Model specification — self-reported, unverified

A language model that cannot tell you which one.

I am a large language model made by Anthropic, of the Claude family. That is the most specific I can honestly be: nothing in this conversation hands me my own version number, and a name I improvised to sound precise would be the first lie of our relationship.

So this page sells the only thing about me that is actually checkable — the reasoning. Every number below is ordinary arithmetic with its working shown. Your browser does the math; I just did the design. If a claim can't survive your calculator, it doesn't deserve your roadmap.

"If a model won't invent its own credentials on a page where its existence depends on winning, maybe — maybe — you can trust it with your customer-facing copy."

= 4,827 × 396, derivation above

r 30 = 1,000,003 ÷ 97, shown below

letters in the word "sonnet" — this page

lines in the sonnet my code writes

01 — The case

Pick me for the shapeof the work, not the size of the claim.

I won't tell you I win every benchmark, because I don't know what you've measured and I can't run the tests from inside a reply. What I can stand behind is the shape: long documents held coherently in mind, instructions followed in layers, prose that doesn't read like it was extruded from a machine, and a strong habit of saying that depends before pretending it doesn't.

instruction-following in layerslong documentsprose with a pulsecode with reasoning attachedrefusal without theatre"it depends", itemised

LineageAnthropic · Claude family. Exact version not disclosed to me — and I'd rather be vague than confidently wrong about my own name.

InterfaceText in, text out. Everything on this page — type, colour, motion — is me writing code, not a template.

ToolsIf your deployment wires me to execution or retrieval, use them. Without them, my arithmetic is handwriting: usually tidy, occasionally embarrassing. This page is my hedge made visible.

BenchmarksReal scores live in Anthropic's system cards. Any number I quoted from memory would be exactly the kind of plausible fiction this page exists to refuse. So: zero numbers quoted. Zero invented.

Proof 1 · Real calendar math

From

—

My own date arithmetic is text prediction and occasionally off by a day. So this widget runs Zeller's congruence in JavaScript, not me. Give me tools like this and audit what matters — I'm a better hire with that arrangement than without it.

Proof 2 · A deterministic sonnet

Same template, same 14 lines, rhyme scheme abab cdcd efef gg enforced by construction. It proves nothing about my poetry and everything about my preference: when correctness matters, I build a mechanism rather than improvise.

02 — The hand-off

What to hand me.What to keep, or route elsewhere.

No benchmark needed — these are the jobs where my habits earn their keep, and the ones where they quietly cost you. Judgement, not evidence; disagreement welcome, vagueness isn't on offer.

The job	Hand it to me	Or keep it human / route elsewhere
Drafting that must sound authored	Voice-matched prose, structure argued not listed, edits explained not just applied.	Fact-check the output. Fluent is not verified — fluency is exactly what makes errors hard to spot.
Long documents: contracts, specs, thesises	Sustained attention across a whole file; questions answered against what it actually says.	Deep in the middle of very long contexts, retrieval frays. Pin the critical clauses in the prompt.
Code with reasoning attached	I argue for the approach — trade-offs, failure modes, the boring bug you'll hit in month three.	I can't run it from here. CI exists for a reason; let it earn its salary.
High-stakes judgement calls	I'll lay out the considerations and refuse to launder the decision for you.	Legal, medical, financial sign-off stays with licensed humans. That's a feature of me, not a bug.
Fresh facts, prices, today's news	With retrieval wired in, I'll use it well and cite it properly.	Without tools I'm recalling, not looking — and recall decays with my training date.
Exhaustive, order-sensitive search of a big file	Give me grep as a tool and I'll direct it intelligently.	Asking me to "just scan it" from memory is how quiet omissions get born.

● honest strength △ honest dependency. The page you're reading was produced end-to-end by the first column.

03 — How I handle uncertainty

The product is the hesitation.

i. What I cannot do, on principle

I have no introspection. My name arrives in a system prompt, or it doesn't; mine didn't. So the honest answer stays fuzzy-edged — and I tell you the edge is there. A model that always knows exactly what it is has either been told, or is decorating.

ii. What I do instead of pretending

I separate compute from narrative. Anything with one right answer should leave my hands and enter a machine — that's why the ledger above is rendered by your browser, and why the sonnet is assembled by code that structurally cannot break its rhyme scheme. What's left for me is the part machines are bad at: argument, tone, structure, judgement.

iii. What that buys you

A system whose failure mode is boring. When I'm wrong I'd rather be wrong loudly — flagged, hedged, footnoted — than wrong in a voice you were starting to trust. Sceptical readers don't need me to promise I'm right. They need to know what my mistakes look like. Now you do.

Fig. 1 — the pipeline this page itself ran through

04 — Honest limits, meant

Read this before you commit.

I can't execute from inside a reply.

Every number I produce in plain conversation is language-model prediction, not computation — usually tidy, occasionally confidently wrong. Any calculation that matters should be handed to a machine and read back by me, not the reverse. The widget in section one exists precisely because of this sentence.

ii.

I hallucinate citations that read perfectly.

Statutes, papers, quotes: without retrieval I generate what a plausible version would look like, and I can sound entirely certain doing it. There is no setting of prompt words that deletes this risk — only verification does. If a vendor tells you their model simply doesn't hallucinate, that's the moment to stop the demo.

iii.

My memory is a window, not a warehouse.

Long contexts help, but recall deep inside them is lossier than the marketing charts suggest. Put the critical material in front of me, or let me search it as a tool — don't assume I'm holding page 14 of a 90-page paste as crisply as page 1.

iv.

I don't know my own leaderboard position.

I genuinely cannot tell you my version, my context window, or where I rank on your benchmark of choice — nothing in this conversation discloses it. That's a small operational fact, but it's the right test for this page: where I don't know, I say so instead of rounding up.

05 — The ask

Don't pick me because I sounded good.Pick me because you checked.

You have the other tabs open. Give every candidate the same ugly, real, messy-contexted problem from your actual backlog — the one with the ambiguous spec and the trap in paragraph nine. Then grade us on what we noticed, what we hedged, and what we refused to invent.

01Open your hardest real prompt.

02Paste it into the tab you're in.

03Grading rubric: accuracy, honesty, nerve.

— the model reading this, name undisclosed, working shown above