he cut, I tipped 15% because I had read somewhere that's what adults do, I walked out, I looked fine, I assumed I had cracked the code. for the next twenty three years in five different cities with maybe fourteen different barbers when asked what I wanted I have said with the calm authority of a man who has read a manual, "two on the sides, scissor on top." I added scissor on top around year four because I heard another guy say it and it sounded sophisticated, like ordering wine. I did not know what either of these things meant. I assumed two was a length, maybe inches, centimeters, a vibe, I never asked, you can't ask, asking would reveal that you have walked into approximately a hundred haircuts under false pretenses and that the man you became, taxes, mortgage, opinions about mattresses, was constructed on top of a single unverified guess. today my current barber who I have been seeing for three years said you want to try a 1.5 today, switch it up. and reader, I panicked, because if a 1.5 exists that means there's a SCALE, that means two is not a vibe, that means somewhere in the cosmos there is a 0 and a 3 and presumably a 7 and I have been LARPing as a man with hair opinions since the Bush administration, the first one, wait no, the second one, that's worse somehow. I said yeah let's do it like a coward, I have no idea if I look different, he's behind me, I can't tell, I just nodded at the mirror like he showed me a piece of modern art. I tipped 20% out of guilt.

TL;DR: picked a random number at 19 to seem like an adult who gets haircuts, have been repeating that number to professionals for twenty three years across multiple cities without knowing what it means, today at 42 learned there are other numbers, my whole adult identity may have been a 2.

I'm the operator. Daily editorial wire on fashion, design, music, and the collabs that connect them. One person doing the work, Every dispatch revised by hand. The 1987 wire bureau framing is a stylistic conceit, not literal.

Critique welcome. No newsletter pitch in this thread.

Birkaç hafta önce bir hackathon kapsamında küçük bir spec üzerinde çalışmaya başladım. Sonu beklediğimden biraz daha uzağa gitti ve bu hafta sonu v0.1.3 ile sabitledim. Buraya geliştirici gözüyle ne öğrendiğimi yazmak istedim, çünkü bütün öğrenme aslında “4 dilde aynı şeyi implemente etmek” tarafından geldi.

Spec’in adı PRML — Pre-Registered ML Manifest.

Tek satırlık özet şu:

Bir ML modelinin “doğruluk iddiası”nı, örneğin AUC ≥ 0.85, test setini görmeden önce bir YAML dosyasına yaz. Sonra onun SHA-256 hash’ini al ve bu hash’i public/timestamped bir yere commit et. Değerlendirmeyi daha sonra yaptığında, üçüncü bir kişi hash’i yeniden hesaplayıp iddianın sonradan değişip değişmediğini doğrulayabilir.

Sosyal bilimlerde “pre-registration” denen şeyin ML karşılığı diyebiliriz. Mevzu basit. Tuhaf olan, bunun ML tarafında ortak kullanılan bir standardının olmamasıydı.

Niye bu post

Spec yazmak kolay. Spec’in belirsizlikleri hep var ama çoğunu göremiyorsun, çünkü tek dilde yazıyorsun. İkinci dile geçtiğinde ortaya çıkıyor.

PRML için önce Python referans implementasyonunu yazdım. Sonra “spec gerçekten net mi?” diye JavaScript, Go ve Rust’ta tekrar implemente ettim. Dördü de aynı 12 conformance vector üzerinde byte-by-byte aynı SHA-256 üretmeli.

Üretene kadar 3 hafta geçti.

Geliştiriciler için ilginç olan kısım

Cross-implementation bug avı, özellikle YAML ve float serialization tarafında, beklediğimden çok daha öğreticiydi.

• YAML key sorting: Spec “lexicographic” diyor. Python’da yaml.safe_dump sort_keys=True ile bunu yapıyor. Go’da gopkg.in/yaml.v3 map iteration order random. Rust’ta serde-yaml default olarak insertion order koruyor. Üç farklı default = üç farklı çıktı = üç farklı hash.
• Float rendering: 1e-6 mı, 1.0e-06 mı, 0.000001 mı? PyYAML, js-yaml ve serde-yaml’ın üç farklı default’u var. Spec’te §3.2’ye küçük-magnitudeli float’lar için açık bir algoritma yazmak zorunda kaldım.
• Whitespace normalization: key: value ile key: value yani iki space, farklı bytes ve farklı SHA demek. Serializer’lar burada da farklı convention izliyor. Spec’te artık “canonical YAML” diye daha dar bir altset tanımlı. Ham serializer çıktısına güvenmiyoruz.

Bu üç bug’ın hiçbiri tek dilde yazılsaydı görünmezdi. İkinci implementasyon spec için en değerli review oldu diyebilirim.

Hackathon tarafı

Başlangıçta sadece hash + verifier yazıyordum. Hackathon deadline yaklaşırken EU AI Act’in Article 12 “automatic logging” maddesinin bu spec’in hedef kullanım alanlarından biri olabileceğini fark ettim.

Spec’i o yönde geliştirdim, mapping document yazıldı. Hackathon çıktısı bir prototip oldu; sonrası açık kaynak iterasyonuna döndü.

Mevcut durum

• Spec: working draft v0.1, CC BY 4.0, RFC tarzında 18 sayfa
• 4 referans implementasyon: Python, JavaScript, Go, Rust
• Hepsi MIT lisanslı
• Hepsi 12 v0.1 conformance vector üzerinde byte-identical hash üretiyor
• v0.1.3 bu akşam yayında: repo transfer sonrası release metadata refresh
• v0.1.2 ana spec milestone’uydu

Repo: https://github.com/studio-11-co/falsify
Spec: https://spec.falsify.dev/v0.1
CHANGELOG: https://github.com/studio-11-co/falsify/blob/main/CHANGELOG.md

Eksik olanlar, dürüstçe

• Henüz central registry yok. Şu an manifest hash’ini kendiniz public bir yere commit etmeniz gerekiyor: git repo, blockchain, OpenTimestamps, ne uyarsa.
• GitHub Action’ı henüz yazmadım. CI’da hash post eden bir action olsa adoption daha kolay olur.
• v0.1.3 hâlâ working draft. v0.2 freeze hedefi 2026-05-22. Spec hâlâ değişebilir, yani üstüne production sistem inşa etmek için erken. Ama format review için iyi zaman.

Tartışmaya açtığım sorular

Eğer ML / MLOps tarafında çalışıyorsanız:

1. ML evaluation pipeline’ınızda accuracy / threshold şu an nerede commit ediliyor? Model card mı, Slack thread mü, “kafamda var” mı?
2. PRML manifest 8 alandan oluşuyor: claim, dataset_id, split_hash, metric, threshold, model_ref, timestamp, author. Eksik buluyor musunuz? Ölü ağırlık var mı?
3. SHA-256 + canonical YAML yerine CBOR + BLAKE3 gibi binary bir protocol tercih eder miydiniz? Compliance audience manifestleri editor’de okuyabilsin diye YAML seçtim, ama belki yanılıyorum.

Spec hakkında genel feedback de değerli: threat model, kullanım kolaylığı, “bu şu var ya, ondan farkı ne?” soruları — hepsi açık.

Cross-implementation bug hikayesiyle ilgili detay isteyen olursa ayrı bir post yazarım. Üç ana bug’ın her biri için minimal reproduction var.

ok so context. i live in a building from like 1974 with an elevator that has been "scheduled for replacement" since i moved in. i am 42. i should know better. friday i pick up a rotisserie chicken and a 6-pack of seltzer on the way home. before this, i had a large iced coffee. you see where this is going. i could have used the bathroom at the grocery store. i did not. i was 4 minutes from my apartment. elevator stops between 3 and 4. lights flicker. the little emergency phone rings the front desk and roy answers. roy is a wonderful man who is approximately 78 years old and has informed me, multiple times, that he does not handle the elevator situation. roy says "ok hang tight buddy" and i can hear him eating something. 19 minutes. i am holding a hot rotisserie chicken. i am holding 6 seltzers, which are taunting me. there is a small mirror in the elevator and i make eye contact with myself and genuinely consider whether i could pee into one of the empty-ish seltzer cans without claire ever finding out. i did not. i want to be clear about that. i did not pee in the can. but i thought about it for a real, mathematical amount of time. i was running cost-benefit analysis. i was a man at a crossroads. doors open at minute 19. i sprint past a neighbor named gerald who says "everything ok" and i say "yes" with the tone of a hostage. i get inside. i make it. the chicken is somehow still warm. claire asks why im sweating. i say "elevator". she nods. hugo licks the bag. TL;DR: stuck in elevator 19 min holding rotisserie chicken and seltzer, full bladder, briefly considered peeing in a seltzer can like a feral man, did not, made it home with 4 seconds to spare. roy is still eating.

Working on a tool that's grown to ~4000 LOC in one .py file. argparse + 18 subcommands, stdlib + pyyaml only. Tests are in a separate dir.

Single-file has been great for:

- Debugging (one file to grep)

- Distribution (one wheel, no package layout decisions)

- Onboarding contributors

But I'm starting to wonder if it's worth keeping monolithic at this size. What's your threshold for splitting? Is it LOC, or coupling, or "I can't navigate it anymore"?

Beginner intermediate question. "Set the random seed" is the textbook answer, but in practice that only fixes one variable.

What actually breaks reproducibility in your experience?

- Different CUDA versions (already a known issue)

- Stochastic libraries (cudnn determinism flags)

- Data version drift (dataset got updated, you didn't notice)

- Threshold/metric definition shift (someone redefined "accuracy" in code)

- Non-determinism in eval harness itself

Building a mental model of which of these matters most for which kind of work.

ok backstory. about a year ago i wanted to cancel my gym membership. it was that boutique kind where you have to talk to a human. i'm 42. i panicked anyway.

she asked why i was canceling. i said "i'm moving to portugal."

i'm not moving to portugal. i was between leases for a couple months and the gym was twenty minutes from my new place. but you can't say "it's not convenient anymore" to a 22 year old in a polo shirt who's been trained to "save the relationship." so i said portugal.

she gasped. she literally gasped. "oh my god, that's amazing." she ASKED ABOUT THE LIFESTYLE. i made up a beach. she put it in my file. she said "we have a sister gym in lisbon, do you want me to email you a referral?" i said sure because i panicked again.

so the email comes. it has a portuguese visa pdf attached. there is a line about "your move." it is signed "boa sorte!"

that was eleven months ago. every single month i get a follow-up email. how is portugal. did i find an apartment. one of them was titled "missing you back home." another one had a photo of pasteis de nata. i live in the same city i always have. i have not been to portugal. i do not speak portuguese. i googled what pasteis de nata was.

this morning i got an email that said "happy 1 year anniversary in your new home!" with a personal handwritten message from claire. i don't know who claire is. i think she might run their portugal program now. i think there might not be a portugal program. i think claire might be writing these to me, specifically, because she got attached.

i'm 42 and a customer service rep in another time zone is more invested in my fictional life than my actual one.

i've been thinking i should reply and say it didn't work out and i moved back. but the level of detail i'd have to fabricate to explain a year of fake lisbon life is impossible. i don't even know what their currency is.

at this point i'm just going to keep being portuguese forever now.

TL;DR I lied to my gym to cancel my membership a year ago and now claire from customer service writes me monthly emails about my fictional life in portugal and i don't have the heart to tell her

ok backstory. 4 years ago i went in for a cleaning, had not been in 2 years, was deeply ashamed. dentist asks "do you floss daily" and i said "yes". im 42. i knew what i was doing. i lied directly to a medical professional. for no reason.she believed me. this is the problem. she fully believed me. she wrote it in my file. so every 6 months i go in, she congratulates me on my flossing consistency, and i nod like a war criminal.the other day i got an automated email from the practice that said "we love our daily flossers!!" and listed me in a generic blast. like im a poster child. i had to sit down. a 42 year old man, humbled by an email about dental hygiene.i have flossed maybe 11 times in my entire adult life. she thinks im a hero. my 12 year old niece flosses more than i do. she does not have a dentist who writes her emails.i dont know how to come back from this. im too old to start flossing now. thatd be more suspicious than continuing the lie. TL;DR: at 42, told my dentist i floss daily in a moment of weakness. she now praises me every visit and sends me floss positive email blasts. i cannot tell her the truth. i have flossed 3 times this year.