Evidence earns attention.
Show the trace, comparison, result, or system state.
Understudy design language · v2.0
Make the work visible.
Understudy Optimizes the Complete LLM Production Route
01 house assistantidentity = shape · activity = breath
actorunderstudy
delegationorbit active
statelearning
01 · principles
Every choice explains the system.
The current website supplies the restraint. The native mood board supplies the living model language. The system keeps both.
Color has a job.
Color identifies an actor, operation, or outcome.
Motion explains work.
Breath, orbit, transfer, and promotion carry meaning.
Structure stays quiet.
Hairlines and spacing create hierarchy before panels do.
Complexity stays inspectable.
Name every actor, threshold, and result.
02 · model color
A small palette with exact jobs.
clay
#D97757
Closed model family.
mint
#9EDBD3
Candidate and primary operation.
amber
#F2B34C
Data-taking and supervised training.
violet
#A78BFA
Search, lineage, and critic.
cyan
#67E8F9
Parallel work and trace flow.
promoted
#6EE7A0
An earned result.
03 · visual grammar
Shape identifies the actor.
Open models expose a circular core. Closed models seal their structure into tiles. Both use the same state and motion system.
03A open modelvisible core
03B closed modelAnthropic
03C closed modelOpenAI
04 · operations
Mint and cyan carry the work.
These patterns come from 04b and 04e. RLM fans one mind into smaller quests. Training loops return only the evidence that improves the policy.
04B RLM · recursivesame compute
one 31b
fifteen 2b · map, reduce
04E RL · policy learningGRPO · group mean
candidate policy
samplegroup-relative signal
six outcomes · ranked together
group mean
additional operations
Use the phase-card shell to compare operational ideas. Select one to inspect the full pattern.
human calibration
Does this answer follow the rubric?
Reject left. Accept right. The next candidate is revealed from the stack.
0 accepted · 0 rejected05 · typography
Mono for structure. Sans for reading.
The current website's mono register stays. Product prose gets a companion built for sustained reading.
Test. Learn.
Deploy. Repeat.
Deploy. Repeat.
IBM Plex Mono carries identity, controls, model labels, data, and short declarative headlines.
Production traces show which model actually works for your workload.
IBM Plex Sans carries explanation, documentation, and dense product prose.
06 · components
Quiet structure. Exact emphasis.
actions and state
evidence before claim
held-out evaluation · run 21+13%
incumbent0.357
candidate v50.409
07 · content design
Say the claim.Show the evidence.Name the action.
VOICE
Use the data you have. Build the model you need.
Direct
Make one clear point per statement. Do not repeat it in clauses.
Earned confidence
Use results and examples that a reader can inspect.
Data-first
Name the metric, baseline, evaluation set, and result.
Unapologetic
Challenge the default. Do not attack the people who chose it.
weak
Unlock powerful optimization capabilities across your AI ecosystem.
understudy
Compare every model on your workload. Deploy the one that wins.
08 · surface profiles
One language. Four registers.
Model theater
Black field. Bright actors. Meaningful motion.
Warm technical
Editorial spacing. Paper, ink, stamp, and mint evidence.
Operational
Surface tiers, native reading type, exact model state.
Readable evidence
LMAB, reports, documents, and long-form proof.