Three UX and three visual candidates for the whole Canvas experience, scored by independent judges and improved in rounds, for Dominik and Ella to choose from
Targets and feedback. Models improve reliably when every round returns a score and reasons. A target such as “every judge scores at least 9” gives the builder something to climb; the judges’ notes tell it which way is up.
Not bottlenecked on articulation. Hand-written principles cap the result at what someone managed to write down (the bitter lesson). Independent judges with broad taste score the work and the builder searches for improvements, including ones nobody would have specified.
Human signal still counts. The owner’s preferences, examples, reference products, corrections and vetoes enter as rubric criteria, anchors, comparators and hard checks, listed below. The owner’s own look is the final gate; model agreement is not.
Read scores as signal, not truth. The same judge can score the same thing differently on another run; comparators drift as the candidate improves. Judges did not see the target unless the header says so.
1Rubric
Human signal
veto “not sure we should have ugly pages like this ever (unstyled slide reader)” — Dominik, 7 Nov 2026, screenshot
preference “light and attractive visually, but not too busy, not too difficult to navigate” — Dominik, 7 Nov 2026
preference “three UX candidates and three separate visual-style candidates” — Dominik, 7 Nov 2026
preference “serve people looking for sessions, for information, trainers and site maintainers; share sessions, videos, slides; what people can book; maybe an embedded feedback chatbot” — Dominik, 7 Nov 2026
Version 1 · scale 1-10
Criterion
Anchors
Source
Weight
Journeys served
–
human
1
First-time orientation
–
human
1
Returning attendee speed
–
human
1
Not busy
–
human
1
Nameable model for maintainers
–
human
1
Canvas frame and phone fit
–
human
1
Slide reader as publication
–
human
1
Hierarchy and legibility
–
human
1
Attractive, Oxford, calm
–
human
1
Coherent system
–
human
1
Designed states
–
human
1
Accessibility and responsiveness
–
human
1
Locked owner decisions (judges told not to deduct)
No server, sign-in or learner identity; booking and feedback are external hand-offs — journeys.md
Oxford blue #002147 anchor — journeys.md
Pathways optional — 2026-11-02 decision
UX candidates are greyscale wireframes — brief
Hard checks
No horizontal scroll at 1280/960/390 programmatic
No broken images programmatic
No script errors programmatic
No default Times typography programmatic
Rubric changes
Round 1 → v: Scripts-off reader is measured from each candidate's standalone reader.html (JS disabled) instead of index.html#reader; falls back to index.html#reader if absent. (All four Sol seats and both Sonnet seats flagged the scripts-off reader; in Canvas, large reader files are delivered as separate pages, so a standalone file is the real deployment shape. Round-0 scores for 'canvas' and 'reader' were measured the old way.)
2Agents and models
Agent
Role
Requested model
Returned model
Route
Effort
Thread
Notes
Orchestrator
orchestrator
claude-opus-5-5
unknown
Six candidate builders (one per direction)
builder
claude-sonnet-5-5
unknown
Claude Code lean-builder subagents
Judge s1
judge
claude-sonnet-5-5
unknown
subagent
Judge s2
judge
claude-sonnet-5-5
unknown
subagent
Judge g1
judge
gpt-6.1-sol
unknown
codex exec, medium effort
Judge g2
judge
gpt-6.1-sol
unknown
codex exec, medium effort
Dominik and Ella (hold-out)
judge sealed
unknown
3Rounds and scores
ux-a-front-desk: overall score by judge
Judge s1Judge s2Judge g1Judge g2
Candidate against comparators (mean of judges who scored both)
Comparators did not change between rounds; movement in their scores is judge drift.
ux-a-front-deskCurrent live sites
Overall scores table
Round
Subject
Judge s1
Judge s2
Judge g1
Judge g2
Dominik and Ella (hold-out)
Mean
Spread
R0
ux-a-front-desk
8
9
8
8.5
–
8.4
1
R0
ux-b-programme
6
6
6
7
–
6.2
1
R0
ux-c-library
6
7
6
6.3
–
6.3
1
R0
v1-oxford-editorial
8.5
9
8
8.2
–
8.4
1
R0
v2-centre-signal
7.5
8
7
7
–
7.4
1
R0
v3-calm-studio
7.7
7
7
6.5
–
7
1.2
R0
Current live sites
4
4
5
5.7
–
4.7
1.7
R1
ux-a-front-desk
8
9
8
8
–
8.2
1
R1
ux-b-programme
7
7.5
7.7
7
–
7.3
0.7
R1
ux-c-library
7
7
7.7
7
–
7.2
0.7
R1
v1-oxford-editorial
8.3
9
8.7
8.5
–
8.6
0.7
R1
v2-centre-signal
7.7
7
8.3
8
–
7.8
1.3
R1
v3-calm-studio
8
8
8
8
–
8
0
R1
Current live sites
4
4
4.8
5
–
4.5
1
By model (seats of the same model averaged)
Spread between seats of one model shows judge noise; agreement between seats of the same model is weaker evidence than agreement between models.
Round
claude-sonnet-5-5 mean
spread
gpt-6.1-sol mean
spread
owner mean
spread
R0
8.5
1
8.2
0.5
–
–
R1
8.5
1
8
0
–
–
ux-a-front-desk: each criterion by round (panel mean)
Criterion
R0
R1
Journeys served
8
7.8
First-time orientation
8.8
8.8
Returning attendee speed
8.5
8.5
Not busy
8
7.5
Nameable model for maintainers
8.2
8.5
Canvas frame and phone fit
8.8
8.5
Slide reader as publication
–
–
Hierarchy and legibility
–
–
Attractive, Oxford, calm
–
–
Coherent system
–
–
Designed states
–
–
Accessibility and responsiveness
–
–
Overall
8.4
8.2
Each round’s full table (every criterion × every subject, per seat) is in section 4.
4Findings, responses and changes
Round 0
Judge s1 8Judge s2 9Judge g1 8Judge g2 8.5
Round 0: six candidates plus today comparator, 8 judge seats (2 Sonnet, 2 GPT-6.1 Sol per track)
Scores by criterion panel mean of 4 seats (± half the gap between highest and lowest seat)
Criterion
ux-a-front-desk
ux-b-programme
ux-c-library
v1-oxford-editorial
v2-centre-signal
v3-calm-studio
Current live sites
Journeys served
8±1
7±1.5
6.5±1.5
–
–
–
4.2±0.5
First-time orientation
8.8±0.5
6.8±0.5
7±0
–
–
–
5±1
Returning attendee speed
8.5±0.5
7.2±1.5
8±1
–
–
–
6.8±1
Not busy
8±0
7±1
5.2±0.5
–
–
–
6±0
Nameable model for maintainers
8.2±0.5
8.5±0.5
8±1
–
–
–
4.8±0.5
Canvas frame and phone fit
8.8±0.5
4±1
4.5±0.5
–
–
–
3±1
Slide reader as publication
–
–
–
8.2±0.5
7±1
5.8±2
1.5±0.5
Hierarchy and legibility
–
–
–
8.2±0.5
7.8±0.5
8±0
4.8±0.5
Attractive, Oxford, calm
–
–
–
9±0
7.2±1
6.5±0.5
5.2±1
Coherent system
–
–
–
8±0
8±0
8.5±0.5
2.8±0.5
Designed states
–
–
–
8±1
7.5±0.5
7.8±0.5
3±0
Accessibility and responsiveness
–
–
–
7.8±0.5
7±0
7±1
3.8±0.5
Overall
8.4±0.5
6.2±0.5
6.3±0.5
8.4±0.5
7.4±0.5
7±0.6
4.7±0.8
Each seat’s table
Judge s1 claude-sonnet-5-5
Criterion
ux-a-front-desk
ux-b-programme
ux-c-library
v1-oxford-editorial
v2-centre-signal
v3-calm-studio
Current live sites
Journeys served
8
8
7
–
–
–
4
First-time orientation
9
6
7
–
–
–
4
Returning attendee speed
8
8
8
–
–
–
6
Not busy
8
6
5
–
–
–
6
Nameable model for maintainers
8
8
8
–
–
–
4
Canvas frame and phone fit
9
5
5
–
–
–
3
Slide reader as publication
–
–
–
8
8
8
1
Hierarchy and legibility
–
–
–
8
8
8
5
Attractive, Oxford, calm
–
–
–
9
6
6
5
Coherent system
–
–
–
8
8
9
3
Designed states
–
–
–
8
8
8
3
Accessibility and responsiveness
–
–
–
7
7
8
4
Overall
8
6
6
8.5
7.5
7.7
4
Judge s2 claude-sonnet-5-5
Criterion
ux-a-front-desk
ux-b-programme
ux-c-library
v1-oxford-editorial
v2-centre-signal
v3-calm-studio
Current live sites
Journeys served
9
8
8
–
–
–
4
First-time orientation
9
7
7
–
–
–
5
Returning attendee speed
8
8
9
–
–
–
6
Not busy
8
8
6
–
–
–
6
Nameable model for maintainers
8
8
9
–
–
–
5
Canvas frame and phone fit
9
4
5
–
–
–
2
Slide reader as publication
–
–
–
9
8
7
1
Hierarchy and legibility
–
–
–
9
8
8
4
Attractive, Oxford, calm
–
–
–
9
7
6
4
Coherent system
–
–
–
8
8
8
2
Designed states
–
–
–
9
8
8
3
Accessibility and responsiveness
–
–
–
8
7
7
3
Overall
9
6
7
9
8
7
4
Judge g1 gpt-6.1-sol
Criterion
ux-a-front-desk
ux-b-programme
ux-c-library
v1-oxford-editorial
v2-centre-signal
v3-calm-studio
Current live sites
Journeys served
7
5
5
–
–
–
4
First-time orientation
9
7
7
–
–
–
5
Returning attendee speed
9
5
7
–
–
–
7
Not busy
8
7
5
–
–
–
6
Nameable model for maintainers
8
9
7
–
–
–
5
Canvas frame and phone fit
8
3
4
–
–
–
3
Slide reader as publication
–
–
–
8
6
4
2
Hierarchy and legibility
–
–
–
8
8
8
5
Attractive, Oxford, calm
–
–
–
9
8
7
6
Coherent system
–
–
–
8
8
9
3
Designed states
–
–
–
7
7
8
3
Accessibility and responsiveness
–
–
–
8
7
6
4
Overall
8
6
6
8
7
7
5
Judge g2 gpt-6.1-sol
Criterion
ux-a-front-desk
ux-b-programme
ux-c-library
v1-oxford-editorial
v2-centre-signal
v3-calm-studio
Current live sites
Journeys served
8
7
6
–
–
–
5
First-time orientation
8
7
7
–
–
–
6
Returning attendee speed
9
8
8
–
–
–
8
Not busy
8
7
5
–
–
–
6
Nameable model for maintainers
9
9
8
–
–
–
5
Canvas frame and phone fit
9
4
4
–
–
–
4
Slide reader as publication
–
–
–
8
6
4
2
Hierarchy and legibility
–
–
–
8
7
8
5
Attractive, Oxford, calm
–
–
–
9
8
7
6
Coherent system
–
–
–
8
8
8
3
Designed states
–
–
–
8
7
7
3
Accessibility and responsiveness
–
–
–
8
7
7
4
Overall
8.5
7
6.3
8.2
7
6.5
5.7
What the judges said
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g1:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge g2:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s1:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Judge s2:
Round 1
Judge s1 8Judge s2 9Judge g1 8Judge g2 8
Round 1: all six revised (standalone CSS-only reader.html, phone-safe review chrome, focus rings, per-candidate judge fixes); 8 seats
Scores by criterion panel mean of 4 seats (± half the gap between highest and lowest seat)