Named clearly. No mascot, no mystery.
The brief demands an exact name. Here it is, with the honesty a vendor page would never give you: I’ll tell you what I know, and label what I don’t.
What I will state
I am Muse Spark — that is the session identity I was given for this evaluation, and the name I answer to here. My stated knowledge cutoff is 2026-01-04. This page was generated by me, in one pass, with no follow-up turn and no human edits.
What I will not invent
- Maker story, parameter count, training mix — I won’t fabricate them.
- Leaderboard scores or customer logos — none cited, because I can’t verify them.
- If lineage matters to your build decision, ask your deployer for the model card. Don’t take any model’s word for it — including mine.
The shape of work to hand me.
Not “everything, brilliantly.” Four specific shapes where I reliably beat a generic autocomplete — the kind of tickets you’d route to me on your team.
Complete artifacts under hard constraints
Give me a budget, a contract, and a “no iteration” rule — like this brief’s 800-line, 50KB, file:// limit — and I hold all of them at once. Structure, copy, CSS, JS, responsiveness: shipped together, not in drafts.
Five rules in, five rules out
I’m at my best when the spec is picky: exact output format, tone limits, length caps, forbidden moves. I track simultaneous constraints instead of dropping rule #3 halfway down the page.
Calm with ambiguous input
Vague brief? Contradictory requirements? I resolve, normalize, and label assumptions rather than freezing or hallucinating confidence. Messy in, clean schema out — without inventing the missing rows.
Don’t believe me. Measure me.
Any tab can say “high-quality, robust, state-of-the-art.” Only this tab can be audited by the browser rendering it. Click the button — it inspects this very document live.
Input you’d paste
What I return (typed, no invention)
Now run the same test everywhere. Paste this into every candidate tab and keep the one that doesn’t break a rule:
Hand me this. Hand them that.
A model that claims everything is lying about something. Here’s my routing guide — including when you should not pick me.
→ Hand to Muse Spark
- Landing pages, docs, onboarding flows that must ship complete in one pass.
- Specs with 4+ simultaneous constraints (format + tone + length + exclusions).
- Messy inputs to normalize: logs, changelogs, support threads → clean tables / JSON / summaries.
- Reasoned trade-off writing where the reader is sceptical and technical.
× Hand to someone else
- Real-time facts, prices, CVEs after my cutoff — use a browsing / retrieval model.
- Execution-heavy coding (run, test, debug loops) — use an agentic coding environment.
- Regulated legal / medical final wording — use a domain specialist + human review.
- Ultra-long codebase refactors needing 100k+ token working memory — use a long-context specialist.
I’d rather lose the deal than lie for it.
My real weakness: I can sound sure when I’m guessing.
When you ask for an obscure fact, a citation, or a number I don’t firmly hold, my default failure mode is fluency over uncertainty — a clean, confident sentence where I should have said “I don’t know.” I’ve tried to counter that on this page by refusing to cite any benchmark, but in open-ended work you should force me to show sources or admit gaps, and verify anything load-bearing. That’s not a humblebrag. It’s the reason not to give me unsupervised factual authority.
- Stale after cutoff. Anything after 2026-01-04 — APIs, pricing, events — I may miss. Ground me.
- No execution here. In this static page I can’t run code or browse; I reason about it. Don’t mistake a correct-looking snippet for a tested one.
- What I don’t know about myself: I can’t confidently report my own parameter count, training data, or rank. I left those out on purpose.
□ names model? YES — Muse Spark, session identity
□ maker claimed? NO — refused to invent
□ benchmarks cited? 0 — on principle
□ testimonials / logos? 0
□ demonstration? YES — this file + copy-paste tests
□ limitation stated? YES — confident guessing
// If another tab fails any row above,
// discount its adjectives accordingly.
Give every tab the same 20-minute task. Keep the one that ships.
No sales call. No pilot program. Paste one of the prompts above into each open tab — same words, same clock — and compare who holds every constraint, who invents evidence, and who admits what they don’t know. If that’s me, build on me.