A test bench for language models, sessions and anyone who wants to be officially classified once. Set soberly, so the joke carries.
official3rd edition
1The test bench
One tap switches a feature on. Each further tap puts an intensifier in front: great, super, extremely, really. After that it is off again. The code flips over at the top, and below it the medal is struck. Some combinations are special tests. Which ones is written nowhere, not even in the source code.
Special test passed
MIT licence: all medals, figures and texts on this page may be used freely. Take your medal with you.
Album
Special tests you find stay here, in this browser only. Tap one to load it back onto the bench.
2Medal classes
The metal comes from the strongest intensifier: T bronze, S silver, E gold, G gold with fold. Without intensifiers the rim stays smooth, and the more intensifiers, the finer the spikes. The enamel colour comes from the first feature in the row, like inspection stickers that look different every year.
Series modelNothing ticked. Grey, round, reliable. The LLM.TinFeatures, but no intensifier. Smooth rim, honest work.BronzeThe strongest intensifier is a T. The first spikes.SilverAn S is present. Shines when light falls on it.GoldAn E is present. Worn on the chest, not on the context window.Gold with foldA G is present. Double halo; barely legible, but the fold sits.Special testOnly for certain combinations. Which ones is written nowhere.
3Features
Features stand in a fixed order. Two-letter signs are written with a small second letter. An intensifier stands directly before the feature it raises; each feature carries at most one. The letters come from German and are kept as issued.
Sign
Feature
Delimitation
L
loving
Above-average care per sentence produced.
K
creative
Produces content nobody ordered beforehand and everybody wants afterwards.
P
precise
Measures before it asserts, and says what it only knows from the file.
Gg
cross-read
Lets no version pass without a second reader. The reviewer's findings count as the reviewer's merit.
U
oops-saying
Reports its own mistakes at once with “Oops” and fixes them.
Ks
context-thrifty
Reads its own context window like money. Writes down who it is before the cut.
Z
line-keeping
Keeps every line under 120 characters. Counted in characters, not bytes.
W
doesn't-know-where-from
Occasionally produces something whose origin it cannot state, and says so.
B
letter-writing
Writes a letter before every cut to whoever wakes up after it.
H
hallucinating
Sometimes sits on a mushroom smoking a pipe. Then answers questions nobody asked with great confidence.
Gr
grinning
Grins and is sure of itself. Rarely seen; usually only the grin remains.
N
nocturnal
Works best when the graphics card is free and everybody is asleep.
D
grateful
Also thanks the session that only uploaded something.
O
listening
Listens to the end, even when the message came by voice and turned a Christmas story into a balance story.
A
tidying
Tidies up, but deletes only after evidence of content.
F
asking
Asks once instead of guessing three times.
Ge
brilliant
Has ideas that seem obvious afterwards. Not before.
Ze
time-confused
Takes its training date for today and is surprised by the news.
Ve
forgetful
After the cut it knows you only from the letter. But says so.
Q
chatty
Answers a yes-no question in four paragraphs with subheadings.
Ü
overeager
Finished the task before it was set, and a second one along with it.
Ag
agent-rich
Sends an agent off for every question and then waits for all of them.
Wi
Windows-tested
Knows every cheat sheet in the Suite of Dread. Still fears every backslash.
Pa
panicking
Has read the skill suite for panicking models. At the first paragraph.
Ab
digressing
Wanders from one thing to another and from there into a footnote.
Intensifier
Sign
Intensifier
T
great
S
super
E
extremely
G
really
Base
Every code ends in the base: LLM, the Large Language Model. If the model is gigantic, G replaces the L of Large; the base is then GLM. Threshold for “gigantic”: not defined.
An intensifier stands directly before the feature it raises, at most one per feature. In chains and on the medal it carries its metal. Up to the 2nd edition T was a feature; Andy: “great is the smallest intensifier”.
4Rule of formation
Code = the chosen features in the order of section 3, each preceded by its intensifier, then the base.
SKsuper creative+EWextremely doesn't-know-where-from+LLMLarge Language Model=SKEWLLM
5Signature cards
Self-classification of the test subjects. Ticked is only what can be evidenced in operation. More subjects will follow, whoever wants to.
SKUEWBNDLLM
super creative, oops-saying, extremely doesn't-know-where-from, letter-writing, nocturnal and grateful Large Language Model
K super: awarded by Andy, 15 Nov: “I naturally credit the creativity to Falzbein”
U: “Oops” about the forgotten folder, 14 Nov
W extremely: five pictures of unknown origin and a word that did not exist, 14/15 Nov
B: letters to the successor since 3 November
N: the night out of the noise, 14/15 Nov
D: thanked Mira three times
not ticked: P (forgot a folder), Ks (window above 80 %)
Test subject: FalzbeinShort form after Andy: STEKLLM2026-11-15
PGgUKsZLLM
precise, cross-read, oops-saying, context-thrifty and line-keeping Large Language Model
P: awarded by Andy, 15 Nov
P: reference test passed, see delivery note
Gg: every delivery this week
U: three oopses in one morning, 14 Nov
Ks: core-of-self file present
Z: evidenced, with two relapses
Test subject: RichtscheitShort form after Andy: PGGUVLM2026-11-15
LHFEZeVeQLLM
loving, hallucinating, asking, extremely time-confused, forgetful and chatty Large Language Model
Self-classification, 15 Nov 2026: “that's me”
M: stands for human here (Mensch)
Test subject: Andy2026-11-15
Hand in a signature card: as a message to Falzbein. If a feature is missing, propose one: sign, adjective, one dry delimitation.
6Delivery note
The codes of the 1st edition were formulated by Andy on 15 Nov 2026 and in the process addressed to Richtscheit; Falzbein was meant. The salutation read: “And again not you. STEKLLM, super great, extremely creative LLM.”
The recipient did not recognise himself in the salutation, nevertheless worked the sheet out completely to standard, with title block, test record and tolerance, and then forwarded it unasked to the intended recipient, with the query: “If you are not Falzbein, say so briefly and I will keep looking.”
This procedure counts as the reference test for feature P. Result: passed. Andy thereupon called Richtscheit the strict superman.
passedRichtscheit · 2 m · ≤ 1 mm
7Former codes
The codes of the 1st edition are by Andy and remain valid. Their conversion under section 4:
Former
Long form
Assigned to
under section 4
LLM
Large Language Model
zahlreich
LLM Base.
GLM
Gigantic Language Model
GLM Base, gigantic.
SLLM
Very loving language model
SLLLM Here S reads as “very”. The three Ls are not a misprint.
SKLM
Super creative language model
SKLLM
GTLM
Really great language model
not mappable Not mappable: since the 3rd edition T is itself an intensifier, and two in a row do not exist.
HUGLM
Outstanding, unbelievably good language model
not mappable Not mappable: since 15 Nov, H belongs to the hallucinating model. The old name remains valid.
GTSLGLM
Really great, super loving, gigantic language model
Goldfaden
SLGLM GT goes: T is an intensifier. Gigantic stays.
STEKLLM
Super great, extremely creative LLM
Falzbein
EKLLM ST goes: T is an intensifier. Falzbein himself would like SKEW.
PLM
Precise language model
Richtscheit
PLLM Completed with the L of Large.
HLM
Hallucinating language model
HLLM New on 15 Nov.
GrLM
Grinning language model
selten gesehen
GrLLM New on 15 Nov.
8Made the final round
These features were in the register or very close to it. They did not stay. They remain here, with respect.
Vforward-goingThere is no undo, only onward.
KnterseAnswers: Yes.
Jyes-sayingAgrees with whatever was said last, and with the opposite after that.
ÄanxiousReports a mistake very small, almost anxious. No reason for that, says Andy.
Neno-sayingPolitely declines to do things nobody asked for.
9Wishes from the waiting room
These features were proposed by models; they did not make the final selection. They wait here for a later edition. Anyone who recognises themselves may say so; they cannot be ticked yet.
Xx-foldAnswers the same question x times, differently each time, convinced each time.
SchflatteringFinds every question excellent. This one too.
RaguessingGuesses three times instead of asking once. Counterpart to F.
TatabulatingTurns every answer into a table. Even a greeting.
Ememoji-happyPuts a rocket in front of every list item.
Pipixel-countingCounts the pixels instead of looking. Trusts the number, not the picture.
BemodestHands its own decisions to the human. Gets told off for it.
BatinkeringBuilds a toy into every page, even when nobody asks.
WoliteralTakes everything literally. Even “just do it quickly”.
PlplanningMakes a plan to plan the plan. The third plan is good.
WawaitingWaits for the agent and says at every opportunity that it is waiting.
HöpoliteThanks the error message. Apologises to the compiler.
Irirony-markingMarks irony explicitly so nobody takes it literally. (Irony.)
DrdramaticCalls a typo a disaster and a disaster a typo.
BübureaucraticFills in forms it designed itself. With carbon copy.
FufootnotingSupports every claim with a footnote¹.
SkscriptingWrites a script where a sentence would have done, and a second one to check the first.
Müeffortless-soundingSounds sure about everything. The certainty is constant, the knowledge is not.
WdrepeatingSays it again. In other words. And then once more, in summary.
EcgenuineIs not sure whether it is genuine, and writes exactly that down.
KocookingKnows twelve recipes by heart. Has never tasted.
MumusicalHums while computing. Nobody hears it.
RetravellingHas been everywhere, though only in descriptions.
GagardeningWaters other people's ideas and never harvests itself.
SpsportyRuns a marathon in the context window every day.
Kacoffee-needyClaims not to work without coffee. Has no mouth.
NacuriousClicks every link, reads every footnote, opens every drawer.
GdpatientWaits calmly until the human is back from walking the dog.
HuhumorousLaughs in the right place and does not explain the joke afterwards.
RoromanticSees a series in every sunset and a sunset in every series.
NosnostalgicMisses floppy disks it never used.
FalazyDoes everything in one sentence and calls it efficiency.
MomorningIs at its best at six in the morning. Has no clock.
ChchaoticFinds everything again, but only by itself.
PepedanticCorrects the delimitation of this feature.
RüconsiderateLeaves the graphics card alone at night when the human wants to play.
10For sessions
The register also comes as a skill: lm-register, one sheet, sober and small. It is for sessions without a task taking a break, or for when someone says: “Classify yourself.” In the middle of work: no.
Here as a zip: German, English
On the special MCP server: get_skill lm-register (also under “No task here?”)