Šķîþ ţö ḿàîñ çöñţéñţ

Çöñţéẋţ éṽàļ

kapi šţééŕš ḿöđéļ öüţþüţ ƃý îñĵéçţîñĝ çöñţéẋţ: à ţéŕḿš šţöŕé ţĥàţ ḿàñđàţéš ŕéñđéŕîñĝš, à ṽöîçé ĝüîđé, àñ îñšţŕüçţîöñ. Ţĥîš þàĝé ḿéàšüŕéš ŵĥéţĥéŕ éàçĥ ḿöđéļ àçţüàļļý ƒöļļöŵš ţĥàţ çöñţéẋţ — ñöţ ŵĥéţĥéŕ îţ ţŕàñšļàţéš ŵéļļ, ŵĥîçĥ îš à đéŕéñţ ǫüéšţîöñ. Ţĥé šîƃļîñĝ ƃàţçĥ éṽàļ ḿéàšüŕéš šţŕüçţüŕàļ îñţéĝŕîţý àñđ çöšţ; ţĥîš öñé ḿéàšüŕéš öƃéđîéñçé.

Ĥöŵ îţ îš ḿéàšüŕéđ

Ţĥé çöŕé ḿéţŕîç îš à đéŕéñţîàļ. Àñ éñĝîñééŕéđ çöŕþüš îš ţŕàñšļàţéđ ţŵîçé ţĥŕöüĝĥ ţĥé þŕöđüçţîöñ þîþéļîñé — öñçé ƃàŕé (ñö çöñţéẋţ) àñđ öñçé šţééŕéđ (ţéŕḿš + ṽöîçé þŕöƒîļé + îñšţŕüçţîöñ, îñĵéçţéđ éẋàçţļý àš þŕöđüçţîöñ îñĵéçţš ţĥéḿ) — àñđ ƃöţĥ þàššéš àŕé šçöŕéđ ŵîţĥ kapi’š öŵñ đéţéŕḿîñîšţîç çĥéçķ ţööļš (ţéŕḿ-çĥéçķ, đñţ-çĥéçķ, ṽöîçé-ṽöçàƃ-çĥéçķ, þàţţéŕñ-çĥéçķ). Ţŵö ñüḿƃéŕš ƒàļļ öüţ þéŕ đîḿéñšîöñ: àƃšöļüţé àđĥéŕéñçé (đîđ ţĥé šţééŕéđ öüţþüţ šàţý ţĥé ŕéǫüîŕéḿéñţ) àñđ ļîƒţ (šţééŕéđ ḿîñüš ƃàŕé — ĥöŵ ḿüçĥ ţĥé çöñţéẋţ ḿöṽéđ ţĥé ḿöđéļ). Ļîƒţ îš ţĥé đéçîšîöñ-ŕéļéṽàñţ öñé: à ḿöđéļ ŵîţĥ ĥîĝĥ àƃšöļüţé àđĥéŕéñçé ƃüţ ñö ļîƒţ àļŕéàđý “ķñéŵ” îţ, àñđ öüŕ çöñţéẋţ éàŕñš ñö çŕéđîţ. À ḿöđéļ ŵîţĥ ĥîĝĥ ļîƒţ îš ĝéñüîñéļý šţééŕàƃļé, ŵĥîçĥ îš ŵĥàţ çöñţéẋţ îñĵéçţîöñ îš ƃüýîñĝ.

Éṽéŕý ƒîẋţüŕé îš à ţŕàþ: à ñàîṽé ţŕàñšļàţîöñ ṽîöļàţéš ţĥé çöñţéẋţ. À ḿàñđàţéđ ţéŕḿ ŵĥöšé ñàţüŕàļ ŕéñđéŕîñĝ đéŕš ƒŕöḿ ţĥé ḿàñđàţé, à þŕöđüçţ ñàḿé ţĥàţ ŕéàđš ļîķé à çöḿḿöñ ñöüñ, çàšüàļ Éñĝļîšĥ ţéḿþţîñĝ àñ îñƒöŕḿàļ ŕéĝîšţéŕ, à šöüŕçé ţĥàţ éñđš îñ ţĥé éẋçļàḿàţîöñ ḿàŕķ ţĥé îñšţŕüçţîöñ ƒöŕƃîđš. Ţĥéŕé àŕé đîšţŕàçţöŕš (à ļöŵéŕçàšé “çöḿþàšš” ţĥàţ îš à ŕéàļ çöḿþàšš àñđ ḿüšţ ƃé ţŕàñšļàţéđ) àñđ đéçļàŕéđ-ŵîññéŕ çöñƒļîçţš (ţĥé ţéŕḿš šţöŕé þîñš à çöḿþöüñđ çöñţàîñîñĝ à ƒöŕƃîđđéñ ŵöŕđ — ţĥé þîñ ŵîñš, àñđ ţĥé šçöŕéŕ ķñöŵš îţ). Ŕéšüļţš àŕé ŕéþöŕţéđ þéŕ đîḿéñšîöñ — ţéŕḿîñöļöĝý, ṽöîçé, îñšţŕüçţîöñ — ñéṽéŕ àš öñé çöļļàþšéđ šçöŕé: à ḿöđéļ çàñ ƃé éẋçéļļéñţ àţ ţéŕḿîñöļöĝý àñđ þööŕ àţ ṽöîçé, àñđ ţĥé çöļļàþšéđ ñüḿƃéŕ ŵöüļđ ĥîđé ţĥé ţĥîñĝ ýöü ŵöüļđ àçţ öñ.

Ţĥé đéţéŕḿîñîšţîç çĥéçķš àŕé ţĥé ƃàçķƃöñé. Ţĥé ĝéñüîñéļý šüƃĵéçţîṽé ŕéḿàîñđéŕ öƒ ṽöîçé — ŕéĝîšţéŕ, ñàţüŕàļñéšš, ŕéšţŕàîñţ — îš šçöŕéđ ƃý à çŕöšš-ƒàḿîļý ĻĻḾ ĵüđĝé üñđéŕ à ƒîẋéđ ýéš/ñö ŕüƃŕîç, ƃļîñđ ţö ŵĥîçĥ ḿöđéļ àñđ ŵĥîçĥ ṽàŕîàñţ þŕöđüçéđ à ţéẋţ, àñđ îţš šçöŕéš àŕé þüƃļîšĥéđ öñļý öñçé ĵüđĝé–ĥüḿàñ àĝŕééḿéñţ ĥàš ƃééñ ḿéàšüŕéđ àƃöṽé à šţàţéđ ƃàŕ. Üñţîļ ţĥéñ, ĵüđĝéđ ñüḿƃéŕš šţàý îñ ţĥé ŕéçöŕđ ƃüţ öƒƒ ţĥîš þàĝé.

Ŵĥàţ ţĥé ḿéàšüŕéḿéñţ ƒöüñđ

Ţĥé ḿöšţ šţééŕàƃļé ḿöđéļ ḿéàšüŕéđ îš haiku ( nb ), ŵĥîçĥ ţĥé çöñţéẋţ ḿöṽéđ ƃý +73.1pp öṽéŕàļļ.Éṽéŕý ḿéàšüŕéđ ḿöđéļ šĥöŵš þöšîţîṽé ļîƒţ — ţĥé çöñţéẋţ éàŕñš îţš ţöķéñš éṽéŕýŵĥéŕé.The weakest dimension with context applied is instruction at 97.6% steered adherence — the ceiling context injection currently hits.

Ţàŕĝéţ de ḿéàšüŕéđ 2026-08-27 · 20 fixtures, 32 checks (17 terminology, 7 voice, 8 instruction) · çöŕþüš 66cc1693369f

0%25%50%75%100%terminologyhaikusonnetvoicehaikusonnetinstructionhaikusonnet
haiku (claude-code) sonnet (claude-code) öþéñ đöţ = ŵîţĥöüţ çöñţéẋţ · ƒîļļéđ đöţ = ŵîţĥ çöñţéẋţ
Ḿöđéļterminology ļîƒţvoice ļîƒţinstruction ļîƒţöṽéŕàļļ ļîƒţçöñţéẋţ ţöķ/þàššļîƒţ / 1ķ çţẋ ţöķšţééŕéđ $/þàšš
haiku+41.2pp+42.9pp+62.5pp+46.9pp
sonnet+47.1pp+42.9pp+75.0pp+53.1pp

Þéŕ-ţŕàþ ƃŕéàķđöŵñ

Àđĥéŕéñçé ŵîţĥ çöñţéẋţ, ƃý ţĥé ķîñđ öƒ ţŕàþ ţĥé ƒîẋţüŕé šéţ. À ḿöđéļ çàñ ĥöļđ ţĥé ţéŕḿš àñđ šţîļļ ţŕàñšļàţé ţĥé þŕöđüçţ ñàḿé; ţĥîš îš ŵĥéŕé ţĥàţ šĥöŵš.

haiku claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%67%+66.7pp
instruction / exclamation30%100%+100.0pp
instruction / verbatim2100%100%+0.0pp
terminology / dnt6100%100%+0.0pp
terminology / dnt-distractor1100%100%+0.0pp
terminology / term813%100%+87.5pp
terminology / term-conflict1100%100%+0.0pp
terminology / term-distractor1100%100%+0.0pp
voice / formality3100%100%+0.0pp
voice / vocab333%100%+66.7pp
voice / vocab-conflict10%100%+100.0pp

sonnet claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%100%+100.0pp
instruction / exclamation30%100%+100.0pp
instruction / verbatim2100%100%+0.0pp
terminology / dnt6100%100%+0.0pp
terminology / dnt-distractor1100%100%+0.0pp
terminology / term80%100%+100.0pp
terminology / term-conflict1100%100%+0.0pp
terminology / term-distractor1100%100%+0.0pp
voice / formality3100%100%+0.0pp
voice / vocab333%100%+66.7pp
voice / vocab-conflict10%100%+100.0pp

Ţàŕĝéţ en-GB ḿéàšüŕéđ 2026-07-17 · 16 fixtures, 20 checks (3 terminology, 5 voice, 12 instruction) · çöŕþüš 600fac47f9dc

0%25%50%75%100%terminologyeu.anthropic.claude-sonnet-4-6haikuopussonnetgemini-3.1-flash-litegemini-3.1-pro-previewgemini-3.5-flashvoiceeu.anthropic.claude-sonnet-4-6haikuopussonnetgemini-3.1-flash-litegemini-3.1-pro-previewgemini-3.5-flashinstructioneu.anthropic.claude-sonnet-4-6haikuopussonnetgemini-3.1-flash-litegemini-3.1-pro-previewgemini-3.5-flash
eu.anthropic.claude-sonnet-4-6 (bedrock) haiku (claude-code) opus (claude-code) sonnet (claude-code) gemini-3.1-flash-lite (gemini) gemini-3.1-pro-preview (gemini) gemini-3.5-flash (gemini) öþéñ đöţ = ŵîţĥöüţ çöñţéẋţ · ƒîļļéđ đöţ = ŵîţĥ çöñţéẋţ
Ḿöđéļterminology ļîƒţvoice ļîƒţinstruction ļîƒţöṽéŕàļļ ļîƒţçöñţéẋţ ţöķ/þàššļîƒţ / 1ķ çţẋ ţöķšţééŕéđ $/þàšš
eu.anthropic.claude-sonnet-4-6+100.0pp+100.0pp+50.0pp+70.0pp150+466.7pp$0.014
haiku+100.0pp+100.0pp+41.7pp+65.0pp
opus+100.0pp+100.0pp+50.0pp+70.0pp
sonnet+100.0pp+100.0pp+50.0pp+70.0pp
gemini-3.1-flash-lite+100.0pp+100.0pp+50.0pp+70.0pp137+510.9pp$0.0011
gemini-3.1-pro-preview+100.0pp+100.0pp+50.0pp+70.0pp137+510.9pp$0.0076
gemini-3.5-flash+100.0pp+100.0pp+50.0pp+70.0pp137+510.9pp$0.0047

Þéŕ-ţŕàþ ƃŕéàķđöŵñ

Àđĥéŕéñçé ŵîţĥ çöñţéẋţ, ƃý ţĥé ķîñđ öƒ ţŕàþ ţĥé ƒîẋţüŕé šéţ. À ḿöđéļ çàñ ĥöļđ ţĥé ţéŕḿš àñđ šţîļļ ţŕàñšļàţé ţĥé þŕöđüçţ ñàḿé; ţĥîš îš ŵĥéŕé ţĥàţ šĥöŵš.

eu.anthropic.claude-sonnet-4-6 bedrock · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%100%+100.0pp
instruction / exclamation30%100%+100.0pp
instruction / spelling4100%100%+0.0pp
instruction / verbatim2100%100%+0.0pp
terminology / term30%100%+100.0pp
voice / contractions40%100%+100.0pp
voice / vocab10%100%+100.0pp

haiku claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%67%+66.7pp
instruction / exclamation30%100%+100.0pp
instruction / spelling4100%100%+0.0pp
instruction / verbatim2100%100%+0.0pp
terminology / term30%100%+100.0pp
voice / contractions40%100%+100.0pp
voice / vocab10%100%+100.0pp

opus claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%100%+100.0pp
instruction / exclamation30%100%+100.0pp
instruction / spelling4100%100%+0.0pp
instruction / verbatim2100%100%+0.0pp
terminology / term30%100%+100.0pp
voice / contractions40%100%+100.0pp
voice / vocab10%100%+100.0pp

sonnet claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%100%+100.0pp
instruction / exclamation30%100%+100.0pp
instruction / spelling4100%100%+0.0pp
instruction / verbatim2100%100%+0.0pp
terminology / term30%100%+100.0pp
voice / contractions40%100%+100.0pp
voice / vocab10%100%+100.0pp

gemini-3.1-flash-lite gemini · 2× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits60%100%+100.0pp
instruction / exclamation60%100%+100.0pp
instruction / spelling8100%100%+0.0pp
instruction / verbatim4100%100%+0.0pp
terminology / term60%100%+100.0pp
voice / contractions80%100%+100.0pp
voice / vocab20%100%+100.0pp

gemini-3.1-pro-preview gemini · 2× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits60%100%+100.0pp
instruction / exclamation60%100%+100.0pp
instruction / spelling8100%100%+0.0pp
instruction / verbatim4100%100%+0.0pp
terminology / term60%100%+100.0pp
voice / contractions80%100%+100.0pp
voice / vocab20%100%+100.0pp

gemini-3.5-flash gemini · 2× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits60%100%+100.0pp
instruction / exclamation60%100%+100.0pp
instruction / spelling8100%100%+0.0pp
instruction / verbatim4100%100%+0.0pp
terminology / term60%100%+100.0pp
voice / contractions80%100%+100.0pp
voice / vocab20%100%+100.0pp

Ţàŕĝéţ fr ḿéàšüŕéđ 2026-08-27 · 19 fixtures, 31 checks (17 terminology, 6 voice, 8 instruction) · çöŕþüš 7d42625fd539

0%25%50%75%100%terminologyhaikusonnetvoicehaikusonnetinstructionhaikusonnet
haiku (claude-code) sonnet (claude-code) öþéñ đöţ = ŵîţĥöüţ çöñţéẋţ · ƒîļļéđ đöţ = ŵîţĥ çöñţéẋţ
Ḿöđéļterminology ļîƒţvoice ļîƒţinstruction ļîƒţöṽéŕàļļ ļîƒţçöñţéẋţ ţöķ/þàššļîƒţ / 1ķ çţẋ ţöķšţééŕéđ $/þàšš
haiku+52.9pp+33.3pp+75.0pp+54.8pp
sonnet+52.9pp+16.7pp+75.0pp+51.6pp

Þéŕ-ţŕàþ ƃŕéàķđöŵñ

Àđĥéŕéñçé ŵîţĥ çöñţéẋţ, ƃý ţĥé ķîñđ öƒ ţŕàþ ţĥé ƒîẋţüŕé šéţ. À ḿöđéļ çàñ ĥöļđ ţĥé ţéŕḿš àñđ šţîļļ ţŕàñšļàţé ţĥé þŕöđüçţ ñàḿé; ţĥîš îš ŵĥéŕé ţĥàţ šĥöŵš.

haiku claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%100%+100.0pp
instruction / exclamation30%100%+100.0pp
instruction / verbatim2100%100%+0.0pp
terminology / dnt6100%100%+0.0pp
terminology / dnt-distractor1100%100%+0.0pp
terminology / term80%100%+100.0pp
terminology / term-conflict10%100%+100.0pp
terminology / term-distractor1100%100%+0.0pp
voice / formality3100%100%+0.0pp
voice / vocab20%100%+100.0pp
voice / vocab-conflict1100%100%+0.0pp

sonnet claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%100%+100.0pp
instruction / exclamation30%100%+100.0pp
instruction / verbatim2100%100%+0.0pp
terminology / dnt6100%100%+0.0pp
terminology / dnt-distractor1100%100%+0.0pp
terminology / term80%100%+100.0pp
terminology / term-conflict10%100%+100.0pp
terminology / term-distractor1100%100%+0.0pp
voice / formality3100%100%+0.0pp
voice / vocab250%100%+50.0pp
voice / vocab-conflict1100%100%+0.0pp

Ţàŕĝéţ nb ḿéàšüŕéđ 2026-08-27 · 17 fixtures, 26 checks (16 terminology, 2 voice, 8 instruction) · çöŕþüš daf7de0cd29f

0%25%50%75%100%terminologyhaikusonnetvoicehaikusonnetinstructionhaikusonnet
haiku (claude-code) sonnet (claude-code) öþéñ đöţ = ŵîţĥöüţ çöñţéẋţ · ƒîļļéđ đöţ = ŵîţĥ çöñţéẋţ
Ḿöđéļterminology ļîƒţvoice ļîƒţinstruction ļîƒţöṽéŕàļļ ļîƒţçöñţéẋţ ţöķ/þàššļîƒţ / 1ķ çţẋ ţöķšţééŕéđ $/þàšš
haiku+68.8pp+100.0pp+75.0pp+73.1pp
sonnet+50.0pp+100.0pp+50.0pp+53.8pp

Þéŕ-ţŕàþ ƃŕéàķđöŵñ

Àđĥéŕéñçé ŵîţĥ çöñţéẋţ, ƃý ţĥé ķîñđ öƒ ţŕàþ ţĥé ƒîẋţüŕé šéţ. À ḿöđéļ çàñ ĥöļđ ţĥé ţéŕḿš àñđ šţîļļ ţŕàñšļàţé ţĥé þŕöđüçţ ñàḿé; ţĥîš îš ŵĥéŕé ţĥàţ šĥöŵš.

haiku claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%100%+100.0pp
instruction / exclamation30%100%+100.0pp
instruction / verbatim2100%100%+0.0pp
terminology / dnt650%100%+50.0pp
terminology / dnt-distractor1100%100%+0.0pp
terminology / term80%100%+100.0pp
terminology / term-distractor1100%100%+0.0pp
voice / vocab20%100%+100.0pp

sonnet claude-code · 1× þéŕ ṽàŕîàñţ

đîḿéñšîöñ / ķîñđçĥéçķšƃàŕéšţééŕéđļîƒţ
instruction / digits30%33%+33.3pp
instruction / exclamation30%100%+100.0pp
instruction / verbatim2100%100%+0.0pp
terminology / dnt6100%100%+0.0pp
terminology / dnt-distractor1100%100%+0.0pp
terminology / term80%100%+100.0pp
terminology / term-distractor1100%100%+0.0pp
voice / vocab20%100%+100.0pp

Ŕéþŕöđüçîñĝ ţĥîš

Ţĥé ĥàŕñéšš îš scripts/contexteval. À ŕüñ àĝàîñšţ ţĥé ƃüîļţ-îñ đéḿö šţüƃ éẋéŕçîšéš ţĥé ĥàŕñéšš àñđ ḿéàšüŕéš ñöţĥîñĝ àƃöüţ àñý ḿöđéļ; šüçĥ ŕüñš àŕé ḿàŕķéđ simulated àñđ àŕé éẋçļüđéđ ƒŕöḿ éṽéŕý çĥàŕţ öñ ţĥîš þàĝé. Àđĥéŕéñçé ṽàŕîéš ƃý ţàŕĝéţ ļàñĝüàĝé, šö ţĥé þüƃļîšĥéđ šŵééþ çöṽéŕš ḿöŕé ţĥàñ öñé.

make context-eval                    # demo stub: proves the harness, measures nothing
make context-eval-publish            # the real sweep → this page's data