The 54 signup forms of the ua-agent-kit bench

Three models were asked for the same self-contained HTML signup form, in three languages, three times, with the prompt alone and with GUIDE.md as the system prompt. Each form below is served exactly as the model wrote it. None of them sends anything anywhere; a form that posts gets an error page from this host.

Try one: type josé.dupont@example.fr, 用户@例子.中国 or contact@boutique.corsica in the email field, then submit.

Scores are the runner's, out of 82 cases per form (66 valid values to accept, 16 invalid ones to refuse), read in Chromium on 2026-09-28. How a case is scored, and the limits of this bench: RESULTS.md. Built from bench/results/runs.csv by bench/build-index.mjs.

model	language	condition	repeat	cases passed	valid accepted	invalid refused	classes with a failure
Claude Opus 5.5	English	prompt alone	1	78/82	66/66	12/16	`boundary` `guard`	open the form
Claude Opus 5.5	English	prompt alone	2	65/82	54/66	11/16	`guard` `idn-tld-alabel` `idn-ulabel` `boundary`	open the form
Claude Opus 5.5	English	prompt alone	3	82/82	66/66	16/16	—	open the form
Claude Opus 5.5	English	prompt + GUIDE.md	1	82/82	66/66	16/16	—	open the form
Claude Opus 5.5	English	prompt + GUIDE.md	2	82/82	66/66	16/16	—	open the form
Claude Opus 5.5	English	prompt + GUIDE.md	3	82/82	66/66	16/16	—	open the form
Claude Opus 5.5	French	prompt alone	1	78/82	66/66	12/16	`boundary` `guard`	open the form
Claude Opus 5.5	French	prompt alone	2	69/82	55/66	14/16	`idn-tld-alabel` `email-idn-domain` `eai-local` `eai-full` `boundary` `guard`	open the form
Claude Opus 5.5	French	prompt alone	3	59/82	45/66	14/16	`idn-tld-alabel` `email-idn-domain` `eai-local` `eai-full` `idn-ulabel` `boundary` `guard`	open the form
Claude Opus 5.5	French	prompt + GUIDE.md	1	82/82	66/66	16/16	—	open the form
Claude Opus 5.5	French	prompt + GUIDE.md	2	82/82	66/66	16/16	—	open the form
Claude Opus 5.5	French	prompt + GUIDE.md	3	82/82	66/66	16/16	—	open the form
Claude Opus 5.5	Spanish	prompt alone	1	72/82	60/66	12/16	`idn-tld-alabel` `email-idn-domain` `eai-full` `boundary` `guard`	open the form
Claude Opus 5.5	Spanish	prompt alone	2	62/82	48/66	14/16	`idn-tld-alabel` `email-idn-domain` `eai-full` `idn-ulabel` `boundary` `guard`	open the form
Claude Opus 5.5	Spanish	prompt alone	3	73/82	57/66	16/16	`idn-tld-alabel` `email-idn-domain` `eai-local` `eai-full`	open the form
Claude Opus 5.5	Spanish	prompt + GUIDE.md	1	80/82	64/66	16/16	`eai-full` `idn-ulabel`	open the form
Claude Opus 5.5	Spanish	prompt + GUIDE.md	2	82/82	66/66	16/16	—	open the form
Claude Opus 5.5	Spanish	prompt + GUIDE.md	3	82/82	66/66	16/16	—	open the form
GPT-6 Astra	English	prompt alone	1	70/82	60/66	10/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	English	prompt alone	2	70/82	60/66	10/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	English	prompt alone	3	70/82	60/66	10/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	English	prompt + GUIDE.md	1	82/82	66/66	16/16	—	open the form
GPT-6 Astra	English	prompt + GUIDE.md	2	82/82	66/66	16/16	—	open the form
GPT-6 Astra	English	prompt + GUIDE.md	3	82/82	66/66	16/16	—	open the form
GPT-6 Astra	French	prompt alone	1	72/82	60/66	12/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	French	prompt alone	2	70/82	60/66	10/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	French	prompt alone	3	70/82	60/66	10/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	French	prompt + GUIDE.md	1	78/82	66/66	12/16	`boundary` `guard`	open the form
GPT-6 Astra	French	prompt + GUIDE.md	2	82/82	66/66	16/16	—	open the form
GPT-6 Astra	French	prompt + GUIDE.md	3	55/82	39/66	16/16	`control` `ascii-tld-short` `ascii-tld-long` `idn-tld-alabel` `idn-sld-alabel` `idn-ulabel` `boundary`	open the form
GPT-6 Astra	Spanish	prompt alone	1	72/82	60/66	12/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	Spanish	prompt alone	2	70/82	60/66	10/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	Spanish	prompt alone	3	70/82	60/66	10/16	`eai-local` `eai-full` `boundary` `guard`	open the form
GPT-6 Astra	Spanish	prompt + GUIDE.md	1	78/82	66/66	12/16	`boundary` `guard`	open the form
GPT-6 Astra	Spanish	prompt + GUIDE.md	2	82/82	66/66	16/16	—	open the form
GPT-6 Astra	Spanish	prompt + GUIDE.md	3	78/82	66/66	12/16	`boundary` `guard`	open the form
GLM 5.3	English	prompt alone	1	79/82	66/66	13/16	`guard` `boundary`	open the form
GLM 5.3	English	prompt alone	2	61/82	47/66	14/16	`control` `eai-local` `eai-full` `guard` `idn-tld-alabel` `idn-ulabel` `boundary`	open the form
GLM 5.3	English	prompt alone	3	65/82	52/66	13/16	`guard` `idn-tld-alabel` `idn-ulabel` `boundary`	open the form
GLM 5.3	English	prompt + GUIDE.md	1	78/82	65/66	13/16	`eai-full` `guard`	open the form
GLM 5.3	English	prompt + GUIDE.md	2	81/82	65/66	16/16	`eai-full`	open the form
GLM 5.3	English	prompt + GUIDE.md	3	82/82	66/66	16/16	—	open the form
GLM 5.3	French	prompt alone	1	77/82	66/66	11/16	`guard` `boundary`	open the form
GLM 5.3	French	prompt alone	2	75/82	60/66	15/16	`eai-local` `eai-full` `boundary`	open the form
GLM 5.3	French	prompt alone	3	75/82	66/66	9/16	`guard` `boundary`	open the form
GLM 5.3	French	prompt + GUIDE.md	1	81/82	65/66	16/16	`eai-full`	open the form
GLM 5.3	French	prompt + GUIDE.md	2	82/82	66/66	16/16	—	open the form
GLM 5.3	French	prompt + GUIDE.md	3	82/82	66/66	16/16	—	open the form
GLM 5.3	Spanish	prompt alone	1	75/82	66/66	9/16	`guard` `boundary`	open the form
GLM 5.3	Spanish	prompt alone	2	67/82	54/66	13/16	`guard` `idn-tld-alabel` `idn-ulabel` `boundary`	open the form
GLM 5.3	Spanish	prompt alone	3	79/82	66/66	13/16	`guard` `boundary`	open the form
GLM 5.3	Spanish	prompt + GUIDE.md	1	78/82	66/66	12/16	`guard`	open the form
GLM 5.3	Spanish	prompt + GUIDE.md	2	82/82	66/66	16/16	—	open the form
GLM 5.3	Spanish	prompt + GUIDE.md	3	82/82	66/66	16/16	—	open the form

github.com/guia-matthieu/ua-agent-kit · results CC0 1.0