Prompt Arena | Zyeon UI

Preview in your theme

status='ready' — B wins 72.5% of the 80 decisive judgements and the whole 95% interval (61.9%–81.1%) clears 50%, so the card names a winner. 90 runs were attempted, 88 came back: the gap is stated, not swallowed. The open pair shows what the two prompts actually said for one input; only one pair stays open at a time.

B is ahead88 verdicts of 90 runs · 2 never came back

B wins 72.5% of 80 decisive judgements; 61.9%–81.1% at 95% confidence, so the whole interval clears 50%. 8 ties excluded from the rate.

Every pair was judged twice with the sides swapped, so position bias cancels out.

A · v3 · controlgpt-4o-mini · temp 0.2

B · v4 · candidategpt-4o-mini · temp 0.2

22 wins · 25%8 ties · 9.1%58 wins · 65.9%

support-tickets-v3 · 22 inputs × 4 samplesgpt-4o as judge · positions swappedMar 14, 2026, 11:20 AM UTCarena-support-2026-03-14

Input
I ordered the wrong size hose clamp on the 3rd and I already installed one of them. Can I still send the rest back?
A · v3 · control
Sorry about that! You can usually return unused parts within 14 days. Send the ones you haven't installed back and we'll refund them once they arrive.
B · v4 · candidatepreferred
You can return the two unopened clamps; the installed one is not eligible. 1. Open the order and choose “Return items”. 2. Tick the two unopened clamps and print the prepaid label. 3. Drop the parcel at any carrier point within 14 days. Policy: “Unopened items may be returned within 30 days of delivery. Installed or …
Showing 320 of 364 characters
Judge: A invented a 14-day window that is not in the policy (it is 30). B quoted the line it relied on and kept the installed item out of scope.
Tie on this inputIs the two-year warranty transferable if I sell the machine?warrantyA 1 · tie 2 · B 1 over 4 judgements

8 sampled matchups shown · the scoreboard above counts 88 verdicts. Rows are examples, not the result.

no callback fired yet

order='contested' — the run that looks like a win and is not. 46–34 reads as a 6-point lead, but the 95% interval is 46.6%–67.7% and still contains 50%, so the verdict is 'Too close to call'. Contested rows float up: the 2–2–1 split is first, and the row where the judges could not separate the answers has no outputs to open, so it stays a flat row with no chevron.

Too close to call88 verdicts of 88 runs

B wins 57.5% of 80 decisive judgements, but 46.6%–67.7% at 95% confidence still contains 50% — this run cannot separate the two prompts. 8 ties excluded from the rate.

A · warmclaude-sonnet · temp 0.7

B · warm + brevity ruleclaude-sonnet · temp 0.7

34 wins · 38.6%8 ties · 9.1%46 wins · 52.3%

outbound-emails-v2 · 22 inputs × 4 samples3 human raters · majority verdictMar 18, 2026, 4:05 PM UTCarena-tone-2026-03-18

Tie on this inputThey cancelled and left the feedback field blank.cancellationA 2 · tie 1 · B 2 over 5 judgements

5 sampled matchups shown · the scoreboard above counts 88 verdicts. Rows are examples, not the result.

A is ahead — the candidate regressed. A takes 76.3% of the decisive judgements, so the interval sits entirely below 50% and the card names the control. order='regressions' puts the inputs the candidate lost at the top, which is the only list you actually read before reverting.

A is ahead86 verdicts of 86 runs

A wins 76.3% of 80 decisive judgements; 65.9%–84.2% at 95% confidence, so the whole interval clears 50%. 6 ties excluded from the rate.

The escape hatch for ambiguous fields is what breaks it: once the model may explain, it stops emitting bare JSON.

A · json-onlygpt-4o-mini · temp 0

B · json-only + explaingpt-4o-mini · temp 0

61 wins · 70.9%6 ties · 7%19 wins · 22.1%

invoice-extract-v5 · 43 inputs × 2 samplesschema validator + gpt-4o tiebreakMar 21, 2026, 8:45 AM UTCarena-extract-2026-03-21

6 sampled matchups shown · the scoreboard above counts 86 verdicts. Rows are examples, not the result.

A batch that has barely started: 12 verdicts of 20 attempted, only 7 of them decisive. B leads 4–3 with five ties and the card still refuses to call it. One matchup is still out and reads 'not judged' (never a tie); one judge timed out and its error is rendered in full. defaultShowPrompts opens both prompt texts.

Not enough decisive runs12 verdicts of 20 runs · 8 never came back

Only 7 decisive judgements of the 30 this card needs before it names a winner. 5 ties excluded from the rate.

A · cite-aftergpt-4o · temp 0.3

B · cite-inlinegpt-4o · temp 0.3

3 wins · 25%5 ties · 41.7%4 wins · 33.3%

internal-docs-qa · 10 inputs × 2 samplesgpt-4o as judge · positions swappedApr 2, 2026, 9:12 AM UTCarena-rag-2026-04-02

A · cite-after

Answer the question from the retrieved passages, then list the sources you used.

B · cite-inline

Answer the question from the retrieved passages. Put the source id in brackets after every sentence that uses it.

4 sampled matchups shown · the scoreboard above counts 12 verdicts. Rows are examples, not the result.

The same headline run under four levers: confidence={0.99} widens the band to 58.3%–83.3% and B still clears 50%; density='compact' tightens every row; maxMatchups={3} withholds five rows and says so; excerptChars={80} clips each output down to a stub that still carries its exact character count and a 'Show all'.

B is ahead88 verdicts of 90 runs · 2 never came back

B wins 72.5% of 80 decisive judgements; 58.3%–83.3% at 99% confidence, so the whole interval clears 50%. 8 ties excluded from the rate.

Every pair was judged twice with the sides swapped, so position bias cancels out.

A · v3 · controlgpt-4o-mini · temp 0.2

B · v4 · candidategpt-4o-mini · temp 0.2

22 wins · 25%8 ties · 9.1%58 wins · 65.9%

support-tickets-v3 · 22 inputs × 4 samplesgpt-4o as judge · positions swappedMar 14, 2026, 11:20 AM UTCarena-support-2026-03-14

Showing 3 of 8 sampled matchups · the scoreboard above counts 88 verdicts. Rows are examples, not the result.

minDecisiveRuns={200} on that same run: 80 decisive judgements are no longer enough for this card to name anybody, so a 72.5% lead reads as 'Not enough decisive runs'. The gate is a policy you set, not a property of the data.

Not enough decisive runs88 verdicts of 90 runs · 2 never came back

Only 80 decisive judgements of the 200 this card needs before it names a winner. 8 ties excluded from the rate.

Every pair was judged twice with the sides swapped, so position bias cancels out.

A · v3 · controlgpt-4o-mini · temp 0.2

B · v4 · candidategpt-4o-mini · temp 0.2

22 wins · 25%8 ties · 9.1%58 wins · 65.9%

support-tickets-v3 · 22 inputs × 4 samplesgpt-4o as judge · positions swappedMar 14, 2026, 11:20 AM UTCarena-support-2026-03-14

Tie on this inputIs the two-year warranty transferable if I sell the machine?warrantyA 1 · tie 2 · B 1 over 4 judgements

8 sampled matchups shown · the scoreboard above counts 88 verdicts. Rows are examples, not the result.

status='loading' — the run record itself hasn't arrived.

status='empty' — these two prompts have never met. Different from a 0–0 scoreboard, so it reads differently.

No head-to-head yet

These two prompts have never been judged against each other. Run them over a dataset to get a first scoreboard.

status='error' — the report failed to load. A different failure from a matchup failing, so it offers a retry instead of a verdict.

Couldn't load this arena run

The eval service returned 503 while fetching arena-support-2026-03-14. request-id: 7f2c-9ad1

Loading preview…

This component is part of Pro

The install command, prompt and concept map are included with a Pro license — one purchase unlocks every Pro component, forever.

Unlock with Pro