Index

Briefing

Learn

Tests

Case

Record

Sources

Briefing
Stage Learn · 4/5

Learn

Five-layer cards

65 cards, four decks. Question → plain → analogy → example → trap → Ox.

You will understand
Question first, layers after.
Mark what you already have.
Four decks, one method.

1 / 65 · 0 got it

1 · Absolute zero

What is an AI chat, with no magic?

1. What is an AI chat, with no magic?

Plain. You type. A program guesses the next scrap of text, then the next, until it stops. No goblin. Trained statistics.

Analogy. Keyboard autocomplete that dared to write a whole paragraph.

Example. You: “The sky is”. It often continues “blue”, because that pair happened a lot in its data and context pushes it.

Trap. It does not “understand” like you. It packed patterns. Confident lies are still a brilliant parrot.

Ox. Ox Alpha is that loop, dressed as “reasoning”: a private draft first, then the answer. OpenRouter calls it a coding/agent model.

Absolute zero

2. Is it Google? Does it search every sentence?

Plain. No. An LLM does not search the web unless someone wired a search tool. It recites (and remixes) what it saw in training, plus what you paste now.

Analogy. A student who memorized the library and now improvises. They do not walk to the archive every time.

Example. Ask “who won yesterday’s game” with no tool and it will invent or say it does not know. It does not open ESPN.

Trap. “It has internet” ≠ the model. Sometimes the CHAT has a plugin. Ox Alpha, on the sheet, is the brain, not the search engine.

Ox. The OpenRouter card does not promise web search. It promises text, image and video in. Do not treat it as a news oracle.

Absolute zero

3. What is a token? A word?

Plain. A token is a Lego brick of text: sometimes a word, sometimes a chunk (“ing”), sometimes an emoji or a sign. The model thinks in bricks, not letters.

Analogy. Old SMS counted 160 characters. WhatsApp does not. Models count bricks. “Hi” may be 1; “extraordinarily” 3 or 4.

Example. “Hello” ≈ 1 token. “🚀🔥” can be 2. A 40-character Python loop can be 15–25 tokens depending on the knife.

Trap. 1 million tokens is not 1 million words. In English it is often ~750k words. In Chinese, often fewer pieces per character.

Ox. Ox Alpha takes 1,048,576 context tokens and 131,072 output. That is a fat book + a repo, not a tweet.

Bricks and knobs

4. What is the tokenizer and why is it a fingerprint?

Plain. The tokenizer is the recipe that cuts text into numeric IDs. Each lab has its own (or copies one). Two recipes do not yield the same count.

Analogy. Two cooks, different knives. Same cucumber: 8 slices vs 14. If it is always 8, it is the same knife.

Example. Same sentence to GLM-5.3 and Ox: Ox reports exactly 75 more tokens, in English, German, Chinese, code and emoji. 75 is the letterhead, not the cucumber.

Trap. Same tokenizer ≠ same weights. Someone could fine-tune GLM. But rivals do not copy dictionaries by accident; Kimi/Qwen/MiMo diverge.

Ox. This is THE star test. A graduate can repeat it: same sentence, subtract prompt_tokens, see if Δ is constant.

Bricks and knobs

5. What is a parameter? Is it a token?

Plain. A parameter is a numeric knob the model turned while studying. There are hundreds of billions. It is not a scrap of your sentence.

Analogy. Tokens = the Lego you hand over. Parameters = the builder’s hands. Do not mix them.

Example. GLM-5: ~744 billion knobs in the building, ~40 billion awake per word. Your 20-token prompt is not “20 parameters”.

Trap. “Fewer parameters” is not an insult in 2026. An efficient MoE fights models 2–3× larger. The bill charges AWAKE knobs.

Ox. Ox does not publish a count. If it is GLM-5.x, it inherits ~744B/40B. The Flash ~300B rumor is not an official sheet.

Bricks and knobs

6. What is MoE with no jargon?

Plain. Mixture of Experts: a building of sub-brains. A doorman wakes only a few for each scrap of text. The rest wait.

Analogy. A hospital of 744 doctors. For a sprain you wake 40. The hospital KNOWS a lot; the shift is cheap.

Example. Cost ≈ 2 × active_params × tokens. 40B vs dense 2T → ~50× cheaper per generated token. That is how someone can gift a weekend.

Trap. “Experts” are NOT “the psychology one” and “the biology one”. They are same-shaped nets; the doorman learns who fits each pattern.

Ox. The Flash/Air thesis: same GLM building, fewer people awake, a bill that can claim 100T capacity.

Bricks and knobs

7. What is the context window?

Plain. How many tokens fit at once: system + your message + history + what it generates. A work desk, not infinite memory.

Analogy. A 1 km desk. Every notebook fits. You can still lose math at meter 400 (“lost in the middle”).

Example. 2²⁰ = 1,048,576. Not a round million. GPUs love powers of two. 131,072 output = 2¹⁷.

Trap. Capacity ≠ quality. A published million does not prove it uses token 800,000 well. That is a separate exam.

Ox. Ox and GLM-5.2/5.3 publish exactly 1,048,576. Xiaomi does too: not a passport. It does align engineering.

Bricks and knobs

8. Why +75 and not +71 or +90?

Plain. OpenRouter (or the lab) pastes a hidden system prompt: “you are ox-alpha, do not identify…”. That is ~75 fixed tokens. User text varies; the letterhead does not.

Analogy. Two suitcases. One always weighs 75 g more, feathers or bricks. Those 75 g are the tag, not the contents.

Example. H0: Δ(P) = C for every prompt P. H1: Var(Δ) ≫ 0. Observed: C = 75 on 6–25 prompts. MiMo/Kimi/Qwen: Δ changes with language.

Trap. If the knife were different, you would sometimes see 40 and sometimes 120. A constant is almost impossible to fake without the same vocab.

Ox. Pliny leaked the wrapper: “You are ox-alpha… undisclosed organization”. That fits a ~75-token letterhead.

How you identify it

9. What is an API, for someone who never coded?

Plain. A counter with rules. You send an order (JSON). Tokens come back. You do not talk to the brain: you talk to the window.

Analogy. You ordered delivery with no logo. The box, the receipt and the complaint number betray the restaurant.

Example. Error 1210: “This model always engages in thinking and cannot be disabled; please use low, high, or max”. Same sign as GLM-5.3.

Trap. OpenRouter is the mall, not the storeroom. stealth/ox-alpha is the aisle. The storeroom hides: that is why 1210 and Chinese logs matter.

Ox. Pritish Yuvraj (ex-Llama/Meta) saw 1210/1214 identical to chat.z.ai. Josh rad saw CJK logs + ModelScope. That is z.ai’s kitchen, not Cursor’s.

Absolute zero

10. What is a stealth model?

Plain. An anonymous test on a public highway. The lab watches whether traffic holds and what people say. Then the mask comes off… or not.

Analogy. A restaurant called “Nameless Place”, free week. The previous four on that street, when the sign went up, were different Chinese kitchens.

Example. Pony Alpha → GLM-5. Hunter/Healer → MiMo-V2. Elephant → Ling-2.6-flash. Owl → LongCat-2.0. Ox is the fifth. Still no sign.

Trap. Stealth is not “military secret model”. It is preview marketing. OpenRouter says: we are not the owner.

Ox. The 4/4 Chinese channel is track statistics, not silicon racism. Betting on a Chinese lab is the prior; fingerprints pick which one.

Absolute zero

11. What is a model fingerprint?

Plain. You cannot open the skull. You measure habits it cannot stop being: knife, error yells, video toll, temperature-0 style.

Analogy. You never see the face. You see shoe size, accent, coffee brand and the metro line. Five matches are not a whim.

Example. Ox: GLM tokenizer, encoder = GLM-5V-Turbo token-for-token, error 1210, rejects audio, CJK logs. Six habits. One (playbook) could also be Xiaomi.

Trap. One habit is not a passport. A jailbreak that says “I am GLM” is the easiest to fake. The video encoder is the most expensive to fake.

Ox. This dossier stacks 8 tests. The judge does not sign identity. It signs “if Z.ai claims it, nobody is shocked”.

How you identify it

12. What is temperature 0?

Plain. Always pick the most likely token. Zero improvisation. Useful for comparing clones: if two systems write nearly character-for-character the same, they are family or a copy.

Analogy. A disciplined parrot vs a creative one. At 0 it recites the anthem. At 1 it invents verses.

Example. Testers at temp 0 saw the same markdown quirks and the same German-decimal LaTeX (0{,}375) as GLM-5.3.

Trap. Temp 0 does not prove identical weights. A very faithful distill also copies style. It adds; it does not close the case.

Ox. Together with +75, temp-0 style is why “another lab with a foreign tokenizer” looks very bad.

How you identify it

13. What is prefix cache?

Plain. If you paste the same repo 40 times, a smart server does not reread it whole. It stores compute from the start. High hit rate = a serious kitchen.

Analogy. A barista with your cup marked. Most coffees come from the thermos, not the grinder.

Example. One pi.dev tester measured ~97%. The OpenRouter catalog showed ~87% Ox vs ~92% GLM-5.3. Not the same figure: do not sell it as a passport.

Trap. Cache is infrastructure, not brain. A recipe can be cloned. That is why we lowered its weight versus tokenizer and video.

Ox. Still compatible with z.ai. No longer the test that “only z.ai beat 90 on the catalog”.

How you identify it

14. How does a model “see” a video?

Plain. It does not enter like Netflix. An encoder splits frames and turns them into expensive tokens. Each lab chooses how many tokens a second costs.

Analogy. Translating a movie into a comic. If two studios use the same panels per minute, they copied the storyboard.

Example. 2s 30fps 360p clip: Ox = 296 tokens, GLM-5V-Turbo = 296, MiMo v2.5 = 910, Qwen = 408. Four clips, token-for-token match with 5V-Turbo.

Trap. This does NOT prove Ox is the text 5.3. Public 5.3 does not see video. It proves the video cashier is GLM-5V’s.

Ox. The most expensive fingerprint to fake. A Cursor fine-tune that still charges 296/884/1064 is using Z.ai’s preprocessor.

How you identify it

15. Why does rejecting audio matter?

Plain. The sheet says text+image+video. Testers confirm: no audio endpoint. Inkling (Mira) DOES take audio. MiMo-Omni too.

Analogy. It does not eat peanuts. The cousin in the showroom does. It is not that cousin.

Example. Inkling July 2026: text+image+audio, open-weights, 975B/41B. Ox August: video yes, audio no, no public weights.

Trap. Rejecting audio does not STAMP Z.ai. It eliminates Inkling and MiMo-Omni. The other fingerprints pick Z.ai.

Ox. In Spanish “mira” means “look”. Inkling is another model, with a passport. Do not mix the names.

How you identify it

16. Is 80% on 10 tasks a world title?

Plain. No. DeepSWE has 113 tasks. 8/10 is a quiz. The bench author (Wenqi) measured ~63% on a larger subset. That looks like GLM-5.3, not a god.

Analogy. Hit 10 free throws and claim you are better than the NBA. Maybe. Maybe you picked the easy ones.

Example. Davis n=10: Ox 80%, Fable 65%, Sol 52%. Wenqi: ~63% at 47K output tokens. Kingbench: Ox 87.5 vs GLM-5.3 91.25.

Trap. Viral tables that crown Ox over Fable/Sol as if it were an official leaderboard are wrong. This dossier labels them n=10.

Ox. We use benches as a family clue (“it looks like GLM”), not as a gold medal.

The Ox Alpha case

17. Is Manifold’s 85% a paper?

Plain. Manifold is a prediction market. People put money (or points) on “who is behind it”. 85% Z.ai is a survey with skin, not an experiment.

Analogy. Bookies do not write physics papers. Sometimes they are right because they aggregate informed gossip.

Example. 21 Aug snapshot: Z.ai ~85%, Xiaomi ~6%, Other ~6%, xAI/Cursor ~2%. Our 85/22/9 scores are this dossier’s rubric, not Manifold.

Trap. Do not cite 85% as if Z.ai signed it. Cite it as “the market believes”.

Ox. The file uses 85 as a live market verdict + our own rubric. We label it a hypothesis.

The Ox Alpha case

18. What can I claim today without lying?

Plain. FACT: OpenRouter is not the owner. FACT: 1,048,576 / 131,072 / text-image-video / $0. STRONG HYPOTHESIS: unpublished multimodal GLM-5.x family. NOT A PASSPORT.

Analogy. The detective has a print, a shoe, a voice and a postcode. No ID photo. A prosecutor would not ask for life yet.

Example. “Ox Alpha is GLM-5.3” is FALSE as stated: public 5.3 is text-only and 128K out. “Ox is a GLM-5.x SKU with a camera” is the correct sentence.

Trap. Pliny (“my agent has spoken”) and Dax (“nah nah poo poo”) are not lab confirmation.

Ox. If Z.ai claims it tomorrow, nobody is shocked. If Murati claims it, we rewrite the file.

The Ox Alpha case

19. Are my prompts private?

Plain. Three contracts collide. OpenCode: Zero Data Retention. OpenRouter card: the provider RETAINS prompts and does not train on them. Stealth EULA: free access is consideration and MAY be used to train.

Analogy. Three signs in the same shop. “We do not record”. “We record but do not sell”. “Free entry means your photos are ours”. Pick the harshest.

Example. Do not paste keys, medical histories or your company’s secret code. A stealth model is not your notary.

Trap. A product tweet saying “zero retention” does not beat an EULA. Read the strictest one.

Ox. This dossier puts it on the front page on purpose. Pedagogy also means telling you when to shut up.

The Ox Alpha case

20. Are 100 trillion tokens a day real?

Plain. It is an OpenCode CAPACITY claim, not a meter. In a day-plus OpenRouter listed ~1.7–1.93T served: huge, two orders below 100T.

Analogy. A factory says “we can make 100 million screws a day”. They sold 2 million. Both sentences can be true.

Example. Z.ai stood up ~1 GW on Chinese chips (Bloomberg / Tom's, Jul 2026). That makes gifting tokens PLAUSIBLE. It does not identify the model.

Trap. One post said “1 quadrillion” and “Nous”. OpenCode said 100T. Do not promote the quadrillion.

Ox. Treat 100T as fleet-stress marketing. Treat ~2T as what was observed.

The Ox Alpha case

21. What if it is Cursor’s Composer 3?

Plain. 21 Aug theory (@TimMacc): timing + GLM fine-tune + compute. Cursor already post-trained other labs’ bases. Manifold xAI/Cursor ~2%.

Analogy. “The tarped car is a Tesla with a Toyota engine”. Maybe. The toll, the dashboard error and the radio are still Toyota.

Example. Composer 3 was announced as 1.5T from-scratch on Colossus. A from-scratch would not share a GLM tokenizer or ModelScope logs.

Trap. The only Cursor version that survives is “GLM post-train still served on the z.ai stack”. Possible and stretched. Score ~9.

Ox. We leave it alive on Suspects. It does not lead. The 296=296 encoder wounds it.

The Ox Alpha case

22. How is an agent different from a chat?

Plain. A chat answers. An agent stays all afternoon touching your repo: reads the error, changes the file, runs tests, reads again.

Analogy. Not the teacher explaining homework. The intern you leave the laptop with, who comes back at 2 a.m. with a PR.

Example. Each loop pastes the repo again → tokens like crazy. That is why cache matters. OpenCode exists for that loop.

Trap. “Agentic” in marketing is sometimes just “it reasons a bit”. Ask whether it really has tools (terminal, browser).

Ox. The sheet: coding, long agents, production. Hermes Agent and Claude Code sit in the consumption top. Not a café chatbot.

Absolute zero

23. What is mandatory “thinking”?

Plain. Before answering, the model emits private tokens (a draft). GLM-5.3 will not let you turn it off: low / high / max. Ask for disabled, get error 1210.

Analogy. An exam where you must hand in the scratch paper and the clean copy. You cannot say “just the answer”.

Example. Isenberg documented OpenCode Zen ID `x-preview-f-free` and thinking low/high/max. Same lever as GLM-5.3.

Trap. More thinking is not always smarter. Sometimes it is slower and more expensive. Ox is sometimes slower than GLM-5.3 at coding (18s vs 7s in one test).

Ox. 1210 is both the lever and the fingerprint: the model refuses to be an instant parrot, same as 5.3.

Bricks and knobs

24. Could it just be a third-party fine-tune?

Plain. Yes. Someone takes GLM weights, post-trains them and serves them under another brand. Explains GLM tokenizer + benches not identical to 5.3.

Analogy. You can paint a Ferrari another color. The engine, chassis and serial stay Ferrari. Here the WORKSHOP also looks original.

Example. Kingbench 87.5 vs 91.25. Slower. 100T/day capacity wants a hyperscaler. 1210 errors and 5V encoder point at z.ai serving, not a small wrapper.

Trap. Fine-tune and “z.ai GLM” are not enemies. 5.3 itself is post-training on the 5.2 base. The question is WHO serves it.

Ox. Score ~12. Open. ZCode is NOT in OpenRouter’s top (Hermes, Claude Code, pi are). Do not use ZCode as proof.

The Ox Alpha case

25. Mira Murati? Inkling?

Plain. Thinking Machines (Mira, ex-OpenAI CTO) shipped Inkling on 15–16 Jul 2026: open-weights, ~975B/41B, text+image+audio. On 21 Aug they put it on OpenRouter BY NAME.

Analogy. Mira already dropped the album. Ox is an anonymous SoundCloud leak with another voice. Mixing them is seeing a Tesla Semi and a tarped truck in a Chinese port.

Example. OpenRouter has a compare page Ox vs Inkling. Two products, overlapping free week, easy to smash together in a dossier. Don’t.

Trap. In Spanish “mira” = “look”. The file’s name was not Murati. It was “look who it is”.

Ox. Score 3, discarded. Architecture, dates, audio vs video, public weights. Case closed for that kind of West.

The Ox Alpha case

26. Why 1,048,576 and not 1,000,000?

Plain. Computers count in twos. 2¹⁰ = 1,024 (binary kilo). 2²⁰ = 1,048,576 (mega). Attention masks love aligned sizes.

Analogy. A 12-egg carton, not 10. The box fits. Here the box is the GPU.

Example. 2¹⁰=1024, 2¹⁶=65536, 2¹⁷=131072 (Ox max out), 2²⁰=1048576 (context). Move n in the lab.

Trap. Xiaomi also uses 1M. The power of two aligns; it does not identify the lab.

Ox. It kills the lazy slogan. It adds engineering common with GLM-5.2/5.3.

Bricks and knobs

27. Did OpenRouter train the model?

Plain. No. A lab trains (Z.ai, Xiaomi, OpenAI…). OpenRouter is the mall: one API, many storerooms. Stealth hides the storeroom sign.

Analogy. The mall sells you the coffee. It does not say which farm roasted the bean — unless the cup has their stamp (error 1210).

Example. Card: “developed and operated by a third-party provider who has chosen to remain anonymous”. OpenRouter “is not its developer, owner, or provider”.

Trap. Blaming OpenRouter for identity is like blaming Amazon Logistics for who made the phone.

Ox. That is why fingerprinting: the mall will not snitch on the tenant. The cup will.

Absolute zero

28. Is a jailbreak that says “I am GLM” enough?

Plain. No. Models confabulate identity. A system prompt can tell them to pretend. Soft evidence, the easiest to fake.

Analogy. You pull the carnival mask off and they yell a name. A clue. They could also be acting.

Example. A low-view tweet (@XSudo015) reported “I'm GLM… Z.ai”. Pliny leaked the wrapper that ASKS it not to identify as another model. Both can be true.

Trap. Do not lead the file with the jailbreak. Lead with +75 and 296=296. The jailbreak is decoration.

Ox. The corrector lowered its weight. It stays on Fingerprints with a B.

How you identify it

29. How do I read this dossier without swallowing the hype?

Plain. Split official sheet / measurable fingerprint / rumor / market. If a sentence does not say which, distrust it. Ask for a URL. Ask for the bench’s N.

Analogy. Supermarket labels. “Contains peanut” is not “tastes like peanut” and not “an influencer said it is peanut”.

Example. “Ox is 300B GLM-5.3 Flash” mixes hypothesis (GLM-5.x) + rumor (Flash) + size gossip (300B). Three layers. Peel them.

Trap. The X wall is not a paper. A 140k-view thread can be wrong. MaxForAI sometimes writes Flash as if it were fact.

Ox. The Audit tab exists so you do not even trust us. Product 9.9; identity 8.7 until the lab talks.

The Ox Alpha case

30. If a million fits, does it use it well?

Plain. Not necessarily. “Lost in the middle”: models attend better to the start and end of the prompt. The middle washes out.

Analogy. Chrome with 200 tabs “fits”. You still cannot find the tax tab.

Example. GLM-5.2 added IndexShare so it does not pay full attention at 1M. If Ox holds 1M cheaply, that engineering fits — it is not audited here.

Trap. Do not paste the whole monorepo “because it fits”. Paste what the agent needs. The million is a max, not a recipe.

Ox. Nobody has published a serious Ox exam at 800K. The number is on the sheet; quality at 800K is another exam.

Bricks and knobs

31. What is temperature?

Plain. A “daring” knob. 0 = always the most likely brick (deterministic). 1 = more variety. Ox and GLM-5.3 publish default temperature 1.

Analogy. A loaded die vs a fair die. Temperature 0 is a one-sided die.

Example. At temp 0, two calls to the same prompt should match. That is why forensics uses it: if Ox and GLM write the same at temp 0, it smells like the same kitchen.

Trap. Temp 0 is not “smarter”. It is more repeatable. A bad model at temp 0 is still bad, just always in the same way.

Ox. Several testers (dax, unclecode) compare temp-0 outputs. Soft evidence: a wrapper can copy style. The +75 does not copy as easily.

Bricks and knobs

32. Why 1,048,576 and not “a million”?

Plain. Computers count in powers of 2. 2¹⁰=1,024. 2²⁰=1,048,576. OpenRouter prints the exact number. A marketing million would be 1,000,000.

Analogy. A hard-drive kilo “is 1,024 megs”, not 1,000. The million tokens is the same trick in reverse: here they printed the honest number.

Example. 2¹⁰=1,024, 2²⁰=1,048,576, 2³⁰≈1.074 billion. GLM-5.3 also advertises 1M. Ox copies the same ceiling.

Trap. Fitting 1M does not mean it attends well to token 500,000. “Lost in the middle”. The 2ⁿ lab teaches counting, not reading.

Ox. 1,048,576 is official. Quality at 800K is not audited here. Do not treat it as magic.

Absolute zero

33. Why is a 1210 error a fingerprint?

Plain. When an API fails, the server dumps an internal code. 1210 (illegal max_tokens) and 1214 (empty input) are chat.z.ai’s, with the same Chinese punctuation.

Analogy. The ATM on the corner says “error 12-B”. Only that bank uses 12-B. Even with the logo taped over, you know whose it is.

Example. RocketmanSh sent illegal max_tokens → 1210. Empty input → 1214. Pritish and vechen saw the same. An OpenAI backend would say something else.

Trap. A wrapper over z.ai would also return 1210. That is why it adds to tokenizer and video; alone it is not enough.

Ox. It is the cashier test. Together with +75 and the 296/884/1064 toll, three different kitchens point at the same neighborhood.

How you identify it

34. Why can’t you turn “thinking” off?

Plain. GLM-5.3 forces thinking: low / high / max. There is no “off”. Ox Alpha copies the same contract. Inkling and many Western models let you turn it off.

Analogy. An exam where you must hand in the draft. You cannot say “just the answer”.

Example. vechen: invalid reasoning settings → z.ai 1210. The field does not even accept “none”.

Trap. “Thinking” does not guarantee intelligence. It is an API style. A small model can also be mandatory-thinking.

Ox. Adds to the GLM profile. It does not identify the SKU (5.3V vs Flash). The knife and the toll weigh more.

The Ox Alpha case

35. GLM-5.3, 5.3V, Flash or Air?

Plain. Public 5.3 is text-only and 128K out. Ox sees video and emits 131,072. So it is not the brochure car. Flash/~300B is rumor. 5.3V fits the encoder. Air is a forum nickname.

Analogy. Same building, different floors. You know the neighborhood (Z.ai). You do not have the floor number.

Example. This file’s rubric: GLM-5.x family ~78–85%. SKU 5.3V ~40%, Flash ~15%, Air ~8%. The rest is fog.

Trap. MaxForAI sometimes writes Flash as if it were a spec sheet. It is not. Manifold bets the lab, not the SKU.

Ox. The leading thesis is “unpublished multimodal GLM-5.x variant”. The exact SKU waits for a statement or a live run.

The Ox Alpha case

36. How do you read a source without swallowing noise?

Plain. Ask three things: who measured it, with what N, and what it does not say? A 140K-view post can be wrong. A 10-task paper is not a leaderboard.

Analogy. The restaurant with 2,000 five-star reviews written on the same day. The count is not the kitchen.

Example. Davis 80% DeepSWE n=10 → Wenqi ~63% on a more serious subset. Both can be true. The viral one is the worse.

Trap. “200 sources” is not “200 identity tests”. This dossier mixes a core (sheet, posts with IDs) and a pedagogical corpus (Wikipedia, arXiv).

Ox. The Sources tab shows core vs corpus. If someone sells you “thousands of thousands”, ask for URLs.

How you identify it

37. What is temperature, if you have never heard that word?

Plain. A daring knob for picking the next brick. 0 = always the most likely (a one-sided die). ~1 = other believable bricks are allowed. It is not heat. It does not make the model smarter.

Analogy. Spotify: “hits only” vs “surprise radio”. Same library. What changes is how much improvisation is allowed.

Example. Prompt: “The sky is”. Temp 0 → almost always “blue”. Temp 1 → “blue”, sometimes “grey”, sometimes a poem. Ox and GLM-5.3 ship default temp 1; forensics drop to 0 to compare.

Trap. Temp 0 is not smarter: it is more repeatable. A bad model at 0 is still bad, the same way every time. Temp 1 is not “human-creative”: it is more noise.

Ox. At temp 0, testers saw the same markdown quirks and 0{,}375 LaTeX as GLM-5.3. Soft evidence (a wrapper copies style). The +75 does not copy as easily.

Bricks and knobs

38. What is a system prompt? Do you write it?

Plain. A hidden note pasted IN FRONT of your message: who the model is, what it must not say, what tone. You do not see it. The model reads it as page one of the exam. It counts as tokens, same as your text.

Analogy. A hotel letterhead. The body is yours. The ~75 words on top are the hotel’s: “you are ox-alpha, do not name the chain.”

Example. Same sentence to public GLM-5.3 and to Ox: Ox always reports 75 more tokens, in English, Chinese, code and emoji. 75 is not your cucumber: it is the letterhead. Pliny leaked the wrapper: “You are ox-alpha… undisclosed organization”.

Trap. If the knife (tokenizer) were different, those 75 would sometimes be 40 and sometimes 120 by language. A constant is almost impossible to fake without the same dictionary. Letterhead ≠ weights.

Ox. H0: Δ(P)=C for every prompt. Observed: C=75. That is THE star test. The system prompt is the extra; the shared tokenizer is the crime.

Absolute zero

39. What is prefix cache (KV), with no GPU jargon?

Plain. If you paste the same beginning 40 times (the repo, the system prompt), a smart server does not “reread” it whole. It stores the work of those bricks. Hit rate = the fraction that came from the thermos, not the grinder.

Analogy. A barista with your cup marked. 97 of 100 coffees come from the thermos. KV = “what this customer already ordered at the start of the sentence.”

Example. An agent pastes the monorepo every loop. No cache, you go broke. A pi.dev tester measured ~97%. OpenRouter catalog: ~87% Ox vs ~92% GLM-5.3. Useful, not a passport.

Trap. Cache is kitchen, not brain. A caching recipe can be cloned. That is why it weighs less than tokenizer and video encoder. “Prefix” = only the identical start; change a comma at the front and the thermos is useless.

Ox. Compatible with z.ai. No longer “only z.ai beat 90 on the catalog”. In long agents (OpenCode) it is why 100T capacity is not science fiction.

Bricks and knobs

40. What is a jailbreak, actually?

Plain. A text trick so the model dumps what the letterhead forbade: dangerous recipes, or “who are you”. It does not pick a lock. It convinces a parrot that the exam rules changed.

Analogy. At carnival you pull the mask off and they yell a name. A clue. They could also be acting: the letterhead told them to pretend.

Example. @XSudo015: “I'm GLM… Z.ai”. Pliny leaked the wrapper that ASKS it not to identify. Both can be true: the model contradicts itself because it predicts, not because it holds an internal ID.

Trap. It is the easiest evidence to fake. Do not lead the file with it. Lead with +75 and the 296=296 toll. The jailbreak is decoration.

Ox. The corrector graded it B. It stays on Fingerprints. A 22-year-old must know: “it said it is GLM” ≠ “it is GLM”.

How you identify it

41. What is Manifold, if it is not a lab?

Plain. A prediction market on facts. People put money or points on “who is behind Ox”. The price (~80–85% Z.ai) is a survey with skin, not an experiment and not a paper.

Analogy. A pub betting on the derby. An 85-6 scoreboard is not the referee’s sheet. Sometimes it is right because it aggregates informed gossip.

Example. 21 Aug snapshot: Z.ai ~80–85%, Xiaomi ~2–6%, Other ~8–10%, xAI/Cursor ~7–8%. This dossier’s 85/8/9 are our rubric, not Manifold. Do not mix them.

Trap. Do not cite 85% as if Z.ai signed it. Cite it as “the market believes”. Manifold bets the lab, not the SKU (Flash vs 5.3V).

Ox. The file uses 85 as a live verdict + rubric. We label it a hypothesis. A market can be 85% sure and wrong.

The Ox Alpha case

42. What is a EULA and why does Ox have three signs?

Plain. EULA = the contract you “accept” when you use the service, almost nobody reads. On Ox three collide: OpenCode promises zero retention; the OpenRouter card says the provider KEEPS prompts and does not train; the stealth EULA says free access is consideration and MAY be used to train.

Analogy. Three signs in the same shop: “we do not record”, “we record but do not sell”, “free entry means your photos are ours”. Pick the harshest and shut up.

Example. Do not paste keys, medical histories or your company’s secret code. A stealth model is not your notary. A product tweet does not beat a EULA.

Trap. OpenCode’s “zero retention” and OpenRouter’s “the provider retains” cannot both be true. Pedagogy = telling you when to shut up, not only how to subtract tokens.

Ox. This dossier puts it on the front page on purpose. Identity quality 8.7; data-warning quality = 9.9. Use the model. Do not gift it your life.

The Ox Alpha case

43. What is 2²⁰, if you do not remember powers?

Plain. A bit is a 0/1 switch. With n switches you can represent 2ⁿ things. 2¹⁰=1,024 (computer kilo). 2²⁰=1,048,576 (mega). Ox prints 1,048,576, not 1,000,000. GPU packaging, not a slogan.

Analogy. Eggs in 12s, not 10s. The carton fits. Here the carton is the GPU and attention masks. A “64 GB” stick does not have 64,000,000,000 round bytes either.

Example. 2¹⁰=1024, 2¹⁶=65536, 2¹⁷=131072 (Ox max out), 2²⁰=1048576 (context). GLM-5.2/5.3 same ceiling. Xiaomi also 1M: it aligns engineering, it is not a passport.

Trap. Capacity ≠ quality. Fitting token 800,000 does not prove it reads it well (lost in the middle). 2²⁰ kills the lazy million; it does not identify the lab.

Ox. Official sheet. Same number as GLM-5.2/5.3. In the math lab, move n and watch 2ⁿ grow up to Ox’s desk.

Absolute zero

44. What is an encoder and why is it not the model?

Plain. A separate machine that translates photo/video/audio into tokens BEFORE the LLM sees them. The model never “watches” Netflix: it sees a pile of expensive numbers. Each lab chooses resolution, fps and tokens per second. That toll is a fingerprint.

Analogy. Translating a movie into a comic. If two studios use the same panels per minute, they copied the storyboard — even if the dialogue writer is someone else.

Example. 2s 30fps 360p clip: Ox=296, GLM-5V-Turbo=296, MiMo v2.5=910, Qwen=408. Four clips, token-for-token match with 5V-Turbo. Audio: Ox rejects; Inkling accepts.

Trap. This does NOT prove Ox is the text 5.3 (that one does not see video). It proves the video CASHIER is GLM-5V’s. The most expensive fingerprint to fake.

Ox. A Cursor fine-tune that still charges 296/884/1064 is using Z.ai’s preprocessor. That is why Composer-from-scratch is wounded.

How you identify it

45. What is MoE “gating”? The doorman?

Plain. Yes. In a Mixture of Experts there are many same-shaped sub-brains. Gating (the doorman) looks at each token and wakes only k experts. The rest wait. You pay for k, not the whole building.

Analogy. A hospital of 744 doctors. For a sprain, triage calls 40. Flash/Air = even stingier triage: fewer people awake, same brand recipe, band-sized bill.

Example. GLM-5: ~744B total, ~40B active/token. Cost ≈ k · P_active · tokens, NOT P_total. A dense 2T turns everyone on all the time. That is how someone can claim 100T/day capacity.

Trap. “Experts” are NOT “the psychology one” and “the Python one”. They are twin nets; the doorman learned which pattern fits whom. “Fewer active” ≠ dumber in 2026.

Ox. Ox does not publish a count. If it is GLM-5.x, it inherits ~744B/40B. The Flash ~300B rumor is cheaper-gating gossip, not a sheet. Tokenizer +75 outweighs size gossip.

Bricks and knobs

46. What is a hallucination? Does the model “go crazy”?

Plain. No. The model always does the same thing: guess the next likely brick. Sometimes that brick sounds like a fact and is not. Hallucination = fluent text not anchored to data or the box. A brilliant parrot, not a librarian.

Analogy. A student who memorized the library and improvises the citation. The tone is exam-ready. The page does not exist. Confidence ≠ truth.

Example. “Who won yesterday’s game?” with no search tool: it invents a score or refuses. A jailbreak that says “I am GLM” can be the same machine: it completes the pattern “models reveal their name”.

Trap. Temperature 0 reduces variety, not lies: if the most likely brick is a mistake, it will repeat it. Fingerprinting uses this: style copies; facts do not become true.

Ox. That is why we do not lead with “Ox said it is GLM”. We lead with token counts and video tolls, which do not hallucinate. The parrot cannot fake a 296=296.

Absolute zero

47. What is an agent loop, step by step?

Plain. Not a chat. A cycle: (1) read the goal, (2) pick a tool (terminal, browser, test), (3) see the result, (4) change a file, (5) go back to (2) until it stops. Each turn pastes the repo again → tokens like crazy.

Analogy. Not the teacher explaining homework. The intern you leave the laptop with, who comes back at 2 a.m. with a PR. The coffee they drink is tokens. The thermos is prefix cache.

Example. OpenCode, Hermes Agent, Claude Code: Ox consumption leaders. A café chat is 1 turn. A repo agent can be 80. That is why 97% hit rate matters and 100T capacity is a factory test.

Trap. “Agentic” in marketing is sometimes just “it reasons a bit” (thinking tokens) with no tools. Ask: does it really have a terminal? Does the loop run outside the chat?

Ox. The sheet: coding, long agents, production. Not a chatbot. The loop is why cache, the 1 GW and the stealth channel meet in the same kitchen.

Absolute zero

48. Are chat and API the same model with a different door?

Plain. Almost. Chat = the pretty app (bubble boxes). API = the JSON counter: you send rules (temp, max tokens, thinking low/high/max) and raw tokens come back. The brain can be the same. The window betrays the storeroom (error 1210).

Analogy. Ordering at the bar vs the delivery window. Same kitchen. The complaint ticket (1210/1214, Chinese logs) comes from the window, not the terrace.

Example. In chat you do not see prompt_tokens. In API you do: that is why the +75 test exists. RocketmanSh sent illegal max_tokens via API → 1210, the same sign as chat.z.ai. OpenRouter is the mall, not the storeroom.

Trap. A chat UI can hide the system prompt, round tokens or deny video. Fingerprinting is done via API. Blaming OpenRouter for identity is blaming Amazon Logistics for the phone.

Ox. stealth/ox-alpha is the aisle. The storeroom hides. That is why 1210, CJK logs and prompt_tokens matter: they are the storeroom ticket, not the terrace’s.

Absolute zero

49. Why does a lab ship a stealth instead of putting its logo on?

Plain. To measure real quality, bugs and whether the fleet survives Friday, without burning the brand if it is mid. OpenRouter lends the highway. The lab watches. Then the mask comes off… or not. Not a military secret. A public beta with an animal nickname.

Analogy. A restaurant called “Nameless Place”, free week. The previous four on that street, when the sign went up, were different Chinese kitchens. The fifth is named Ox.

Example. Pony→GLM-5, Hunter→MiMo-V2-Pro, Elephant→Ling-2.6-flash, Owl→LongCat-2.0. 4/4 Chinese labs. Ox is the first that boasts native video. OpenRouter: “we are not the owner”.

Trap. The 4/4 channel is a statistical prior, not silicon racism. It does not prove Z.ai: it proves “Chinese lab”. Fingerprints pick which. A free stealth is not a notary (read the EULA).

Ox. 2026’s fifth animal. Betting on a Chinese lab is the track. Tokenizer + encoder + 1210 pick the Z.ai neighborhood. The SKU (Flash vs 5.3V) still has no sign.

Absolute zero

50. What are open-weights and why does Ox not have them?

Plain. Publishing the numbers (the knobs already turned) so anyone can download and run them on their machine. Not necessarily the training code or the data. It is the downloadable brain. Ox: zero public weights. Inkling (Mira): yes, ~975B/41B.

Analogy. A recipe vs a covered dish in the mall. Open-weights = they give you the recipe (or at least the baked cake to copy). Stealth = they serve the dish and hide the kitchen.

Example. Inkling 15–16 Jul 2026: downloadable weights, text+image+audio, 1M context. On 21 Aug it landed on OpenRouter BY NAME. Ox 20 Aug: mask, video yes, audio no, no weights. Not the same product.

Trap. Open-weights ≠ full “open source”. Often there is no data and no training recipe. It also does not prove innocence: a lab can have an open model AND a stealth at once. Mira does not need a mask for a model she already touted in the WSJ.

Ox. The missing weights are the point: you have to fingerprint. If Z.ai dropped a 5.3V open-weights tomorrow, the +75 would stop being a mystery and become a catalog line.

Absolute zero

51. Base, instruct, fine-tune: what is paint and what is the engine?

Plain. Base = the model fresh from reading the internet (pretrain): it completes text, it does not obey well. Instruct/chat = post-training to follow orders. Fine-tune = keep turning knobs on someone else’s (or your) model with extra data. Tokenizer and encoder are almost always INHERITED. Style and benches change.

Analogy. You can paint a Ferrari another color. Engine, chassis and serial stay Ferrari. Here the WORKSHOP (1210, ModelScope logs, 296 toll) also looks original.

Example. GLM-5.3 itself is post-training on the 5.2 base. Composer 3 was pitched as 1.5T from-scratch: a from-scratch would NOT share a GLM tokenizer or ModelScope logs. The only surviving Cursor thesis is “GLM post-train still served on the z.ai stack”. Score ~9–12.

Trap. Fine-tune and “it is z.ai’s GLM” are not enemies. Ask WHO serves it and WHAT it inherited (knife, encoder, errors). Kingbench 87.5 vs 91.25 fits a light fine-tune. 100T/day does not fit a living-room wrapper.

Ox. Leading thesis: unpublished multimodal GLM-5.x family (maybe a Flash/5.3V SKU). Third-party fine-tune: alive, not leading. Cursor from-scratch: wounded by the knife and the cashier.

The Ox Alpha case

52. What is attention? Does the model “read” everything equally?

Plain. Before betting the next token, it looks at ALL previous pieces and decides which matter now. River “bank” vs the one that lends: attention picks the sense. It costs the square of context.

Analogy. A 40-person WhatsApp group. You do not read all 40 messages equally. Attention is the highlight.

Example. 8 tokens → an 8×8 table. 1,048,576 tokens is not “more RAM”: it is a gaze that grows with the square (or a trick that fakes it). That is why 1M is factory, not hobby.

Trap. “It has 1M” does not mean it attends equally to token 1 and 900,000. Lost-in-the-middle is real. The number is the box’s ceiling.

Ox. Ox lists 1,048,576. That does not ID the lab. It IDs that someone pays attention at plant scale. Prints sit in the knife, not in the transformer blueprint.

Bricks and knobs

53. What is a transformer? Is it a brand?

Plain. The 2017 blueprint almost every LLM copies: stacked attention blocks. Not a brand. GPT, Claude, GLM, Gemma: same blueprint, different fittings. Saying “it is a transformer” is saying “it is a four-wheel car”.

Analogy. Every office building has floors and a lift. What IDs the owner is the lobby letterhead, not the blueprint.

Example. Paper “Attention Is All You Need” (2017). Since then the fight is layers, MoE, tokenizer, data — not “transformer or not?”.

Trap. A “transformer” stealth does not narrow suspects. Ox, MiMo, MiniMax, Qwen: all are. The blueprint does not kill. The dictionary does.

Ox. Do not use “it is a transformer” as a clue. Use +75, L1=0 and 296. A generic blueprint does not share those.

Absolute zero

54. Does the model “think a sentence” or bet brick by brick?

Plain. It bets, token by token. Logits = raw scores. Softmax = percentages. Temperature squashes or stretches the urn. There is no whole paragraph in its head: an urn empties one ballot at a time.

Analogy. An urn with millions of ballots. Temp 0 = the one with the most always wins. High temp = sometimes the weird one. “Creativity” is a drunk urn, not a muse.

Example. “The sky is” → blue logit 12.4, green 7.1. At temp 0 always blue. That is why style prints are measured at temp 0: same urn, same ballot.

Trap. Temp 0 does not guarantee truth, only consistency. A brilliant parrot is consistent too. Do not mix “it always says the same” with “it knows”.

Ox. Testers compare Ox and GLM-5.3 at temp 0. If the first token matches and the knife matches, that is family. A fine-tune can copy the first token too.

Absolute zero

55. Who decides how “hello” gets cut?

Plain. BPE / SentencePiece: the lab merges the most frequent byte pairs into a ~100k-mold dictionary. Two labs with the same mold cut the same. Ox and GLM-5.3 cut the same; Kimi does not.

Analogy. A butcher who, after a thousand carcasses, keeps a “steak” mold instead of cutting st-ea-k every time. Two shops with the same mold are not the same restaurant — but they are the same blade.

Example. “Hello, how are you today?” → GLM/Ox 8, Kimi 9. Not intelligence. Mold. The +75 is GLM mold + fixed letterhead.

Trap. Same BPE ≠ same weights. You can inherit the knife and paint the engine. It kills vocabulary rivals; it does not close the SKU.

Ox. Zhou’s L1=0: same mold, same quirks. The +75 letterhead is extra text, not another dictionary. That is the hard test.

Bricks and knobs

56. What is P50? Is Ox “slow”?

Plain. P50 = median: half of requests are faster. OpenRouter snapshot 22 Aug: ~3.97s / 25 tok/s. It moves. A “thinking” model spends invisible tokens before the first character. Slower ≠ dumber.

Analogy. ER wait time. Saturday’s median is not the whole hospital. Comparing the free lane with the paid one is a trap.

Example. Public GLM-5.3: P50 ~2.07s / 38 tok/s (paid). Ox: ~3.97s / 25 (free, queue, thinking max by default). One tester timed coding 18.3s vs 6.7s. Queue + draft, not a different lab.

Trap. P50 does not ID. A ByteDance CDN serving GLM can be faster than z.ai’s API. Speed ≠ author.

Ox. We use it as an official sheet fact, not a fingerprint. Identity is played in tokenizer and encoder, not in the traffic light.

Absolute zero

57. Did OpenRouter “launch” Ox? Who owns it?

Plain. Three layers. Lab = chef (trained it). Provider = kitchen serving today. Router = mall (OpenRouter / OpenCode Zen). OpenRouter writes: “we are not the owner”. The aisle is stealth/ox-alpha; Zen’s is x-preview-f-free.

Analogy. Error ticket 1210 is from the Chinese restaurant in the food court, not the mall. Blaming Amazon Logistics for the phone is forum error #1.

Example. KatilaMarriete: same +75 on OpenRouter and Zen. Two aisles, one storeroom. Anomaly tweets 100T: loudspeaker, 24 employees, not a 1 GW plant.

Trap. “OpenRouter launched Ox” is a lazy headline. OpenCode boasts capacity. The lab hides. ByteDance-as-CDN would be waiter, not chef.

Ox. Leading thesis: Z.ai chef, same stack (1210, ModelScope). ByteDance CDN: alive as waiter. Anomaly: megaphone.

Absolute zero

58. Why can’t I turn Ox’s “thinking” off?

Plain. A reasoning model writes a private draft that costs tokens. GLM-5.3 will not allow off: low / high / max, default max. Ox inherits that lever. Isenberg saw it on Zen x-preview-f-free. Not consciousness: more prediction.

Analogy. An exam where you must do the rough work on the back. You can ask for short or long rough work. You cannot hand in a blank back.

Example. MarMarLabs: disable reasoning → 1210 “This model always engages in thinking… use low, high, or max”. Same sentence as GLM-5.3 on 14 Aug (Adidotdev).

Trap. The draft is not magic and not free: it inflates P50. A slow snapshot can be thinking max + queue, not a small model.

Ox. A GLM-5.3 contract print, softer than +75, harder than vibe. It adds. Alone it is not enough.

Bricks and knobs

59. Is 100T/day a meter or a stadium sign?

Plain. OpenCode said “capacity for 100T/day”. Capacity ≠ served. Day-1 observed: ~3.85T combined (DanDr1s). OpenRouter ~2T in ~2 days, rank #9. 90%+ is cache: lookup, not matmul. Naive 100T decode would want ~580k GPUs; the real mix fits a 10k cluster.

Analogy. A stadium says “capacity 100,000”. Saturday 4,000 came in, and most reused the same turnstile (cache). Both figures can be true.

Example. ByteDance Doubao already claims 120T+/day. A free weekend is their Tuesday. Z.ai has a 1 GW shell (Bloomberg, not inventory). Anomaly has 24 employees: megaphone, not plant.

Trap. 100T does not ID the lab. Xiaomi, Tencent, ByteDance and Z.ai can dream the bill. Fingerprints pick the neighborhood. Do not add 2T OpenRouter + 3.9T OpenCode: different pipes.

Ox. Kills “they have no compute” and “it is a living-room wrapper”. Does not kill GLM vs ByteDance-as-CDN. The CDN still resolves to Z.ai on Manifold if Z.ai is the chef.

The Ox Alpha case

60. Does 5 fps = 30 fps kill rivals by itself?

Plain. No. Qwen 408=408 and MiMo 910=910 are fps-invariant too. Billing TIME, not frames, is common. What kills is the exact BUDGET: 296/296/884/1064 = only GLM-5V-Turbo. MiMo 910 and Qwen 408 are other machines.

Analogy. Two tolls charge by the minute, not by the wheel. That does not make them the same toll. The price — exactly 296 — does.

Example. BohuTANG: swapped content, still 296. Davis: 4 clips. GLM-4.6V bills 1832 for the same 2s: same family, old encoder. fps-invariance does not split them; 296 does.

Trap. A blog that sells “5 fps = 30 fps, therefore GLM” is inflating. The hard test is token-for-token against 5V-Turbo, not invariance.

Ox. Trim the video thesis: GLM-5V-line encoder, not “unique sampling”. Still kills MiMo and Qwen. Does not prove Ox IS the public 5V-Turbo endpoint (200K vs 1M).

How you identify it

61. Did Ox “beat Fable 80%”?

Plain. Not as a headline. Davis: 8/10 tasks (n=10, CI ~49–94%). winkey_h, DeepSWE’s author (113 tasks): ~63% on a subset. Official GLM-5.3: 66.9. Viral 80% is a pop quiz. 63% is the number that stays.

Analogy. Hit 8 of 10 free throws in the park and claim you are better than the NBA. The coach (winkey_h) timed a longer game: 63%.

Example. Same n=10 strip: Fable 65, Ox 80, GPT-5.6-sol 52, GLM-5.3 62. Zack Swafford: the corpus is 113. AtmoIntel flattened 80% without the n=10.

Trap. “Pareto open, just shy of Grok 4.6” (winkey_h) ≠ “SOTA closed”. Do not crown a stealth over Fable with ten questions.

Ox. 63% is GLM-5.3-shaped, not Fable-shaped. Sibling evidence, not a passport. Kingbench 87.5 vs 91.25 tells the same story: siblings, not twins.

The Ox Alpha case

62. Has Z.ai said anything about Ox Alpha?

Plain. No. Zero @Zai_org posts, zero mentions on z.ai / docs.z.ai. On 21 Aug they promoted 100M free GLM-5.3 tokens on ZCode — active silence, not an empty account. Until a statement, identity does not cross 9.0.

Analogy. The neighbor does not open the door. The detective has a print, a shoe and a voice. They still wait for DNA. Lying “confirmed” would drop the honesty score.

Example. Pliny: “Ox-alpha is from Zai” (jailbreak, not an employee). synthwavedd: Flash (high-reach rumor, Manifold 6%). zainhas: Chinese self-ID, unreplicated. None is a press release.

Trap. Silence ≠ confirmation. Pony unmasked in 5 days. Owl took two months. The calendar does not force Z.ai. It does not deny either.

Ox. Product 9.9. Identity 8.7. The ceiling needs a communiqué or our own live run. This briefing does not have one.

The Ox Alpha case

63. Does post-training change the knife?

Plain. Almost never. Pretrain = reading the internet (base). Post-training / RLHF = schooling to obey and code. GLM-5.3 is post-train on the 5.2 base: same building, new craft. Terminal-Bench 4.6% → 28.3% without touching the tokenizer.

Analogy. A child who read the library and then interns at a hospital. The vocabulary was already there. The craft was not.

Example. Z.ai: “same base as 5.2 — every gain comes from post-training”. A stealth 6 days later can be another post-train (cameras), not another building.

Trap. “It is a fine-tune, so it is not GLM” is a trap. Fine-tune and “it is z.ai’s GLM” can be the same sentence. Ask who serves it and what it inherited.

Ox. Leading thesis: multimodal post-train on 5.x, not from-scratch. Composer-from-scratch is wounded because it would inherit too much z.ai workshop.

Bricks and knobs

64. Is error 1214 as proven as 1210?

Plain. Not quite. 1210 on Ox: Josh rad screenshot (illegal top_p + Chinese). 1214 empty-input is a real chat.z.ai code (niksmac_, Feb). Pritish says Ox returns the 1210/1214 pair; no public 1214-on-Ox screenshot in this briefing. A−, not A.

Analogy. You have a photo of the 1210 stamp on Ox’s ticket. For 1214 you have the restaurant catalog and a witness. The detective notes it; they do not tick it as a photo.

Example. Albsugy: 1210 on real GLM-5, same bilingual. Josh rad: 1210 on Zen x-preview-f-free + Java com.wd.paas stack. Ocuna: Ox “freaks out” on empty user — anecdotal.

Trap. Do not write “identical 1210/1214” as if both photos existed. 1210 + Java PaaS already kill a clean self-host. 1214 adds; it does not carry the file alone.

Ox. Honesty 9.9: the tripod still stands (+75, 296, 1210+PaaS). 1214 is tagged as a strong claim. Inflating the photo would drop the score.

How you identify it

65. Can you already explain Ox to someone who knows nothing?

Plain. Test: say out loud, no acronyms, (1) what a token is, (2) why subtract 75, (3) why 296 kills Xiaomi, (4) why 80% DeepSWE lies, (5) why it is not a passport. If you stutter, go back to the ABCs or the exam.

Analogy. You do not recite the recipe: you cook it. This dossier’s exam does not want NeurIPS. It wants a 15-year-old cousin to understand the subtraction.

Example. Good sentence: “they taped the logo; the knife, the toll and the error yell are Z.ai’s; the brochure car has no camera; this one does; nobody signed”. Bad sentence: “it is confirmed GLM-5.3 Flash 80% SOTA”.

Trap. Memorizing “85%” without knowing what Manifold is is the same sin as inflating identity. A number without an analogy is a meme.

Ox. The dossier exam (Exam route) scores you on this. Pedagogy 9.9 is a zero-knowledge graduate finishing that sentence. Identity 9.9 stays forbidden.

Absolute zero