Release v1.1.2 benchmark runtime evidence

Add fail-loud Flash/KV benchmark controls, requested-versus-observed runtime reporting, controlled JSPI/compat selection, per-tier DFlash disposition, tests, provenance, and a prebuilt static bundle.

README.md

@@ -74,12 +74,20 @@ prefill/decode throughput, graph placement, device limits, engine revisions,
 the exact runtime artifact hash, and a fixed temperature-0 prompt. The report
 exports as JSON; generated model text is deliberately excluded.
 
-`n_batch` and `n_ubatch` are available there as explicitly experimental shape
-controls. The browser result reports both requested and actual llama.cpp
-values. A pinned 1.7B WebGPU smoke proved `256/128` end to end with one graph,
-29/29 GPU layers, and zero CPU ops. Flash Attention remains locked off and KV
-remains F16/F16 on this page until candidate settings pass graph and output
-parity checks.
+`n_batch`, `n_ubatch`, Flash Attention, paired K/V cache types, and the pinned
+JSPI/compat WASM flavor are available there as explicitly experimental
+controls. The browser report separates requested and native-observed values,
+records the actually loaded artifact hashes, and fails the run when they
+disagree. A pinned local 1.7B WebGPU sweep proved Flash off/auto with F16, Q8_0,
+and Q4_0 KV: every row retained one graph, 29/29 GPU layers, zero CPU ops, and
+the same short deterministic output hash. Observed KV allocation fell from
+112 MiB for F16 to 59.5 MiB for Q8_0 and 31.5 MiB for Q4_0.
+
+A same-Chrome A/B with one warmup and three alternating measured runs per arm
+also proved explicit JSPI/compat switching. Median warm load was 2,257.7 ms for
+JSPI and 2,799.1 ms for compat; all eight outputs and graph tripwires matched.
+These are experimental single-device results, not release recommendations;
+normal chat keeps Flash off, F16/F16, and automatic WASM selection.
 
 The public prewarmed-cache baseline below is one Apple `metal-3` / headless
 Chrome sample, not a cross-device study and not a first-download claim:
@@ -117,9 +125,11 @@ stay in the browser; only model/static asset downloads contact Hugging Face.
 - Dawn native WebGPU validation: `18eb229ef5f707c1464cc581252e7603c73a3ef0`
 - custom wllama source: `912c18b75d4358c1405a64646b8dbe43a205943b`
 - custom nested llama.cpp: `00fa7cb284cbf133fc426733bd64238a3588a33e`
-- custom patch set: `62efebdb3071ba49b87c835c1b45716257fe5ea6cb7eec6ee82f6a2f33bc581c`
+- custom patch set: `ad8786a295eaeb75ee03752d7f84d30f14ea373407844d5c589f7ab46800bb4b`
+- custom ESM: `24bcef8aea8e27fb7b7e2d9e6ea94ba8ced7bfbffd09a0675821b9eb1b4a4c9f`
 - JSPI WASM: `dd71f58c75f32c677a64eeb2e56efab7dedbe7395f730bb08bc80021bd182caa`
 - compat WASM: `eca9754d0d8a490b5c7b58cb384d67e31aa8ab26dd330e4ad1460982df6dd82e`
+- compat worker: `0b667db536815fb9e5dc9ffbfbd6ba29affdda38a4ba892d5372ddb41ce9a8c8`
 
 The vendored runtime exposes the validated token-embedding WebGPU placement
 needed for single-graph Bonsai inference and fails closed when placement or the
@@ -128,9 +138,10 @@ backend tripwire disagrees. Its exact JS/WASM hashes are checked during
 `vendor/wllama-bonsai/SOURCE.json`. Third-party notices are in
 `THIRD_PARTY_NOTICES.md` and `public/licenses/`.
 
-The pinned WebGPU source implements Flash Attention with Q4_0/Q8_0 K/V, but
-those combinations are not yet validated for Bonsai browser parity. Flash
-Attention therefore remains off for the release path.
+The pinned WebGPU source implements Flash Attention with Q4_0/Q8_0 K/V. The
+benchmark lane has short deterministic-output, memory, and graph-placement
+evidence for those combinations, but long-state, 27B, and cross-device gates
+remain open. Flash Attention therefore remains off for the release path.
 
 ## Licenses and attribution
 

dist/assets/{index-Do-82RUc.css → index-C-2tm91j.css}

@@ -1 +1 @@
-:root{--lightningcss-light: ;--lightningcss-dark:initial;color-scheme:dark;--ink:#111512;--ink-raised:#191e1a;--ink-soft:#252b25;--paper:#e8dfca;--paper-deep:#d8cdb4;--paper-ghost:#e8dfca9e;--phosphor:#b9ff45;--phosphor-dim:#b9ff4529;--moss:#6f8062;--rust:#c96f43;--line-light:#e8dfca24;--line-dark:#1115122e;--display:"Bodoni 72", Didot, "Iowan Old Style", Baskerville, serif;--body:"Avenir Next Condensed", "Gill Sans", "Trebuchet MS", sans-serif;--mono:"Berkeley Mono", Iosevka, "SFMono-Regular", Consolas, monospace;font-family:var(--body);font-synthesis:none;text-rendering:optimizelegibility}*{box-sizing:border-box}html{background:var(--ink)}body{background:linear-gradient(90deg, #1115120f 1px, transparent 1px) 0 0 / 22px 22px, linear-gradient(#1115120b 1px, transparent 1px) 0 0 / 22px 22px, var(--paper);min-width:320px;min-height:100vh;color:var(--paper);margin:0}button,textarea,select{font:inherit}button{color:inherit}button:focus-visible,textarea:focus-visible,select:focus-visible,summary:focus-visible{outline:2px solid var(--phosphor);outline-offset:3px}button:disabled{cursor:not-allowed}::selection{color:var(--ink);background:var(--phosphor)}.station-frame{isolation:isolate;min-height:100dvh;animation:.58s cubic-bezier(.16,1,.3,1) both station-in}.station-frame:before{z-index:20;pointer-events:none;content:"";opacity:.14;mix-blend-mode:soft-light;background-image:url("data:image/svg+xml,%3Csvg viewBox='0 0 180 180' xmlns='http://www.w3.org/2000/svg'%3E%3Cfilter id='n'%3E%3CfeTurbulence type='fractalNoise' baseFrequency='.85' numOctaves='3' stitchTiles='stitch'/%3E%3C/filter%3E%3Crect width='100%25' height='100%25' filter='url(%23n)' opacity='.3'/%3E%3C/svg%3E");position:fixed;inset:0}.station-header{background:var(--ink);border-bottom:1px solid var(--phosphor);z-index:10;grid-template-columns:minmax(330px,1fr) auto minmax(220px,1fr);align-items:stretch;min-height:76px;display:grid;position:relative;box-shadow:0 12px 38px #11151238}.brand-block{align-items:center;gap:14px;padding:12px 24px;display:flex}.brand-block h1{font:400 clamp(25px, 2vw, 34px) / .9 var(--display);letter-spacing:-.035em;margin:0}.brand-block h1 i{color:var(--phosphor);font-weight:400}.brand-mark{background:repeating-radial-gradient(at 52% 62%,#0000 0 4px,#b9ff456b 4.5px 5px);border:1px solid #b9ff457a;border-radius:50%;flex:0 0 42px;place-items:center;width:42px;height:42px;display:grid;transform:rotate(-11deg)}.brand-mark:before,.brand-mark:after,.brand-mark span:before,.brand-mark span:after{content:"";background:var(--phosphor);transform-origin:bottom;width:1px;position:absolute}.brand-mark:before{height:25px;transform:translate(-1px,4px)rotate(0)}.brand-mark:after{height:13px;transform:translate(4px,-2px)rotate(43deg)}.brand-mark span:before{height:11px;transform:translate(-5px,-1px)rotate(-47deg)}.brand-mark span:after{height:8px;transform:translate(3px,-9px)rotate(72deg)}.eyebrow{color:#e8dfca85;font:600 9px/1 var(--mono);letter-spacing:.18em;text-transform:uppercase;margin:0 0 5px}.eyebrow.ink{color:#11151285}.local-promise{border-inline:1px solid var(--line-light);background:linear-gradient(115deg,#b9ff4514,#0000 66%);align-items:center;gap:11px;min-width:360px;padding:0 28px;display:flex}.signal-dot,.live-label i,.sampling-label i{background:var(--phosphor);width:7px;height:7px;box-shadow:0 0 0 5px var(--phosphor-dim), 0 0 17px #b9ff459e;border-radius:50%;display:inline-block}.local-promise span:last-child{gap:3px;display:grid}.local-promise strong{color:var(--phosphor);font:600 10px/1 var(--mono);letter-spacing:.13em;text-transform:uppercase}.local-promise small{color:var(--paper-ghost);font:11px/1.2 var(--body)}.header-actions{justify-self:end;align-items:center;padding-right:18px;display:flex}.station-clock{color:var(--paper-ghost);font-family:var(--mono);justify-items:end;padding:0 18px;display:grid}.station-clock small{letter-spacing:.18em;font-size:8px}.station-clock span{letter-spacing:.09em;margin-top:4px;font-size:10px}.icon-button{cursor:pointer;background:0 0;border:1px solid #e8dfca3d;border-radius:50%;place-items:center;width:38px;height:38px;transition:color .16s,border-color .16s,transform .16s;display:grid}.icon-button:hover{color:var(--phosphor);border-color:var(--phosphor);transform:rotate(-7deg)}.icon-button svg{fill:none;stroke:currentColor;stroke-width:1.4px;width:18px}.workspace-grid{grid-template-columns:clamp(240px,18vw,286px) minmax(510px,1fr) clamp(310px,23vw,370px);grid-template-areas:"models chat inspector";min-height:calc(100dvh - 76px);display:grid}.model-rail,.inspector{background:var(--ink);min-width:0}.model-rail{border-right:1px solid var(--line-light);flex-direction:column;grid-area:models;padding:24px 16px 14px;animation:.52s cubic-bezier(.16,1,.3,1) .1s both rise-in;display:flex;position:relative;overflow:hidden}.model-rail:before{z-index:0;content:"";opacity:.18;background:repeating-radial-gradient(ellipse at 51% 47%, transparent 0 7px, var(--moss) 7.5px 8.5px);border-radius:43% 57% 52% 48%;width:265px;height:265px;position:absolute;bottom:44px;left:-103px;transform:rotate(18deg)}.rail-heading,.section-heading{z-index:1;align-items:flex-start;gap:11px;display:flex;position:relative}.rail-heading>span,.section-heading>span:first-child{color:var(--phosphor);font:500 9px/1 var(--mono);border-top:1px solid var(--phosphor);padding-top:5px}.rail-heading h2,.section-heading h2{font:400 22px/1 var(--display);letter-spacing:-.025em;margin:0}.model-list{z-index:1;gap:5px;margin-top:21px;display:grid;position:relative}.model-card{border:1px solid #0000;transition:border-color .18s,background .18s}.model-card.is-active{border-color:var(--phosphor);background:var(--phosphor-dim);box-shadow:inset 3px 0 0 var(--phosphor)}.model-select{text-align:left;width:100%;color:var(--paper);cursor:pointer;background:0 0;border:0;grid-template-columns:23px 1fr auto;gap:8px;padding:12px 8px 11px;display:grid}.model-card:not(.is-active) .model-select:hover{background:#e8dfca0e}.model-index{color:#e8dfca61;font:8px/1 var(--mono);padding-top:4px}.model-copy{gap:3px;min-width:0;display:grid}.model-title-line{justify-content:space-between;align-items:center;gap:7px;display:flex}.model-title-line strong{font:500 20px/1 var(--display)}.model-copy em{color:#e8dfca75;font:italic 10px/1.2 var(--display);text-overflow:ellipsis;white-space:nowrap;overflow:hidden}.model-copy>span:last-child{color:var(--paper-ghost);margin-top:3px;font-size:10px}.availability{font:600 7px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;border:1px solid var(--line-light);padding:2px 4px}.availability-loaded,.availability-loading,.availability-cached{color:var(--phosphor);border-color:#b9ff4552}.availability-remote,.availability-limited{color:var(--rust);border-color:#c96f4366}.model-footprint{color:#e8dfca63;font:8px/1 var(--mono);white-space:nowrap;align-self:end}.text-action{width:calc(100% - 30px);color:var(--rust);font:8px/1 var(--mono);text-align:left;cursor:pointer;background:0 0;border:0;border-top:1px dashed #c96f4359;margin:-3px 8px 9px 30px;padding:7px 0}.text-action span{float:right}.text-action.is-status{cursor:default;text-overflow:ellipsis;white-space:nowrap;display:block;overflow:hidden}.model-limit-reason{color:#e8dfca94;font:8px/1.35 var(--mono);margin:-4px 8px 9px 30px;display:block}.dflash-card{z-index:1;background:#111512b8;border:1px dashed #e8dfca3b;margin-top:auto;padding:13px 12px 11px;position:relative}.dflash-card.is-available{border-color:#b9ff457a}.dflash-heading{justify-content:space-between;gap:12px;display:flex}.dflash-heading h3{font:500 14px/1 var(--display);margin:7px 0 0}.experimental-stamp{color:var(--rust);font:700 7px/1 var(--mono);letter-spacing:.12em;text-transform:uppercase;border:1px solid;padding:3px 5px 2px;display:inline-block;transform:rotate(-2deg)}.switch{cursor:pointer;background:0 0;border:1px solid #e8dfca4d;border-radius:10px;width:35px;height:18px;padding:2px}.switch span{background:#e8dfca6b;border-radius:50%;width:12px;height:12px;transition:transform .17s,background .17s;display:block}.switch[aria-checked=true]{border-color:var(--phosphor)}.switch[aria-checked=true] span{background:var(--phosphor);transform:translate(17px)}.switch:disabled{opacity:.38}.dflash-card p{color:#e8dfca80;font:9px/1.45 var(--body);margin:9px 0 5px}.dflash-contract{color:#c96f43b8;font:7px/1.3 var(--mono);text-transform:uppercase;margin-bottom:11px;display:block}.dflash-axis{border-top:1px solid #e8dfca1f;justify-content:space-between;align-items:center;display:flex}.dflash-axis span{background:#e8dfca47;width:1px;height:5px;margin-top:-1px}.rail-footnote{z-index:1;color:#e8dfca4d;font:7px/1 var(--mono);letter-spacing:.03em;justify-content:space-between;gap:8px;margin-top:12px;display:flex;position:relative}.chat-surface{min-width:0;height:calc(100dvh - 76px);min-height:calc(100dvh - 76px);color:var(--ink);background:radial-gradient(ellipse at 82% 16%, transparent 0 77px, #1115120e 78px 79px, transparent 80px 89px, #1115120b 90px 91px, transparent 92px), linear-gradient(100deg, transparent 0 67%, #11151206 67.2% 67.35%, transparent 67.5%), var(--paper);grid-area:chat;grid-template-rows:auto 19px minmax(0,1fr) auto;animation:.56s cubic-bezier(.16,1,.3,1) .16s both rise-in;display:grid;position:relative;overflow:hidden}.chat-header{border-bottom:1px solid var(--line-dark);justify-content:space-between;align-items:center;gap:20px;min-height:82px;padding:17px clamp(21px,3vw,44px) 13px;display:flex}.chat-header h2{font:500 clamp(25px, 2.4vw, 37px)/1 var(--display);letter-spacing:-.045em;margin:0;display:inline}.botanical-subtitle{color:#11151275;font:italic 12px/1 var(--display);margin-left:10px}.chat-header-meta{color:#1115128c;font:8px/1 var(--mono);text-transform:uppercase;letter-spacing:.07em;grid-template-columns:repeat(2,auto);gap:5px 13px;display:grid}.header-load-button{color:var(--ink);border:1px solid var(--ink);background:var(--phosphor);font:700 7px/1 var(--mono);letter-spacing:.06em;text-transform:uppercase;cursor:pointer;grid-column:span 2;justify-self:end;padding:6px 9px}.header-load-button:disabled{opacity:.42}.live-label{color:#3d551d;grid-column:span 2;justify-self:end;align-items:center;gap:7px;display:flex}.live-label i{background:#5d7e2e;width:5px;height:5px;box-shadow:0 0 0 3px #5a782b21}.live-label.is-dormant{color:#1115127a}.live-label.is-dormant i{background:#1115125c;box-shadow:0 0 0 3px #11151217}.runtime-error{z-index:4;color:var(--paper);border:1px solid var(--rust);background:#111512f5;justify-content:space-between;align-items:center;gap:16px;padding:11px 13px;display:flex;position:absolute;top:105px;left:clamp(21px,3vw,44px);right:clamp(21px,3vw,44px);box-shadow:4px 4px #c96f433d}.runtime-error div{gap:3px;display:grid}.runtime-error strong{color:#ff9a68;font:700 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase}.runtime-error span{font:10px/1.35 var(--mono);overflow-wrap:anywhere}.runtime-error button{font:7px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;border:1px solid #e8dfca4d;flex:none;padding:5px 8px}.runtime-error button:disabled{cursor:wait;opacity:.5}.runtime-error.download-retry button{max-width:132px;line-height:1.25}.chat-ruler{color:#11151266;font:7px/1 var(--mono);background:repeating-linear-gradient(90deg,#0000 0 14px,#11151233 14px 15px) 0 100%/100% 4px no-repeat;border-bottom:1px solid #11151217;justify-content:space-between;padding-inline:clamp(21px,3vw,44px);display:flex}.transcript{scrollbar-color:#11151247 transparent;min-height:0;padding:20px clamp(21px,4vw,64px) 32px;overflow-y:auto}.transcript-date{color:#11151273;font:8px/1 var(--mono);letter-spacing:.1em;text-transform:uppercase;align-items:center;gap:10px;margin-bottom:22px;display:flex}.transcript-date:before,.transcript-date:after{content:"";background:#11151224;flex:1;height:1px}.transcript-empty{text-align:center;color:#11151280;max-width:430px;margin:12vh auto 0}.transcript-empty>span{color:#507324;font:42px/1 var(--display);display:block}.transcript-empty h3{color:var(--ink);font:400 25px/1 var(--display);margin:8px 0 6px}.transcript-empty p{margin:0;font-size:11px;line-height:1.5}.message{max-width:690px;margin-bottom:25px;animation:.43s cubic-bezier(.16,1,.3,1) both note-in}.message>header{color:#11151278;font:8px/1 var(--mono);text-transform:uppercase;letter-spacing:.07em;align-items:center;gap:8px;margin-bottom:8px;display:flex}.message>header strong{color:var(--ink)}.message>header time{margin-left:auto}.message-ordinal{color:#4d6728;min-width:22px}.message-content{font:15px/1.55 var(--body)}.message-content>p{margin:0}.markdown-content>:first-child{margin-top:0}.markdown-content>:last-child{margin-bottom:0}.markdown-content p,.markdown-content ul,.markdown-content ol,.markdown-content blockquote{margin:0 0 10px}.markdown-content ul,.markdown-content ol{padding-left:22px}.markdown-content h3,.markdown-content h4,.markdown-content h5{font-family:var(--display);margin:16px 0 7px;font-weight:500;line-height:1.1}.markdown-content blockquote{color:#111512ad;white-space:pre-wrap;border-left:2px solid #5073248c;padding-left:12px}.markdown-content code{font:.78em/1.4 var(--mono);background:#11151214;padding:1px 4px}.markdown-content pre{max-width:100%;color:var(--paper);background:var(--ink);border-left:2px solid var(--phosphor);margin:11px 0;padding:12px;position:relative;overflow:auto}.markdown-content pre code{white-space:pre;background:0 0;padding:0}.markdown-content a{color:#3f6417;text-underline-offset:2px;text-decoration-thickness:1px}.message-user .markdown-content code{background:#e8dfca1f}.reasoning-block{color:#1115129e;background:#11151209;border:1px dashed #1115123d;margin:0 0 11px;padding:8px 10px;font-size:11px}.reasoning-block summary{color:#4d6728;font:700 7px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;cursor:pointer}.reasoning-block[open] summary{margin-bottom:8px}.generation-placeholder{color:#11151273;margin:0;font-style:italic}.message-error{color:#8d321b;font:9px/1.4 var(--mono);margin:9px 0 0}.message.has-error{border-color:var(--rust)}.message-assistant{border-left:2px solid var(--ink);padding-left:22px}.message-assistant .message-content>p:first-letter{float:left;color:#38511d;font:42px/.75 var(--display);margin:1px 6px 0 0}.message-user{max-width:min(78%,610px);color:var(--paper);background:var(--ink);clip-path:polygon(0 0,calc(100% - 10px) 0,100% 10px,100% 100%,0 100%);margin-left:auto;padding:12px 15px 13px;box-shadow:5px 5px #1115121f}.message-user>header,.message-user>header strong{color:#e8dfca8f}.message-user .message-ordinal{color:var(--phosphor)}.message>footer{color:#1115126e;font:8px/1 var(--mono);border-top:1px dashed #11151230;gap:12px;margin-top:9px;padding-top:7px;display:flex}.stream-cursor{vertical-align:-2px;background:#507324;width:7px;height:14px;margin-left:4px;animation:.86s step-end infinite cursor-blink;display:inline-block}.tool-call{background:#1115120b;border:1px solid #1115122e;margin-top:13px}.tool-call summary{cursor:pointer;grid-template-columns:auto 1fr auto;align-items:center;gap:8px;padding:8px 10px;list-style:none;display:grid}.tool-call summary::-webkit-details-marker{display:none}.tool-call code{font:9px/1.4 var(--mono)}.tool-call summary>span:last-child{color:#1115127a;font:7px/1 var(--mono);text-transform:uppercase}.tool-state,.event-pip{background:var(--moss);border-radius:50%;width:6px;height:6px}.tool-state-complete,.event-complete{background:#55772c}.tool-state-running,.event-running{background:var(--phosphor);box-shadow:0 0 0 3px #5a782b26}.tool-state-error,.event-error{background:var(--rust)}.tool-call>div{border-top:1px solid #11151221;padding:9px 10px}.tool-call>div p{color:#11151294;margin:5px 0 0;font-size:10px}.composer{border-top:1px solid var(--line-dark);-webkit-backdrop-filter:blur(9px);backdrop-filter:blur(9px);background:#d8cdb4d1;padding:10px clamp(21px,3vw,44px) 18px;position:relative;box-shadow:0 -15px 38px #1115120f}.composer:before{content:"";background:repeating-linear-gradient(90deg, var(--ink) 0 1px, transparent 1px 12px);opacity:.23;height:5px;position:absolute;top:-5px;left:0;right:0}.composer-meta{color:#11151275;font:8px/1 var(--mono);justify-content:space-between;align-items:center;gap:12px;margin-bottom:7px;display:flex}.tool-toggle{color:#1115128a;font:700 8px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;border:0;align-items:center;gap:5px;padding:0;display:flex}.tool-toggle.is-on{color:#425f1f}.regenerate-button{color:#1115128a;font:700 8px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;border:0;margin-right:auto;padding:0}.regenerate-button:disabled{cursor:default;opacity:.45}.tool-toggle svg{fill:none;stroke:currentColor;stroke-width:1.3px;width:12px}.composer-field{border:1px solid var(--ink);background:#e8dfcad4;grid-template-columns:1fr auto;min-height:76px;display:grid;position:relative;box-shadow:4px 4px #1115121a}.composer-field label{color:#11151273;font:7px/1 var(--mono);letter-spacing:.09em;text-transform:uppercase;position:absolute;top:7px;left:11px}.composer-field textarea{resize:none;min-width:0;color:var(--ink);font:14px/1.4 var(--body);background:0 0;border:0;outline:none;padding:23px 13px 9px}.composer-field textarea::placeholder{color:#11151257}.send-button{min-width:82px;color:var(--ink);border:1px solid var(--ink);background:var(--phosphor);font:700 9px/1 var(--mono);text-transform:uppercase;cursor:pointer;justify-content:center;align-items:center;gap:7px;margin:7px;transition:transform .15s,box-shadow .15s,background .15s;display:flex}.send-button:hover:not(:disabled){box-shadow:3px 3px 0 var(--ink);transform:translate(-2px,-2px)}.send-button:disabled{opacity:.34}.send-button svg{fill:none;stroke:currentColor;stroke-width:1.4px;width:15px}.send-button.stop{color:var(--paper);background:var(--ink)}.send-button.stop span{background:var(--rust);width:7px;height:7px}.inspector{border-left:1px solid var(--line-light);scrollbar-color:#e8dfca2e transparent;grid-area:inspector;height:calc(100dvh - 76px);min-height:calc(100dvh - 76px);animation:.57s cubic-bezier(.16,1,.3,1) .22s both rise-in;overflow-y:auto}.inspector-section{border-bottom:1px solid var(--line-light);padding:20px 18px}.section-heading{align-items:center;margin-bottom:17px}.section-heading.compact{margin-bottom:14px}.section-heading>div{flex:1}.section-heading h2{font-size:20px}.sampling-label,.persistence{color:#e8dfca6b;font:7px/1 var(--mono);text-transform:uppercase;letter-spacing:.06em;align-items:center;gap:6px;display:flex}.sampling-label i{width:4px;height:4px}.speed-gauge{grid-template-columns:116px 1fr;align-items:center;gap:15px;display:grid}.gauge-ring{background:radial-gradient(circle, var(--ink) 0 51%, transparent 52%), conic-gradient(var(--phosphor) 0 var(--gauge), #e8dfca1c var(--gauge) 100%);border-radius:49% 51% 47% 53%;place-items:center;width:116px;height:116px;display:grid;transform:rotate(-9deg);box-shadow:inset 0 0 0 1px #e8dfca1a,0 0 30px #b9ff450e}.gauge-ring:after{content:"";border:1px dashed #e8dfca21;border-radius:45% 55% 52% 48%;width:128px;height:128px;position:absolute}.gauge-ring>div{z-index:1;justify-items:center;display:grid;position:relative;transform:rotate(9deg)}.gauge-ring strong{color:var(--phosphor);font:400 29px/1 var(--display)}.gauge-ring small{color:var(--paper-ghost);font:7px/1 var(--mono);text-transform:uppercase;margin-top:3px}.speed-gauge dl{margin:0}.speed-gauge dl div{border-bottom:1px solid var(--line-light);font:8px/1 var(--mono);justify-content:space-between;padding:7px 0;display:flex}.speed-gauge dt{color:#e8dfca6e}.speed-gauge dd{color:var(--paper);margin:0}.device-card{background:linear-gradient(90deg,#b9ff4514,#0000);border:1px solid #b9ff453b;grid-template-columns:auto 1fr auto;align-items:center;gap:10px;margin:19px 0 14px;padding:10px;display:grid}.device-icon{border:1px solid #b9ff4540;align-items:flex-end;gap:3px;width:28px;height:28px;padding:5px;display:flex}.device-icon i{background:var(--phosphor);width:4px}.device-icon i:first-child{opacity:.45;height:7px}.device-icon i:nth-child(2){height:15px}.device-icon i:nth-child(3){opacity:.7;height:11px}.device-card>div{gap:3px;display:grid}.device-card small,.memory-grid small,.meter small,.download-item small{color:#e8dfca6b;font:7px/1 var(--mono);text-transform:uppercase}.device-card strong{font:500 11px/1 var(--mono)}.device-card>span:last-child{color:var(--phosphor);font:600 13px/1 var(--mono);justify-items:end;display:grid}.device-card>span:last-child small{color:#e8dfca5e;margin-top:3px;font-size:6px}.meter{margin-top:13px}.meter>div:first-child{font:8px/1 var(--mono);justify-content:space-between;margin-bottom:6px;display:flex}.meter-track{border:1px solid #e8dfca36;height:5px;padding:1px}.meter-track>span{background:repeating-linear-gradient(90deg, var(--phosphor) 0 5px, transparent 5px 7px);height:100%;display:block}.memory-grid{background:var(--line-light);grid-template-columns:1fr 1fr;gap:1px;margin-top:14px;display:grid}.memory-grid>div{background:var(--ink);gap:6px;padding:10px;display:grid}.memory-grid strong{font:500 12px/1 var(--mono)}.event-log{gap:0;margin:0;padding:0;list-style:none;display:grid;position:relative}.event-log:before{content:"";background:var(--line-light);width:1px;position:absolute;top:5px;bottom:5px;left:3px}.event-log li{grid-template-columns:8px 64px 1fr;align-items:start;gap:8px;padding:7px 0;display:grid}.event-log li.event-empty{color:#e8dfca5c;font:8px/1.4 var(--mono);padding:5px 0;display:block}.event-pip{z-index:1;border:1px solid var(--ink);margin-top:2px}.event-log time{color:#e8dfca57;font:7px/1.2 var(--mono)}.event-log div{gap:3px;display:grid}.event-log strong{font:500 9px/1 var(--mono)}.event-log small{color:#e8dfca6b;font-size:8px}.persistence{border:1px solid var(--line-light);padding:3px 5px}.persistence.is-persistent{color:var(--phosphor);border-color:#b9ff4547}.download-item{border:1px solid var(--line-light);grid-template-columns:1fr auto;gap:8px;margin-top:16px;padding:11px 10px 10px;display:grid;position:relative}.download-item>div:first-child{gap:5px;display:grid}.download-item strong{font:500 10px/1 var(--mono)}.download-state{width:fit-content;color:var(--rust);font:700 6px/1 var(--mono);letter-spacing:.09em;text-transform:uppercase}.state-downloading{color:var(--phosphor)}.download-percent{color:var(--phosphor);font:500 15px/1 var(--display)}.download-track{background:#e8dfca1a;grid-column:span 2;height:3px}.download-track span{background:var(--phosphor);height:100%;display:block;box-shadow:0 0 9px #b9ff4559}.artifact-close{border:1px solid var(--line-light);cursor:pointer;background:0 0;border-radius:50%;width:25px;height:25px}.artifact-empty{color:#e8dfca66;font:8px/1.5 var(--mono);margin:0}.artifact-pane{border:1px solid var(--line-light);background:#f8f5eb}.artifact-tabs{color:var(--paper);border-bottom:1px solid var(--line-light);background:var(--ink-raised);grid-template-columns:auto auto 1fr;align-items:center;display:grid}.artifact-tabs button{border:0;border-right:1px solid var(--line-light);font:7px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;padding:7px 9px}.artifact-tabs button[aria-selected=true]{color:var(--ink);background:var(--phosphor)}.artifact-tabs span{color:#e8dfca73;font:7px/1 var(--mono);text-align:right;text-overflow:ellipsis;white-space:nowrap;padding:0 8px;overflow:hidden}.artifact-pane iframe{background:#fff;border:0;width:100%;height:260px;display:block}.artifact-running{position:relative}.artifact-running>button{z-index:2;color:var(--paper);border:1px solid var(--rust);font:7px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:#111512e6;padding:5px 7px;position:absolute;top:7px;right:7px}.artifact-launch{min-height:180px;color:var(--ink);text-align:center;align-content:center;justify-items:center;gap:13px;padding:20px;display:grid}.artifact-launch p{color:#1115129e;max-width:250px;font:9px/1.5 var(--mono);margin:0}.artifact-launch button{color:var(--ink);border:1px solid var(--ink);background:var(--phosphor);font:700 7px/1 var(--mono);text-transform:uppercase;cursor:pointer;padding:8px 10px}.artifact-code{height:260px;color:var(--paper);font:8px/1.45 var(--mono);white-space:pre-wrap;background:#0d100e;margin:0;padding:10px;overflow:auto}.settings-layer{z-index:50;-webkit-backdrop-filter:blur(4px);backdrop-filter:blur(4px);background:#080b09a3;justify-content:flex-end;animation:.16s both station-in;display:flex;position:fixed;inset:0}.settings-sheet{width:min(430px,100%);min-height:100%;color:var(--paper);border-left:1px solid var(--phosphor);background:repeating-radial-gradient(ellipse at 104% 101%, transparent 0 13px, #b9ff450e 13.5px 14.5px), var(--ink);padding:30px 27px;animation:.3s cubic-bezier(.16,1,.3,1) both sheet-in;overflow-y:auto;box-shadow:-24px 0 70px #00000052}.settings-sheet>header{border-bottom:1px solid var(--line-light);justify-content:space-between;align-items:flex-start;gap:20px;margin-bottom:37px;padding-bottom:20px;display:flex}.settings-sheet h2{font:400 31px/1 var(--display);margin:0}.sheet-close{border:1px solid var(--line-light);width:35px;height:35px;font:300 22px/1 var(--body);cursor:pointer;background:0 0;border-radius:50%}.sheet-close:hover{color:var(--phosphor);border-color:var(--phosphor)}.setting-group{border-bottom:1px solid var(--line-light);grid-template-columns:24px 1fr auto;align-items:center;gap:12px;padding:15px 0;display:grid}.setting-number{color:var(--phosphor);font:8px/1 var(--mono)}.setting-group>div,.range-setting>span:first-child{gap:5px;display:grid}.setting-group small,.range-setting small{color:#e8dfca6e;font:7px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase}.setting-group strong,.range-setting strong{font:500 13px/1 var(--mono)}.setting-locked{color:#e8dfca5c;font:7px/1 var(--mono);text-transform:uppercase}.setting-select select{max-width:112px;color:var(--paper);border:1px solid var(--line-light);background:var(--ink-raised);font:8px/1 var(--mono);border-radius:0;padding:6px 7px}.range-setting{gap:13px;margin-top:29px;display:grid}.range-setting>span:first-child{justify-content:space-between;align-items:end;display:flex}.range-setting input{appearance:none;background:linear-gradient(90deg, var(--phosphor) 0 35%, #e8dfca1c 35%);border:1px solid #e8dfca3d;border-radius:0;width:100%;height:4px;margin:0}.range-setting:first-of-type input{background:var(--phosphor)}.range-setting input::-webkit-slider-thumb{appearance:none;border:2px solid var(--ink);background:var(--phosphor);width:12px;height:18px;box-shadow:0 0 0 1px var(--phosphor);cursor:grab}.range-setting input::-moz-range-thumb{border:2px solid var(--ink);background:var(--phosphor);width:10px;height:16px;box-shadow:0 0 0 1px var(--phosphor);cursor:grab;border-radius:0}.range-ticks{pointer-events:none;justify-content:space-between;margin-top:-13px;display:flex}.range-limits{color:#e8dfca59;font:7px/1 var(--mono);justify-content:space-between;display:flex}.range-limits i{font-style:normal}.system-prompt-setting{gap:9px;margin-top:29px;display:grid}.system-prompt-setting>span{color:#e8dfca6e;font:7px/1 var(--mono);letter-spacing:.05em;text-transform:uppercase;justify-content:space-between;display:flex}.system-prompt-setting textarea{resize:vertical;color:var(--paper);border:1px solid var(--line-light);background:var(--ink-raised);font:10px/1.45 var(--mono);padding:10px}.settings-actions{grid-template-columns:1fr 1fr;gap:7px;margin-top:29px;display:grid}.settings-actions button{color:var(--paper);border:1px solid var(--line-light);font:7px/1.2 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;padding:9px}.settings-actions button:hover{color:var(--phosphor);border-color:var(--phosphor)}.settings-actions .danger-action{color:#e78a62;border-color:#c96f4373;grid-column:span 2}.boot-screen{min-height:100dvh;color:var(--paper);background:var(--ink);justify-content:center;align-items:center;gap:18px;padding:24px;display:flex}.boot-screen h1{color:var(--phosphor);font:400 34px/1 var(--display);margin:0}.boot-screen p:last-child{color:var(--paper-ghost);font:10px/1.5 var(--mono)}.range-ticks i{background:#e8dfca40;width:1px;height:5px}.settings-note{background:#b9ff450f;border:1px solid #b9ff4540;align-items:flex-start;gap:13px;margin-top:42px;padding:15px;display:flex}.settings-note .signal-dot{flex:0 0 7px;margin-top:6px}.settings-note p{color:#e8dfca94;margin:0;font-size:10px;line-height:1.5}.settings-note strong{color:var(--phosphor)}@keyframes station-in{0%{opacity:0}to{opacity:1}}@keyframes rise-in{0%{opacity:0;transform:translateY(11px)}to{opacity:1;transform:translateY(0)}}@keyframes note-in{0%{opacity:0;transform:translateY(7px)}to{opacity:1;transform:translateY(0)}}@keyframes cursor-blink{0%,52%{opacity:1}53%,to{opacity:0}}@keyframes sheet-in{0%{opacity:0;transform:translate(34px)}to{opacity:1;transform:translate(0)}}@media (width<=1279px){.station-header{grid-template-columns:minmax(320px,1fr) auto auto}.local-promise{min-width:auto}.local-promise small{display:none}.workspace-grid{grid-template-columns:260px minmax(500px,1fr);grid-template-areas:"models chat""inspector inspector"}.model-rail,.chat-surface{height:calc(100dvh - 76px);min-height:calc(100dvh - 76px)}.inspector{border-top:1px solid var(--phosphor);border-left:0;grid-template-columns:repeat(2,minmax(0,1fr));height:auto;min-height:auto;display:grid;overflow:visible}.inspector-section{border-right:1px solid var(--line-light);border-bottom:0}}@media (width<=820px){.station-header{min-height:70px;display:block}.brand-block{padding:11px 70px 11px 15px}.brand-mark{flex-basis:36px;width:36px;height:36px}.local-promise{border:0;border-top:1px solid var(--line-light);min-height:31px;padding:0 16px}.header-actions{position:absolute;top:10px;right:0}.station-clock{display:none}.workspace-grid{display:block}.model-rail,.chat-surface,.inspector{height:auto;min-height:auto}.model-rail{border-right:0;border-bottom:1px solid var(--phosphor);padding:18px 15px;display:block}.model-list{scroll-snap-type:x mandatory;gap:7px;padding-bottom:7px;display:flex;overflow-x:auto}.model-card{scroll-snap-align:start;min-width:205px}.dflash-card{margin-top:13px}.rail-footnote{display:none}.chat-surface{grid-template-rows:auto 19px minmax(450px,auto) auto;display:grid;overflow:visible}.transcript{padding-inline:22px;overflow:visible}.chat-header{padding:17px 19px}.chat-header-meta span:not(.live-label){display:none}.chat-header-meta .header-load-button{display:block}.live-label{grid-column:1}.composer{z-index:5;padding-inline:14px;position:sticky;bottom:0}.inspector{border-top:1px solid var(--phosphor);display:block}.inspector-section{border-right:0;border-bottom:1px solid var(--line-light)}}@media (width<=520px){.brand-block h1{font-size:24px}.brand-block .eyebrow,.local-promise strong{font-size:7px}.chat-header{align-items:flex-end}.botanical-subtitle{margin:4px 0 0;display:block}.message-user{max-width:91%}.composer-meta>span{display:none}.composer-field{grid-template-columns:1fr;padding-bottom:50px}.send-button{height:35px;position:absolute;bottom:0;right:0}.speed-gauge{grid-template-columns:106px 1fr}.gauge-ring{width:102px;height:102px}.gauge-ring:after{width:113px;height:113px}}@media (prefers-reduced-motion:reduce){*,:before,:after{scroll-behavior:auto!important;transition-duration:.01ms!important;animation-duration:.01ms!important;animation-iteration-count:1!important}}.bench-page{min-height:100dvh;color:var(--ink);background:var(--paper)}.bench-masthead{min-height:86px;color:var(--paper);border-bottom:1px solid var(--phosphor);background:var(--ink);grid-template-columns:minmax(310px,1fr) minmax(360px,1.2fr) auto;align-items:stretch;display:grid}.bench-masthead .brand-block h1{color:var(--paper)}.bench-masthead-note{border-inline:1px solid var(--line-light);background:linear-gradient(110deg,#b9ff4517,#0000 72%);flex-direction:column;justify-content:center;gap:5px;padding:14px 28px;display:flex}.bench-masthead-note strong{color:var(--phosphor);font:700 9px/1 var(--mono);letter-spacing:.11em;text-transform:uppercase}.bench-masthead-note span{color:#e8dfca8c;max-width:520px;font:10px/1.35 var(--body)}.bench-return{color:var(--paper);font:700 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;align-items:center;gap:8px;padding:0 25px;text-decoration:none;display:flex}.bench-return:hover{color:var(--phosphor)}.bench-layout{grid-template-columns:minmax(275px,330px) minmax(0,1fr);min-height:calc(100dvh - 86px);display:grid}.bench-controls{color:var(--paper);border-right:1px solid var(--phosphor);background:repeating-radial-gradient(ellipse at -12% 102%, transparent 0 24px, #b9ff450b 25px 26px), var(--ink);padding:28px 22px}.bench-controls h2{font:400 30px/1 var(--display);letter-spacing:-.03em;margin:0}.bench-control-intro{color:#e8dfca80;margin:9px 0 22px;font-size:11px;line-height:1.45}.bench-controls form{gap:15px;display:grid}.bench-field{gap:6px;min-width:0;display:grid}.bench-field>span{color:#e8dfcac7;font:700 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase}.bench-field select,.bench-field input{width:100%;min-width:0;height:39px;color:var(--paper);background:var(--ink-raised);font:10px/1 var(--mono);border:1px solid #e8dfca38;border-radius:0;outline:0;padding:0 10px}.bench-field input::placeholder{color:#e8dfca40}.bench-field select:disabled,.bench-field input:disabled{opacity:.5}.bench-field small{color:#e8dfca59;font:8px/1.2 var(--mono)}.bench-field-pair{grid-template-columns:repeat(2,minmax(0,1fr));gap:10px;display:grid}.bench-experimental{background:#c96f430d;border:1px dashed #c96f437a;padding:11px}.bench-experimental summary{color:#e19a76;font:700 8px/1 var(--mono);letter-spacing:.07em;text-transform:uppercase;cursor:pointer}.bench-experimental>p{color:#e8dfca75;margin:10px 0;font-size:9px;line-height:1.4}.bench-locks{color:#e8dfca70;font:7px/1.35 var(--mono);gap:5px;margin:10px 0 0;display:grid}.bench-locks div{justify-content:space-between;gap:8px;display:flex}.bench-locks dt,.bench-locks dd{margin:0}.bench-locks dt{color:#e8dfcab8}.bench-gate{border-left:2px solid var(--rust);background:#c96f4314;gap:5px;padding:10px;display:grid}.bench-gate.pass{border-color:var(--phosphor);background:var(--phosphor-dim)}.bench-gate strong{color:#ff9a68;font:700 8px/1 var(--mono);letter-spacing:.06em;text-transform:uppercase}.bench-gate.pass strong{color:var(--phosphor)}.bench-gate span{color:#e8dfca80;font:8px/1.4 var(--mono);overflow-wrap:anywhere}.bench-run-actions{grid-template-columns:1fr auto;gap:8px;display:grid}.bench-run-button,.bench-stop-button,.bench-section-heading button{border:1px solid var(--ink);background:var(--phosphor);min-height:38px;font:800 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;cursor:pointer;border-radius:0;padding:0 13px}.bench-run-button:disabled{opacity:.35}.bench-stop-button{color:#ffb08a;border-color:var(--rust);background:0 0}.bench-workbench{background:linear-gradient(90deg, #1115120b 1px, transparent 1px) 0 0 / 22px 22px, linear-gradient(#1115120a 1px, transparent 1px) 0 0 / 22px 22px, var(--paper);min-width:0;padding:clamp(26px,4vw,58px)}.bench-hero{border-bottom:1px solid var(--line-dark);grid-template-columns:minmax(320px,1.3fr) minmax(260px,.7fr);align-items:end;gap:34px;padding-bottom:29px;display:grid}.bench-hero h2{max-width:690px;font:400 clamp(42px, 6vw, 82px)/.84 var(--display);letter-spacing:-.065em;margin:0}.bench-hero h2 em{color:#4f7125;font-weight:400}.bench-hero>p{color:#11151294;max-width:430px;margin:0;font-size:11px;line-height:1.55}.bench-progress{gap:8px;padding:16px 0;display:grid}.bench-progress>div{font:8px/1 var(--mono);letter-spacing:.05em;text-transform:uppercase;justify-content:space-between;gap:20px;display:flex}.bench-progress>div span{color:#11151273}.bench-progress-track{background:#11151221;height:3px}.bench-progress-track i{background:#527426;height:100%;transition:width .18s;display:block}.bench-error{color:var(--paper);border-left:3px solid var(--rust);background:var(--ink);gap:5px;margin-bottom:18px;padding:12px 14px;display:grid}.bench-error strong{color:#ff9a68;font:700 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase}.bench-error span{font:9px/1.4 var(--mono);overflow-wrap:anywhere}.bench-core-section{margin-top:13px}.bench-section-heading{justify-content:space-between;align-items:end;gap:20px;margin-bottom:13px;display:flex}.bench-section-heading h3,.bench-output h3{font:500 24px/1 var(--display);margin:0}.bench-section-heading button{min-height:31px;color:var(--paper);border-color:var(--ink);background:var(--ink)}.bench-core-sample{border:1px solid var(--ink);background:var(--paper-deep);grid-template-columns:repeat(4,minmax(0,1fr));display:grid;position:relative;overflow:hidden}.bench-core-sample:before{content:"";pointer-events:none;opacity:.35;background:repeating-radial-gradient(#0000 0 34px,#1115122b 35px 36px);position:absolute;inset:-130% -5%}.bench-core-sample article{background:#e8dfca91;border-right:1px solid #11151242;flex-direction:column;justify-content:space-between;min-height:128px;padding:17px;display:flex;position:relative}.bench-core-sample article:nth-child(2n){background:#d8cdb499}.bench-core-sample article:last-child{color:var(--paper);background:#111512e8;border-right:0}.bench-core-sample article>span{font:800 8px/1 var(--mono);letter-spacing:.09em;text-transform:uppercase}.bench-core-sample article strong{font:400 clamp(28px, 3vw, 43px)/1 var(--display);letter-spacing:-.04em;margin-block:auto}.bench-core-sample article small{color:#11151280;font:7px/1.2 var(--mono);text-transform:uppercase}.bench-core-sample article:last-child small{color:#e8dfca7a}.bench-detail-grid{grid-template-columns:repeat(2,minmax(0,1fr));gap:12px;margin-top:12px;display:grid}.bench-data-card{border:1px solid var(--line-dark);background:#e8dfcaad;padding:16px}.bench-runtime-card{grid-column:span 2}.bench-data-card dl{grid-template-columns:repeat(2,minmax(0,1fr));gap:12px 26px;margin:0;display:grid}.bench-runtime-card dl{grid-template-columns:minmax(180px,1.5fr) repeat(3,minmax(100px,.7fr))}.bench-data-card dl div{border-top:1px solid #11151221;gap:4px;min-width:0;padding-top:8px;display:grid}.bench-data-card dt{color:#11151273;font:7px/1 var(--mono);letter-spacing:.07em;text-transform:uppercase}.bench-data-card dd{min-width:0;font:500 16px/1.2 var(--body);text-overflow:ellipsis;white-space:nowrap;margin:0;overflow:hidden}.bench-output{border:1px dashed #1115124d;grid-template-columns:repeat(2,minmax(0,1fr));gap:34px;margin-top:12px;padding:21px;display:grid}.bench-output>div:last-child{border-left:1px solid var(--line-dark);padding-left:34px}.bench-output p:not(.eyebrow){color:#11151299;white-space:pre-wrap;margin:8px 0 0;font-size:11px;line-height:1.5}@media (width<=1050px){.bench-masthead{grid-template-columns:minmax(310px,1fr) auto}.bench-masthead-note{display:none}.bench-layout{grid-template-columns:285px minmax(0,1fr)}.bench-core-sample{grid-template-columns:repeat(2,minmax(0,1fr))}.bench-core-sample article:nth-child(2){border-right:0}.bench-core-sample article:nth-child(-n+2){border-bottom:1px solid #11151242}.bench-runtime-card dl{grid-template-columns:repeat(2,minmax(0,1fr))}}@media (width<=760px){.bench-masthead{align-items:center;display:flex}.bench-masthead .brand-block{flex:1;min-width:0}.bench-return{padding:0 15px;font-size:0}.bench-return span{font-size:17px}.bench-layout{display:block}.bench-controls{border-right:0;border-bottom:1px solid var(--phosphor)}.bench-workbench{padding:28px 16px}.bench-hero,.bench-output{grid-template-columns:1fr;gap:18px}.bench-hero h2{font-size:clamp(42px,14vw,68px)}.bench-detail-grid{display:block}.bench-data-card+.bench-data-card{margin-top:10px}.bench-output>div:last-child{border-top:1px solid var(--line-dark);border-left:0;padding-top:18px;padding-left:0}}@media (width<=480px){.bench-masthead .brand-mark{display:none}.bench-core-sample,.bench-data-card dl,.bench-runtime-card dl{grid-template-columns:1fr}.bench-core-sample article{border-bottom:1px solid #11151242;border-right:0;min-height:105px}.bench-core-sample article:last-child{border-bottom:0}}
+:root{--lightningcss-light: ;--lightningcss-dark:initial;color-scheme:dark;--ink:#111512;--ink-raised:#191e1a;--ink-soft:#252b25;--paper:#e8dfca;--paper-deep:#d8cdb4;--paper-ghost:#e8dfca9e;--phosphor:#b9ff45;--phosphor-dim:#b9ff4529;--moss:#6f8062;--rust:#c96f43;--line-light:#e8dfca24;--line-dark:#1115122e;--display:"Bodoni 72", Didot, "Iowan Old Style", Baskerville, serif;--body:"Avenir Next Condensed", "Gill Sans", "Trebuchet MS", sans-serif;--mono:"Berkeley Mono", Iosevka, "SFMono-Regular", Consolas, monospace;font-family:var(--body);font-synthesis:none;text-rendering:optimizelegibility}*{box-sizing:border-box}html{background:var(--ink)}body{background:linear-gradient(90deg, #1115120f 1px, transparent 1px) 0 0 / 22px 22px, linear-gradient(#1115120b 1px, transparent 1px) 0 0 / 22px 22px, var(--paper);min-width:320px;min-height:100vh;color:var(--paper);margin:0}button,textarea,select{font:inherit}button{color:inherit}button:focus-visible,textarea:focus-visible,select:focus-visible,summary:focus-visible{outline:2px solid var(--phosphor);outline-offset:3px}button:disabled{cursor:not-allowed}::selection{color:var(--ink);background:var(--phosphor)}.station-frame{isolation:isolate;min-height:100dvh;animation:.58s cubic-bezier(.16,1,.3,1) both station-in}.station-frame:before{z-index:20;pointer-events:none;content:"";opacity:.14;mix-blend-mode:soft-light;background-image:url("data:image/svg+xml,%3Csvg viewBox='0 0 180 180' xmlns='http://www.w3.org/2000/svg'%3E%3Cfilter id='n'%3E%3CfeTurbulence type='fractalNoise' baseFrequency='.85' numOctaves='3' stitchTiles='stitch'/%3E%3C/filter%3E%3Crect width='100%25' height='100%25' filter='url(%23n)' opacity='.3'/%3E%3C/svg%3E");position:fixed;inset:0}.station-header{background:var(--ink);border-bottom:1px solid var(--phosphor);z-index:10;grid-template-columns:minmax(330px,1fr) auto minmax(220px,1fr);align-items:stretch;min-height:76px;display:grid;position:relative;box-shadow:0 12px 38px #11151238}.brand-block{align-items:center;gap:14px;padding:12px 24px;display:flex}.brand-block h1{font:400 clamp(25px, 2vw, 34px) / .9 var(--display);letter-spacing:-.035em;margin:0}.brand-block h1 i{color:var(--phosphor);font-weight:400}.brand-mark{background:repeating-radial-gradient(at 52% 62%,#0000 0 4px,#b9ff456b 4.5px 5px);border:1px solid #b9ff457a;border-radius:50%;flex:0 0 42px;place-items:center;width:42px;height:42px;display:grid;transform:rotate(-11deg)}.brand-mark:before,.brand-mark:after,.brand-mark span:before,.brand-mark span:after{content:"";background:var(--phosphor);transform-origin:bottom;width:1px;position:absolute}.brand-mark:before{height:25px;transform:translate(-1px,4px)rotate(0)}.brand-mark:after{height:13px;transform:translate(4px,-2px)rotate(43deg)}.brand-mark span:before{height:11px;transform:translate(-5px,-1px)rotate(-47deg)}.brand-mark span:after{height:8px;transform:translate(3px,-9px)rotate(72deg)}.eyebrow{color:#e8dfca85;font:600 9px/1 var(--mono);letter-spacing:.18em;text-transform:uppercase;margin:0 0 5px}.eyebrow.ink{color:#11151285}.local-promise{border-inline:1px solid var(--line-light);background:linear-gradient(115deg,#b9ff4514,#0000 66%);align-items:center;gap:11px;min-width:360px;padding:0 28px;display:flex}.signal-dot,.live-label i,.sampling-label i{background:var(--phosphor);width:7px;height:7px;box-shadow:0 0 0 5px var(--phosphor-dim), 0 0 17px #b9ff459e;border-radius:50%;display:inline-block}.local-promise span:last-child{gap:3px;display:grid}.local-promise strong{color:var(--phosphor);font:600 10px/1 var(--mono);letter-spacing:.13em;text-transform:uppercase}.local-promise small{color:var(--paper-ghost);font:11px/1.2 var(--body)}.header-actions{justify-self:end;align-items:center;padding-right:18px;display:flex}.station-clock{color:var(--paper-ghost);font-family:var(--mono);justify-items:end;padding:0 18px;display:grid}.station-clock small{letter-spacing:.18em;font-size:8px}.station-clock span{letter-spacing:.09em;margin-top:4px;font-size:10px}.icon-button{cursor:pointer;background:0 0;border:1px solid #e8dfca3d;border-radius:50%;place-items:center;width:38px;height:38px;transition:color .16s,border-color .16s,transform .16s;display:grid}.icon-button:hover{color:var(--phosphor);border-color:var(--phosphor);transform:rotate(-7deg)}.icon-button svg{fill:none;stroke:currentColor;stroke-width:1.4px;width:18px}.workspace-grid{grid-template-columns:clamp(240px,18vw,286px) minmax(510px,1fr) clamp(310px,23vw,370px);grid-template-areas:"models chat inspector";min-height:calc(100dvh - 76px);display:grid}.model-rail,.inspector{background:var(--ink);min-width:0}.model-rail{border-right:1px solid var(--line-light);flex-direction:column;grid-area:models;padding:24px 16px 14px;animation:.52s cubic-bezier(.16,1,.3,1) .1s both rise-in;display:flex;position:relative;overflow:hidden}.model-rail:before{z-index:0;content:"";opacity:.18;background:repeating-radial-gradient(ellipse at 51% 47%, transparent 0 7px, var(--moss) 7.5px 8.5px);border-radius:43% 57% 52% 48%;width:265px;height:265px;position:absolute;bottom:44px;left:-103px;transform:rotate(18deg)}.rail-heading,.section-heading{z-index:1;align-items:flex-start;gap:11px;display:flex;position:relative}.rail-heading>span,.section-heading>span:first-child{color:var(--phosphor);font:500 9px/1 var(--mono);border-top:1px solid var(--phosphor);padding-top:5px}.rail-heading h2,.section-heading h2{font:400 22px/1 var(--display);letter-spacing:-.025em;margin:0}.model-list{z-index:1;gap:5px;margin-top:21px;display:grid;position:relative}.model-card{border:1px solid #0000;transition:border-color .18s,background .18s}.model-card.is-active{border-color:var(--phosphor);background:var(--phosphor-dim);box-shadow:inset 3px 0 0 var(--phosphor)}.model-select{text-align:left;width:100%;color:var(--paper);cursor:pointer;background:0 0;border:0;grid-template-columns:23px 1fr auto;gap:8px;padding:12px 8px 11px;display:grid}.model-card:not(.is-active) .model-select:hover{background:#e8dfca0e}.model-index{color:#e8dfca61;font:8px/1 var(--mono);padding-top:4px}.model-copy{gap:3px;min-width:0;display:grid}.model-title-line{justify-content:space-between;align-items:center;gap:7px;display:flex}.model-title-line strong{font:500 20px/1 var(--display)}.model-copy em{color:#e8dfca75;font:italic 10px/1.2 var(--display);text-overflow:ellipsis;white-space:nowrap;overflow:hidden}.model-copy>span:last-child{color:var(--paper-ghost);margin-top:3px;font-size:10px}.availability{font:600 7px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;border:1px solid var(--line-light);padding:2px 4px}.availability-loaded,.availability-loading,.availability-cached{color:var(--phosphor);border-color:#b9ff4552}.availability-remote,.availability-limited{color:var(--rust);border-color:#c96f4366}.model-footprint{color:#e8dfca63;font:8px/1 var(--mono);white-space:nowrap;align-self:end}.text-action{width:calc(100% - 30px);color:var(--rust);font:8px/1 var(--mono);text-align:left;cursor:pointer;background:0 0;border:0;border-top:1px dashed #c96f4359;margin:-3px 8px 9px 30px;padding:7px 0}.text-action span{float:right}.text-action.is-status{cursor:default;text-overflow:ellipsis;white-space:nowrap;display:block;overflow:hidden}.model-limit-reason{color:#e8dfca94;font:8px/1.35 var(--mono);margin:-4px 8px 9px 30px;display:block}.dflash-card{z-index:1;background:#111512b8;border:1px dashed #e8dfca3b;margin-top:auto;padding:13px 12px 11px;position:relative}.dflash-card.is-available{border-color:#b9ff457a}.dflash-heading{justify-content:space-between;gap:12px;display:flex}.dflash-heading h3{font:500 14px/1 var(--display);margin:7px 0 0}.experimental-stamp{color:var(--rust);font:700 7px/1 var(--mono);letter-spacing:.12em;text-transform:uppercase;border:1px solid;padding:3px 5px 2px;display:inline-block;transform:rotate(-2deg)}.switch{cursor:pointer;background:0 0;border:1px solid #e8dfca4d;border-radius:10px;width:35px;height:18px;padding:2px}.switch span{background:#e8dfca6b;border-radius:50%;width:12px;height:12px;transition:transform .17s,background .17s;display:block}.switch[aria-checked=true]{border-color:var(--phosphor)}.switch[aria-checked=true] span{background:var(--phosphor);transform:translate(17px)}.switch:disabled{opacity:.38}.dflash-card p{color:#e8dfca80;font:9px/1.45 var(--body);margin:9px 0 5px}.dflash-contract{color:#c96f43b8;font:7px/1.3 var(--mono);text-transform:uppercase;margin-bottom:11px;display:block}.dflash-axis{border-top:1px solid #e8dfca1f;justify-content:space-between;align-items:center;display:flex}.dflash-axis span{background:#e8dfca47;width:1px;height:5px;margin-top:-1px}.rail-footnote{z-index:1;color:#e8dfca4d;font:7px/1 var(--mono);letter-spacing:.03em;justify-content:space-between;gap:8px;margin-top:12px;display:flex;position:relative}.chat-surface{min-width:0;height:calc(100dvh - 76px);min-height:calc(100dvh - 76px);color:var(--ink);background:radial-gradient(ellipse at 82% 16%, transparent 0 77px, #1115120e 78px 79px, transparent 80px 89px, #1115120b 90px 91px, transparent 92px), linear-gradient(100deg, transparent 0 67%, #11151206 67.2% 67.35%, transparent 67.5%), var(--paper);grid-area:chat;grid-template-rows:auto 19px minmax(0,1fr) auto;animation:.56s cubic-bezier(.16,1,.3,1) .16s both rise-in;display:grid;position:relative;overflow:hidden}.chat-header{border-bottom:1px solid var(--line-dark);justify-content:space-between;align-items:center;gap:20px;min-height:82px;padding:17px clamp(21px,3vw,44px) 13px;display:flex}.chat-header h2{font:500 clamp(25px, 2.4vw, 37px)/1 var(--display);letter-spacing:-.045em;margin:0;display:inline}.botanical-subtitle{color:#11151275;font:italic 12px/1 var(--display);margin-left:10px}.chat-header-meta{color:#1115128c;font:8px/1 var(--mono);text-transform:uppercase;letter-spacing:.07em;grid-template-columns:repeat(2,auto);gap:5px 13px;display:grid}.header-load-button{color:var(--ink);border:1px solid var(--ink);background:var(--phosphor);font:700 7px/1 var(--mono);letter-spacing:.06em;text-transform:uppercase;cursor:pointer;grid-column:span 2;justify-self:end;padding:6px 9px}.header-load-button:disabled{opacity:.42}.live-label{color:#3d551d;grid-column:span 2;justify-self:end;align-items:center;gap:7px;display:flex}.live-label i{background:#5d7e2e;width:5px;height:5px;box-shadow:0 0 0 3px #5a782b21}.live-label.is-dormant{color:#1115127a}.live-label.is-dormant i{background:#1115125c;box-shadow:0 0 0 3px #11151217}.runtime-error{z-index:4;color:var(--paper);border:1px solid var(--rust);background:#111512f5;justify-content:space-between;align-items:center;gap:16px;padding:11px 13px;display:flex;position:absolute;top:105px;left:clamp(21px,3vw,44px);right:clamp(21px,3vw,44px);box-shadow:4px 4px #c96f433d}.runtime-error div{gap:3px;display:grid}.runtime-error strong{color:#ff9a68;font:700 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase}.runtime-error span{font:10px/1.35 var(--mono);overflow-wrap:anywhere}.runtime-error button{font:7px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;border:1px solid #e8dfca4d;flex:none;padding:5px 8px}.runtime-error button:disabled{cursor:wait;opacity:.5}.runtime-error.download-retry button{max-width:132px;line-height:1.25}.chat-ruler{color:#11151266;font:7px/1 var(--mono);background:repeating-linear-gradient(90deg,#0000 0 14px,#11151233 14px 15px) 0 100%/100% 4px no-repeat;border-bottom:1px solid #11151217;justify-content:space-between;padding-inline:clamp(21px,3vw,44px);display:flex}.transcript{scrollbar-color:#11151247 transparent;min-height:0;padding:20px clamp(21px,4vw,64px) 32px;overflow-y:auto}.transcript-date{color:#11151273;font:8px/1 var(--mono);letter-spacing:.1em;text-transform:uppercase;align-items:center;gap:10px;margin-bottom:22px;display:flex}.transcript-date:before,.transcript-date:after{content:"";background:#11151224;flex:1;height:1px}.transcript-empty{text-align:center;color:#11151280;max-width:430px;margin:12vh auto 0}.transcript-empty>span{color:#507324;font:42px/1 var(--display);display:block}.transcript-empty h3{color:var(--ink);font:400 25px/1 var(--display);margin:8px 0 6px}.transcript-empty p{margin:0;font-size:11px;line-height:1.5}.message{max-width:690px;margin-bottom:25px;animation:.43s cubic-bezier(.16,1,.3,1) both note-in}.message>header{color:#11151278;font:8px/1 var(--mono);text-transform:uppercase;letter-spacing:.07em;align-items:center;gap:8px;margin-bottom:8px;display:flex}.message>header strong{color:var(--ink)}.message>header time{margin-left:auto}.message-ordinal{color:#4d6728;min-width:22px}.message-content{font:15px/1.55 var(--body)}.message-content>p{margin:0}.markdown-content>:first-child{margin-top:0}.markdown-content>:last-child{margin-bottom:0}.markdown-content p,.markdown-content ul,.markdown-content ol,.markdown-content blockquote{margin:0 0 10px}.markdown-content ul,.markdown-content ol{padding-left:22px}.markdown-content h3,.markdown-content h4,.markdown-content h5{font-family:var(--display);margin:16px 0 7px;font-weight:500;line-height:1.1}.markdown-content blockquote{color:#111512ad;white-space:pre-wrap;border-left:2px solid #5073248c;padding-left:12px}.markdown-content code{font:.78em/1.4 var(--mono);background:#11151214;padding:1px 4px}.markdown-content pre{max-width:100%;color:var(--paper);background:var(--ink);border-left:2px solid var(--phosphor);margin:11px 0;padding:12px;position:relative;overflow:auto}.markdown-content pre code{white-space:pre;background:0 0;padding:0}.markdown-content a{color:#3f6417;text-underline-offset:2px;text-decoration-thickness:1px}.message-user .markdown-content code{background:#e8dfca1f}.reasoning-block{color:#1115129e;background:#11151209;border:1px dashed #1115123d;margin:0 0 11px;padding:8px 10px;font-size:11px}.reasoning-block summary{color:#4d6728;font:700 7px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;cursor:pointer}.reasoning-block[open] summary{margin-bottom:8px}.generation-placeholder{color:#11151273;margin:0;font-style:italic}.message-error{color:#8d321b;font:9px/1.4 var(--mono);margin:9px 0 0}.message.has-error{border-color:var(--rust)}.message-assistant{border-left:2px solid var(--ink);padding-left:22px}.message-assistant .message-content>p:first-letter{float:left;color:#38511d;font:42px/.75 var(--display);margin:1px 6px 0 0}.message-user{max-width:min(78%,610px);color:var(--paper);background:var(--ink);clip-path:polygon(0 0,calc(100% - 10px) 0,100% 10px,100% 100%,0 100%);margin-left:auto;padding:12px 15px 13px;box-shadow:5px 5px #1115121f}.message-user>header,.message-user>header strong{color:#e8dfca8f}.message-user .message-ordinal{color:var(--phosphor)}.message>footer{color:#1115126e;font:8px/1 var(--mono);border-top:1px dashed #11151230;gap:12px;margin-top:9px;padding-top:7px;display:flex}.stream-cursor{vertical-align:-2px;background:#507324;width:7px;height:14px;margin-left:4px;animation:.86s step-end infinite cursor-blink;display:inline-block}.tool-call{background:#1115120b;border:1px solid #1115122e;margin-top:13px}.tool-call summary{cursor:pointer;grid-template-columns:auto 1fr auto;align-items:center;gap:8px;padding:8px 10px;list-style:none;display:grid}.tool-call summary::-webkit-details-marker{display:none}.tool-call code{font:9px/1.4 var(--mono)}.tool-call summary>span:last-child{color:#1115127a;font:7px/1 var(--mono);text-transform:uppercase}.tool-state,.event-pip{background:var(--moss);border-radius:50%;width:6px;height:6px}.tool-state-complete,.event-complete{background:#55772c}.tool-state-running,.event-running{background:var(--phosphor);box-shadow:0 0 0 3px #5a782b26}.tool-state-error,.event-error{background:var(--rust)}.tool-call>div{border-top:1px solid #11151221;padding:9px 10px}.tool-call>div p{color:#11151294;margin:5px 0 0;font-size:10px}.composer{border-top:1px solid var(--line-dark);-webkit-backdrop-filter:blur(9px);backdrop-filter:blur(9px);background:#d8cdb4d1;padding:10px clamp(21px,3vw,44px) 18px;position:relative;box-shadow:0 -15px 38px #1115120f}.composer:before{content:"";background:repeating-linear-gradient(90deg, var(--ink) 0 1px, transparent 1px 12px);opacity:.23;height:5px;position:absolute;top:-5px;left:0;right:0}.composer-meta{color:#11151275;font:8px/1 var(--mono);justify-content:space-between;align-items:center;gap:12px;margin-bottom:7px;display:flex}.tool-toggle{color:#1115128a;font:700 8px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;border:0;align-items:center;gap:5px;padding:0;display:flex}.tool-toggle.is-on{color:#425f1f}.regenerate-button{color:#1115128a;font:700 8px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;border:0;margin-right:auto;padding:0}.regenerate-button:disabled{cursor:default;opacity:.45}.tool-toggle svg{fill:none;stroke:currentColor;stroke-width:1.3px;width:12px}.composer-field{border:1px solid var(--ink);background:#e8dfcad4;grid-template-columns:1fr auto;min-height:76px;display:grid;position:relative;box-shadow:4px 4px #1115121a}.composer-field label{color:#11151273;font:7px/1 var(--mono);letter-spacing:.09em;text-transform:uppercase;position:absolute;top:7px;left:11px}.composer-field textarea{resize:none;min-width:0;color:var(--ink);font:14px/1.4 var(--body);background:0 0;border:0;outline:none;padding:23px 13px 9px}.composer-field textarea::placeholder{color:#11151257}.send-button{min-width:82px;color:var(--ink);border:1px solid var(--ink);background:var(--phosphor);font:700 9px/1 var(--mono);text-transform:uppercase;cursor:pointer;justify-content:center;align-items:center;gap:7px;margin:7px;transition:transform .15s,box-shadow .15s,background .15s;display:flex}.send-button:hover:not(:disabled){box-shadow:3px 3px 0 var(--ink);transform:translate(-2px,-2px)}.send-button:disabled{opacity:.34}.send-button svg{fill:none;stroke:currentColor;stroke-width:1.4px;width:15px}.send-button.stop{color:var(--paper);background:var(--ink)}.send-button.stop span{background:var(--rust);width:7px;height:7px}.inspector{border-left:1px solid var(--line-light);scrollbar-color:#e8dfca2e transparent;grid-area:inspector;height:calc(100dvh - 76px);min-height:calc(100dvh - 76px);animation:.57s cubic-bezier(.16,1,.3,1) .22s both rise-in;overflow-y:auto}.inspector-section{border-bottom:1px solid var(--line-light);padding:20px 18px}.section-heading{align-items:center;margin-bottom:17px}.section-heading.compact{margin-bottom:14px}.section-heading>div{flex:1}.section-heading h2{font-size:20px}.sampling-label,.persistence{color:#e8dfca6b;font:7px/1 var(--mono);text-transform:uppercase;letter-spacing:.06em;align-items:center;gap:6px;display:flex}.sampling-label i{width:4px;height:4px}.speed-gauge{grid-template-columns:116px 1fr;align-items:center;gap:15px;display:grid}.gauge-ring{background:radial-gradient(circle, var(--ink) 0 51%, transparent 52%), conic-gradient(var(--phosphor) 0 var(--gauge), #e8dfca1c var(--gauge) 100%);border-radius:49% 51% 47% 53%;place-items:center;width:116px;height:116px;display:grid;transform:rotate(-9deg);box-shadow:inset 0 0 0 1px #e8dfca1a,0 0 30px #b9ff450e}.gauge-ring:after{content:"";border:1px dashed #e8dfca21;border-radius:45% 55% 52% 48%;width:128px;height:128px;position:absolute}.gauge-ring>div{z-index:1;justify-items:center;display:grid;position:relative;transform:rotate(9deg)}.gauge-ring strong{color:var(--phosphor);font:400 29px/1 var(--display)}.gauge-ring small{color:var(--paper-ghost);font:7px/1 var(--mono);text-transform:uppercase;margin-top:3px}.speed-gauge dl{margin:0}.speed-gauge dl div{border-bottom:1px solid var(--line-light);font:8px/1 var(--mono);justify-content:space-between;padding:7px 0;display:flex}.speed-gauge dt{color:#e8dfca6e}.speed-gauge dd{color:var(--paper);margin:0}.device-card{background:linear-gradient(90deg,#b9ff4514,#0000);border:1px solid #b9ff453b;grid-template-columns:auto 1fr auto;align-items:center;gap:10px;margin:19px 0 14px;padding:10px;display:grid}.device-icon{border:1px solid #b9ff4540;align-items:flex-end;gap:3px;width:28px;height:28px;padding:5px;display:flex}.device-icon i{background:var(--phosphor);width:4px}.device-icon i:first-child{opacity:.45;height:7px}.device-icon i:nth-child(2){height:15px}.device-icon i:nth-child(3){opacity:.7;height:11px}.device-card>div{gap:3px;display:grid}.device-card small,.memory-grid small,.meter small,.download-item small{color:#e8dfca6b;font:7px/1 var(--mono);text-transform:uppercase}.device-card strong{font:500 11px/1 var(--mono)}.device-card>span:last-child{color:var(--phosphor);font:600 13px/1 var(--mono);justify-items:end;display:grid}.device-card>span:last-child small{color:#e8dfca5e;margin-top:3px;font-size:6px}.meter{margin-top:13px}.meter>div:first-child{font:8px/1 var(--mono);justify-content:space-between;margin-bottom:6px;display:flex}.meter-track{border:1px solid #e8dfca36;height:5px;padding:1px}.meter-track>span{background:repeating-linear-gradient(90deg, var(--phosphor) 0 5px, transparent 5px 7px);height:100%;display:block}.memory-grid{background:var(--line-light);grid-template-columns:1fr 1fr;gap:1px;margin-top:14px;display:grid}.memory-grid>div{background:var(--ink);gap:6px;padding:10px;display:grid}.memory-grid strong{font:500 12px/1 var(--mono)}.event-log{gap:0;margin:0;padding:0;list-style:none;display:grid;position:relative}.event-log:before{content:"";background:var(--line-light);width:1px;position:absolute;top:5px;bottom:5px;left:3px}.event-log li{grid-template-columns:8px 64px 1fr;align-items:start;gap:8px;padding:7px 0;display:grid}.event-log li.event-empty{color:#e8dfca5c;font:8px/1.4 var(--mono);padding:5px 0;display:block}.event-pip{z-index:1;border:1px solid var(--ink);margin-top:2px}.event-log time{color:#e8dfca57;font:7px/1.2 var(--mono)}.event-log div{gap:3px;display:grid}.event-log strong{font:500 9px/1 var(--mono)}.event-log small{color:#e8dfca6b;font-size:8px}.persistence{border:1px solid var(--line-light);padding:3px 5px}.persistence.is-persistent{color:var(--phosphor);border-color:#b9ff4547}.download-item{border:1px solid var(--line-light);grid-template-columns:1fr auto;gap:8px;margin-top:16px;padding:11px 10px 10px;display:grid;position:relative}.download-item>div:first-child{gap:5px;display:grid}.download-item strong{font:500 10px/1 var(--mono)}.download-state{width:fit-content;color:var(--rust);font:700 6px/1 var(--mono);letter-spacing:.09em;text-transform:uppercase}.state-downloading{color:var(--phosphor)}.download-percent{color:var(--phosphor);font:500 15px/1 var(--display)}.download-track{background:#e8dfca1a;grid-column:span 2;height:3px}.download-track span{background:var(--phosphor);height:100%;display:block;box-shadow:0 0 9px #b9ff4559}.artifact-close{border:1px solid var(--line-light);cursor:pointer;background:0 0;border-radius:50%;width:25px;height:25px}.artifact-empty{color:#e8dfca66;font:8px/1.5 var(--mono);margin:0}.artifact-pane{border:1px solid var(--line-light);background:#f8f5eb}.artifact-tabs{color:var(--paper);border-bottom:1px solid var(--line-light);background:var(--ink-raised);grid-template-columns:auto auto 1fr;align-items:center;display:grid}.artifact-tabs button{border:0;border-right:1px solid var(--line-light);font:7px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;padding:7px 9px}.artifact-tabs button[aria-selected=true]{color:var(--ink);background:var(--phosphor)}.artifact-tabs span{color:#e8dfca73;font:7px/1 var(--mono);text-align:right;text-overflow:ellipsis;white-space:nowrap;padding:0 8px;overflow:hidden}.artifact-pane iframe{background:#fff;border:0;width:100%;height:260px;display:block}.artifact-running{position:relative}.artifact-running>button{z-index:2;color:var(--paper);border:1px solid var(--rust);font:7px/1 var(--mono);text-transform:uppercase;cursor:pointer;background:#111512e6;padding:5px 7px;position:absolute;top:7px;right:7px}.artifact-launch{min-height:180px;color:var(--ink);text-align:center;align-content:center;justify-items:center;gap:13px;padding:20px;display:grid}.artifact-launch p{color:#1115129e;max-width:250px;font:9px/1.5 var(--mono);margin:0}.artifact-launch button{color:var(--ink);border:1px solid var(--ink);background:var(--phosphor);font:700 7px/1 var(--mono);text-transform:uppercase;cursor:pointer;padding:8px 10px}.artifact-code{height:260px;color:var(--paper);font:8px/1.45 var(--mono);white-space:pre-wrap;background:#0d100e;margin:0;padding:10px;overflow:auto}.settings-layer{z-index:50;-webkit-backdrop-filter:blur(4px);backdrop-filter:blur(4px);background:#080b09a3;justify-content:flex-end;animation:.16s both station-in;display:flex;position:fixed;inset:0}.settings-sheet{width:min(430px,100%);min-height:100%;color:var(--paper);border-left:1px solid var(--phosphor);background:repeating-radial-gradient(ellipse at 104% 101%, transparent 0 13px, #b9ff450e 13.5px 14.5px), var(--ink);padding:30px 27px;animation:.3s cubic-bezier(.16,1,.3,1) both sheet-in;overflow-y:auto;box-shadow:-24px 0 70px #00000052}.settings-sheet>header{border-bottom:1px solid var(--line-light);justify-content:space-between;align-items:flex-start;gap:20px;margin-bottom:37px;padding-bottom:20px;display:flex}.settings-sheet h2{font:400 31px/1 var(--display);margin:0}.sheet-close{border:1px solid var(--line-light);width:35px;height:35px;font:300 22px/1 var(--body);cursor:pointer;background:0 0;border-radius:50%}.sheet-close:hover{color:var(--phosphor);border-color:var(--phosphor)}.setting-group{border-bottom:1px solid var(--line-light);grid-template-columns:24px 1fr auto;align-items:center;gap:12px;padding:15px 0;display:grid}.setting-number{color:var(--phosphor);font:8px/1 var(--mono)}.setting-group>div,.range-setting>span:first-child{gap:5px;display:grid}.setting-group small,.range-setting small{color:#e8dfca6e;font:7px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase}.setting-group strong,.range-setting strong{font:500 13px/1 var(--mono)}.setting-locked{color:#e8dfca5c;font:7px/1 var(--mono);text-transform:uppercase}.setting-select select{max-width:112px;color:var(--paper);border:1px solid var(--line-light);background:var(--ink-raised);font:8px/1 var(--mono);border-radius:0;padding:6px 7px}.range-setting{gap:13px;margin-top:29px;display:grid}.range-setting>span:first-child{justify-content:space-between;align-items:end;display:flex}.range-setting input{appearance:none;background:linear-gradient(90deg, var(--phosphor) 0 35%, #e8dfca1c 35%);border:1px solid #e8dfca3d;border-radius:0;width:100%;height:4px;margin:0}.range-setting:first-of-type input{background:var(--phosphor)}.range-setting input::-webkit-slider-thumb{appearance:none;border:2px solid var(--ink);background:var(--phosphor);width:12px;height:18px;box-shadow:0 0 0 1px var(--phosphor);cursor:grab}.range-setting input::-moz-range-thumb{border:2px solid var(--ink);background:var(--phosphor);width:10px;height:16px;box-shadow:0 0 0 1px var(--phosphor);cursor:grab;border-radius:0}.range-ticks{pointer-events:none;justify-content:space-between;margin-top:-13px;display:flex}.range-limits{color:#e8dfca59;font:7px/1 var(--mono);justify-content:space-between;display:flex}.range-limits i{font-style:normal}.system-prompt-setting{gap:9px;margin-top:29px;display:grid}.system-prompt-setting>span{color:#e8dfca6e;font:7px/1 var(--mono);letter-spacing:.05em;text-transform:uppercase;justify-content:space-between;display:flex}.system-prompt-setting textarea{resize:vertical;color:var(--paper);border:1px solid var(--line-light);background:var(--ink-raised);font:10px/1.45 var(--mono);padding:10px}.settings-actions{grid-template-columns:1fr 1fr;gap:7px;margin-top:29px;display:grid}.settings-actions button{color:var(--paper);border:1px solid var(--line-light);font:7px/1.2 var(--mono);text-transform:uppercase;cursor:pointer;background:0 0;padding:9px}.settings-actions button:hover{color:var(--phosphor);border-color:var(--phosphor)}.settings-actions .danger-action{color:#e78a62;border-color:#c96f4373;grid-column:span 2}.boot-screen{min-height:100dvh;color:var(--paper);background:var(--ink);justify-content:center;align-items:center;gap:18px;padding:24px;display:flex}.boot-screen h1{color:var(--phosphor);font:400 34px/1 var(--display);margin:0}.boot-screen p:last-child{color:var(--paper-ghost);font:10px/1.5 var(--mono)}.range-ticks i{background:#e8dfca40;width:1px;height:5px}.settings-note{background:#b9ff450f;border:1px solid #b9ff4540;align-items:flex-start;gap:13px;margin-top:42px;padding:15px;display:flex}.settings-note .signal-dot{flex:0 0 7px;margin-top:6px}.settings-note p{color:#e8dfca94;margin:0;font-size:10px;line-height:1.5}.settings-note strong{color:var(--phosphor)}@keyframes station-in{0%{opacity:0}to{opacity:1}}@keyframes rise-in{0%{opacity:0;transform:translateY(11px)}to{opacity:1;transform:translateY(0)}}@keyframes note-in{0%{opacity:0;transform:translateY(7px)}to{opacity:1;transform:translateY(0)}}@keyframes cursor-blink{0%,52%{opacity:1}53%,to{opacity:0}}@keyframes sheet-in{0%{opacity:0;transform:translate(34px)}to{opacity:1;transform:translate(0)}}@media (width<=1279px){.station-header{grid-template-columns:minmax(320px,1fr) auto auto}.local-promise{min-width:auto}.local-promise small{display:none}.workspace-grid{grid-template-columns:260px minmax(500px,1fr);grid-template-areas:"models chat""inspector inspector"}.model-rail,.chat-surface{height:calc(100dvh - 76px);min-height:calc(100dvh - 76px)}.inspector{border-top:1px solid var(--phosphor);border-left:0;grid-template-columns:repeat(2,minmax(0,1fr));height:auto;min-height:auto;display:grid;overflow:visible}.inspector-section{border-right:1px solid var(--line-light);border-bottom:0}}@media (width<=820px){.station-header{min-height:70px;display:block}.brand-block{padding:11px 70px 11px 15px}.brand-mark{flex-basis:36px;width:36px;height:36px}.local-promise{border:0;border-top:1px solid var(--line-light);min-height:31px;padding:0 16px}.header-actions{position:absolute;top:10px;right:0}.station-clock{display:none}.workspace-grid{display:block}.model-rail,.chat-surface,.inspector{height:auto;min-height:auto}.model-rail{border-right:0;border-bottom:1px solid var(--phosphor);padding:18px 15px;display:block}.model-list{scroll-snap-type:x mandatory;gap:7px;padding-bottom:7px;display:flex;overflow-x:auto}.model-card{scroll-snap-align:start;min-width:205px}.dflash-card{margin-top:13px}.rail-footnote{display:none}.chat-surface{grid-template-rows:auto 19px minmax(450px,auto) auto;display:grid;overflow:visible}.transcript{padding-inline:22px;overflow:visible}.chat-header{padding:17px 19px}.chat-header-meta span:not(.live-label){display:none}.chat-header-meta .header-load-button{display:block}.live-label{grid-column:1}.composer{z-index:5;padding-inline:14px;position:sticky;bottom:0}.inspector{border-top:1px solid var(--phosphor);display:block}.inspector-section{border-right:0;border-bottom:1px solid var(--line-light)}}@media (width<=520px){.brand-block h1{font-size:24px}.brand-block .eyebrow,.local-promise strong{font-size:7px}.chat-header{align-items:flex-end}.botanical-subtitle{margin:4px 0 0;display:block}.message-user{max-width:91%}.composer-meta>span{display:none}.composer-field{grid-template-columns:1fr;padding-bottom:50px}.send-button{height:35px;position:absolute;bottom:0;right:0}.speed-gauge{grid-template-columns:106px 1fr}.gauge-ring{width:102px;height:102px}.gauge-ring:after{width:113px;height:113px}}@media (prefers-reduced-motion:reduce){*,:before,:after{scroll-behavior:auto!important;transition-duration:.01ms!important;animation-duration:.01ms!important;animation-iteration-count:1!important}}.bench-page{min-height:100dvh;color:var(--ink);background:var(--paper)}.bench-masthead{min-height:86px;color:var(--paper);border-bottom:1px solid var(--phosphor);background:var(--ink);grid-template-columns:minmax(310px,1fr) minmax(360px,1.2fr) auto;align-items:stretch;display:grid}.bench-masthead .brand-block h1{color:var(--paper)}.bench-masthead-note{border-inline:1px solid var(--line-light);background:linear-gradient(110deg,#b9ff4517,#0000 72%);flex-direction:column;justify-content:center;gap:5px;padding:14px 28px;display:flex}.bench-masthead-note strong{color:var(--phosphor);font:700 9px/1 var(--mono);letter-spacing:.11em;text-transform:uppercase}.bench-masthead-note span{color:#e8dfca8c;max-width:520px;font:10px/1.35 var(--body)}.bench-return{color:var(--paper);font:700 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;align-items:center;gap:8px;padding:0 25px;text-decoration:none;display:flex}.bench-return:hover{color:var(--phosphor)}.bench-layout{grid-template-columns:minmax(275px,330px) minmax(0,1fr);min-height:calc(100dvh - 86px);display:grid}.bench-controls{color:var(--paper);border-right:1px solid var(--phosphor);background:repeating-radial-gradient(ellipse at -12% 102%, transparent 0 24px, #b9ff450b 25px 26px), var(--ink);padding:28px 22px}.bench-controls h2{font:400 30px/1 var(--display);letter-spacing:-.03em;margin:0}.bench-control-intro{color:#e8dfca80;margin:9px 0 22px;font-size:11px;line-height:1.45}.bench-controls form{gap:15px;display:grid}.bench-field{gap:6px;min-width:0;display:grid}.bench-field>span{color:#e8dfcac7;font:700 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase}.bench-field select,.bench-field input{width:100%;min-width:0;height:39px;color:var(--paper);background:var(--ink-raised);font:10px/1 var(--mono);border:1px solid #e8dfca38;border-radius:0;outline:0;padding:0 10px}.bench-field input::placeholder{color:#e8dfca40}.bench-field select:disabled,.bench-field input:disabled{opacity:.5}.bench-field small{color:#e8dfca59;font:8px/1.2 var(--mono)}.bench-field-pair{grid-template-columns:repeat(2,minmax(0,1fr));gap:10px;display:grid}.bench-experimental{background:#c96f430d;border:1px dashed #c96f437a;padding:11px}.bench-experimental summary{color:#e19a76;font:700 8px/1 var(--mono);letter-spacing:.07em;text-transform:uppercase;cursor:pointer}.bench-experimental>p{color:#e8dfca75;margin:10px 0;font-size:9px;line-height:1.4}.bench-locks{color:#e8dfca70;font:7px/1.35 var(--mono);gap:5px;margin:10px 0 0;display:grid}.bench-locks div{justify-content:space-between;gap:8px;display:flex}.bench-locks dt,.bench-locks dd{margin:0}.bench-locks dt{color:#e8dfcab8}.bench-gate{border-left:2px solid var(--rust);background:#c96f4314;gap:5px;padding:10px;display:grid}.bench-gate.pass{border-color:var(--phosphor);background:var(--phosphor-dim)}.bench-gate strong{color:#ff9a68;font:700 8px/1 var(--mono);letter-spacing:.06em;text-transform:uppercase}.bench-gate.pass strong{color:var(--phosphor)}.bench-gate span{color:#e8dfca80;font:8px/1.4 var(--mono);overflow-wrap:anywhere}.bench-run-actions{grid-template-columns:1fr auto;gap:8px;display:grid}.bench-run-button,.bench-stop-button,.bench-section-heading button{border:1px solid var(--ink);background:var(--phosphor);min-height:38px;font:800 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;cursor:pointer;border-radius:0;padding:0 13px}.bench-run-button:disabled{opacity:.35}.bench-stop-button{color:#ffb08a;border-color:var(--rust);background:0 0}.bench-workbench{background:linear-gradient(90deg, #1115120b 1px, transparent 1px) 0 0 / 22px 22px, linear-gradient(#1115120a 1px, transparent 1px) 0 0 / 22px 22px, var(--paper);min-width:0;padding:clamp(26px,4vw,58px)}.bench-hero{border-bottom:1px solid var(--line-dark);grid-template-columns:minmax(320px,1.3fr) minmax(260px,.7fr);align-items:end;gap:34px;padding-bottom:29px;display:grid}.bench-hero h2{max-width:690px;font:400 clamp(42px, 6vw, 82px)/.84 var(--display);letter-spacing:-.065em;margin:0}.bench-hero h2 em{color:#4f7125;font-weight:400}.bench-hero>p{color:#11151294;max-width:430px;margin:0;font-size:11px;line-height:1.55}.bench-progress{gap:8px;padding:16px 0;display:grid}.bench-progress>div{font:8px/1 var(--mono);letter-spacing:.05em;text-transform:uppercase;justify-content:space-between;gap:20px;display:flex}.bench-progress>div span{color:#11151273}.bench-progress-track{background:#11151221;height:3px}.bench-progress-track i{background:#527426;height:100%;transition:width .18s;display:block}.bench-error{color:var(--paper);border-left:3px solid var(--rust);background:var(--ink);gap:5px;margin-bottom:18px;padding:12px 14px;display:grid}.bench-error strong{color:#ff9a68;font:700 8px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase}.bench-error span{font:9px/1.4 var(--mono);overflow-wrap:anywhere}.bench-core-section{margin-top:13px}.bench-section-heading{justify-content:space-between;align-items:end;gap:20px;margin-bottom:13px;display:flex}.bench-section-heading h3,.bench-output h3{font:500 24px/1 var(--display);margin:0}.bench-section-heading button{min-height:31px;color:var(--paper);border-color:var(--ink);background:var(--ink)}.bench-core-sample{border:1px solid var(--ink);background:var(--paper-deep);grid-template-columns:repeat(4,minmax(0,1fr));display:grid;position:relative;overflow:hidden}.bench-core-sample:before{content:"";pointer-events:none;opacity:.35;background:repeating-radial-gradient(#0000 0 34px,#1115122b 35px 36px);position:absolute;inset:-130% -5%}.bench-core-sample article{background:#e8dfca91;border-right:1px solid #11151242;flex-direction:column;justify-content:space-between;min-height:128px;padding:17px;display:flex;position:relative}.bench-core-sample article:nth-child(2n){background:#d8cdb499}.bench-core-sample article:last-child{color:var(--paper);background:#111512e8;border-right:0}.bench-core-sample article>span{font:800 8px/1 var(--mono);letter-spacing:.09em;text-transform:uppercase}.bench-core-sample article strong{font:400 clamp(28px, 3vw, 43px)/1 var(--display);letter-spacing:-.04em;margin-block:auto}.bench-core-sample article small{color:#11151280;font:7px/1.2 var(--mono);text-transform:uppercase}.bench-core-sample article:last-child small{color:#e8dfca7a}.bench-detail-grid{grid-template-columns:repeat(2,minmax(0,1fr));gap:12px;margin-top:12px;display:grid}.bench-data-card{border:1px solid var(--line-dark);background:#e8dfcaad;padding:16px}.bench-runtime-card,.bench-tuning-card{grid-column:span 2}.bench-data-card dl{grid-template-columns:repeat(2,minmax(0,1fr));gap:12px 26px;margin:0;display:grid}.bench-runtime-card dl{grid-template-columns:minmax(180px,1.5fr) repeat(4,minmax(100px,.7fr))}.bench-tuning-card dl{grid-template-columns:repeat(4,minmax(0,1fr))}.bench-data-card dl div{border-top:1px solid #11151221;gap:4px;min-width:0;padding-top:8px;display:grid}.bench-data-card dt{color:#11151273;font:7px/1 var(--mono);letter-spacing:.07em;text-transform:uppercase}.bench-data-card dd{min-width:0;font:500 16px/1.2 var(--body);text-overflow:ellipsis;white-space:nowrap;margin:0;overflow:hidden}.bench-output{border:1px dashed #1115124d;grid-template-columns:repeat(2,minmax(0,1fr));gap:34px;margin-top:12px;padding:21px;display:grid}.bench-output>div:last-child{border-left:1px solid var(--line-dark);padding-left:34px}.bench-output p:not(.eyebrow){color:#11151299;white-space:pre-wrap;margin:8px 0 0;font-size:11px;line-height:1.5}@media (width<=1050px){.bench-masthead{grid-template-columns:minmax(310px,1fr) auto}.bench-masthead-note{display:none}.bench-layout{grid-template-columns:285px minmax(0,1fr)}.bench-core-sample{grid-template-columns:repeat(2,minmax(0,1fr))}.bench-core-sample article:nth-child(2){border-right:0}.bench-core-sample article:nth-child(-n+2){border-bottom:1px solid #11151242}.bench-runtime-card dl{grid-template-columns:repeat(2,minmax(0,1fr))}}@media (width<=760px){.bench-masthead{align-items:center;display:flex}.bench-masthead .brand-block{flex:1;min-width:0}.bench-return{padding:0 15px;font-size:0}.bench-return span{font-size:17px}.bench-layout{display:block}.bench-controls{border-right:0;border-bottom:1px solid var(--phosphor)}.bench-workbench{padding:28px 16px}.bench-hero,.bench-output{grid-template-columns:1fr;gap:18px}.bench-hero h2{font-size:clamp(42px,14vw,68px)}.bench-detail-grid{display:block}.bench-data-card+.bench-data-card{margin-top:10px}.bench-output>div:last-child{border-top:1px solid var(--line-dark);border-left:0;padding-top:18px;padding-left:0}}@media (width<=480px){.bench-masthead .brand-mark{display:none}.bench-core-sample,.bench-data-card dl,.bench-runtime-card dl,.bench-tuning-card dl{grid-template-columns:1fr}.bench-core-sample article{border-bottom:1px solid #11151242;border-right:0;min-height:105px}.bench-core-sample article:last-child{border-bottom:0}}

dist/index.html

@@ -10,8 +10,8 @@
       content="Bonsai local inference field station — private WebGPU language models in your browser."
     />
     <title>Bonsai Field Station</title>
-    <script type="module" crossorigin src="/assets/index-AGrRq4Jz.js"></script>
-    <link rel="stylesheet" crossorigin href="/assets/index-Do-82RUc.css">
+    <script type="module" crossorigin src="/assets/index-jtL5gUVS.js"></script>
+    <link rel="stylesheet" crossorigin href="/assets/index-C-2tm91j.css">
   </head>
   <body>
     <div id="root"></div>

scripts/verify-wllama-assets.mjs

@@ -40,7 +40,14 @@ if (!declarations.includes('offload_token_embedding?: boolean')) {
 if (!declarations.includes('n_ubatch?: number')) {
   throw new Error('Vendored Bonsai wllama declarations lack n_ubatch');
 }
+const wllamaDeclarations = await readFile(resolve(root, 'vendor/wllama-bonsai/esm/wllama.d.ts'), 'utf8');
+if (!wllamaDeclarations.includes('forceCompat?: boolean')) {
+  throw new Error('Vendored Bonsai wllama declarations lack forceCompat');
+}
 const moduleSource = await readFile(resolve(root, 'vendor/wllama-bonsai/esm/index.js'), 'utf8');
 if (!moduleSource.includes('n_ubatch: params.n_ubatch')) {
   throw new Error('Vendored Bonsai wllama module does not forward n_ubatch');
 }
+if (!moduleSource.includes('const forceCompat = this.config.forceCompat === true')) {
+  throw new Error('Vendored Bonsai wllama module does not implement forceCompat');
+}

src/app/styles.css

@@ -2590,7 +2590,8 @@ button:disabled {
   background: rgba(232, 223, 202, 0.68);
 }
 
-.bench-runtime-card {
+.bench-runtime-card,
+.bench-tuning-card {
   grid-column: span 2;
 }
 
@@ -2602,7 +2603,11 @@ button:disabled {
 }
 
 .bench-runtime-card dl {
-  grid-template-columns: minmax(180px, 1.5fr) repeat(3, minmax(100px, 0.7fr));
+  grid-template-columns: minmax(180px, 1.5fr) repeat(4, minmax(100px, 0.7fr));
+}
+
+.bench-tuning-card dl {
+  grid-template-columns: repeat(4, minmax(0, 1fr));
 }
 
 .bench-data-card dl div {
@@ -2747,7 +2752,8 @@ button:disabled {
 
   .bench-core-sample,
   .bench-data-card dl,
-  .bench-runtime-card dl {
+  .bench-runtime-card dl,
+  .bench-tuning-card dl {
     grid-template-columns: 1fr;
   }
 

src/bench/BenchApp.tsx

@@ -4,6 +4,9 @@ import {
   EngineClientError,
   evaluateModelGate,
   loadModelManifestV2,
+  type BenchmarkFlashMode,
+  type BenchmarkKvCacheType,
+  type BenchmarkWasmFlavor,
   type EngineCapabilities,
   type ModelManifestV2,
   type ModelTierId,
@@ -21,6 +24,7 @@ import {
 
 const MANIFEST_PATH = 'manifest/models.json';
 const DEFAULT_MODEL: ModelTierId = '1_7b';
+const MAX_GLUE_INT = 2_147_483_647;
 
 interface RunProgress {
   label: string;
@@ -57,7 +61,7 @@ function throwIfAborted(signal: AbortSignal): void {
 function parseOptionalInteger(value: string): number | null {
   if (value.trim() === '') return null;
   const parsed = Number(value);
-  return Number.isSafeInteger(parsed) && parsed > 0 ? parsed : null;
+  return Number.isSafeInteger(parsed) && parsed > 0 && parsed <= MAX_GLUE_INT ? parsed : null;
 }
 
 function formatDuration(value: number | null | undefined): string {
@@ -91,6 +95,9 @@ export function BenchApp() {
   const [maxTokens, setMaxTokens] = useState(64);
   const [nBatchInput, setNBatchInput] = useState('');
   const [nUbatchInput, setNUbatchInput] = useState('');
+  const [flashMode, setFlashMode] = useState<BenchmarkFlashMode>('off');
+  const [kvCacheType, setKvCacheType] = useState<BenchmarkKvCacheType>('f16');
+  const [wasmFlavor, setWasmFlavor] = useState<BenchmarkWasmFlavor>('auto');
   const [running, setRunning] = useState(false);
   const [progress, setProgress] = useState<RunProgress>({
     label: 'Waiting for an operator-initiated run',
@@ -99,7 +106,8 @@ export function BenchApp() {
   });
   const [report, setReport] = useState<BenchmarkReport | null>(null);
   const [output, setOutput] = useState('');
-  const [error, setError] = useState<string | null>(null);
+  const [bootstrapError, setBootstrapError] = useState<string | null>(null);
+  const [runError, setRunError] = useState<string | null>(null);
   const operationRef = useRef<AbortController | null>(null);
 
   useEffect(() => {
@@ -125,7 +133,7 @@ export function BenchApp() {
       });
     }).catch((bootstrapError) => {
       if (cancelled || controller.signal.aborted) return;
-      setError(errorMessage(bootstrapError));
+      setBootstrapError(errorMessage(bootstrapError));
       setProgress((current) => ({ ...current, label: 'Runtime inspection failed' }));
     });
     return () => {
@@ -156,17 +164,44 @@ export function BenchApp() {
     : model?.contextLength ?? 4_096;
   const nBatch = parseOptionalInteger(nBatchInput);
   const nUbatch = parseOptionalInteger(nUbatchInput);
-  const tuningValid = (nBatchInput.trim() === '' || nBatch !== null)
-    && (nUbatchInput.trim() === '' || nUbatch !== null);
+  const batchTuningError = nBatchInput.trim() !== '' && nBatch === null
+    ? `nBatch must be an integer from 1 to ${MAX_GLUE_INT.toLocaleString()}.`
+    : nUbatchInput.trim() !== '' && nUbatch === null
+      ? `nUbatch must be an integer from 1 to ${MAX_GLUE_INT.toLocaleString()}.`
+      : nBatch !== null && nUbatch !== null && nUbatch > nBatch
+        ? 'nUbatch must not exceed nBatch.'
+        : null;
+  const runtimeTuningError = flashMode === 'off' && kvCacheType !== 'f16'
+    ? 'Quantized KV requires Flash Attention auto.'
+    : flashMode === 'auto' && backend !== 'webgpu'
+      ? 'Flash Attention experiments require the explicit WebGPU backend.'
+      : null;
+  const wasmTuningError = wasmFlavor === 'jspi' && capabilities?.runtime.wasmFlavor === 'compat'
+    ? 'JSPI is unavailable in this browser; use auto or compat.'
+    : null;
+  const configurationTuningError = batchTuningError ?? runtimeTuningError ?? wasmTuningError;
   const configurationValid = Number.isSafeInteger(contextSize)
     && contextSize > 0
     && contextSize <= contextLimit
     && Number.isSafeInteger(maxTokens)
     && maxTokens >= 8
     && maxTokens <= 512
-    && tuningValid;
+    && configurationTuningError === null;
   const canRun = Boolean(manifest && capabilities && model && deviceGate?.allowed && configurationValid);
 
+  const clearPreviousRun = () => {
+    setReport(null);
+    setOutput('');
+    setRunError(null);
+    if (manifest && capabilities) {
+      setProgress({
+        label: 'Ready · settings changed; run again to produce matching evidence',
+        loadedBytes: 0,
+        totalBytes: 0,
+      });
+    }
+  };
+
   const handleModelChange = (nextId: ModelTierId) => {
     if (running || !manifest) return;
     const next = manifest.models.find((candidate) => candidate.id === nextId);
@@ -174,9 +209,7 @@ export function BenchApp() {
     setModelId(nextId);
     setContextSize(next.defaultContext);
     if (!next.cpuFallback && backend === 'wasm') setBackend('auto');
-    setReport(null);
-    setOutput('');
-    setError(null);
+    clearPreviousRun();
   };
 
   const runBenchmark = () => {
@@ -186,7 +219,7 @@ export function BenchApp() {
     setRunning(true);
     setReport(null);
     setOutput('');
-    setError(null);
+    setRunError(null);
     const startedAt = new Date().toISOString();
 
     const run = async () => {
@@ -209,9 +242,13 @@ export function BenchApp() {
         modelId: model.id,
         backend,
         contextSize,
-        ...(nBatch === null ? {} : { nBatch }),
-        ...(nUbatch === null ? {} : { nUbatch }),
-        flashAttention: false as const,
+        benchmarkTuning: {
+          ...(nBatch === null ? {} : { nBatch }),
+          ...(nUbatch === null ? {} : { nUbatch }),
+          flashMode,
+          kvCacheType,
+          wasmFlavor,
+        },
       };
 
       const coldStarted = performance.now();
@@ -252,6 +289,11 @@ export function BenchApp() {
       if (coldLoad.backend !== warmLoad.backend) {
         throw new Error(`Backend changed between cold and warm loads (${coldLoad.backend} → ${warmLoad.backend}).`);
       }
+      if (coldLoad.tuning.observed.wasmFlavor !== warmLoad.tuning.observed.wasmFlavor) {
+        throw new Error(
+          `WASM flavor changed between cold and warm loads (${coldLoad.tuning.observed.wasmFlavor} → ${warmLoad.tuning.observed.wasmFlavor}).`,
+        );
+      }
       throwIfAborted(controller.signal);
 
       let firstTokenAt: number | null = null;
@@ -294,8 +336,6 @@ export function BenchApp() {
         requestedBackend: backend,
         contextSize: warmLoad.context.size,
         maxTokens,
-        nBatch,
-        nUbatch,
         coldLoadMs,
         warmLoadMs,
         coldCachedBytes,
@@ -321,7 +361,7 @@ export function BenchApp() {
         setProgress({ label: 'Benchmark stopped by operator', loadedBytes: 0, totalBytes: 0 });
         return;
       }
-      setError(errorMessage(runError));
+      setRunError(errorMessage(runError));
       setProgress({ label: 'Benchmark failed safely · no report exported', loadedBytes: 0, totalBytes: 0 });
     }).finally(() => {
       if (operationRef.current === controller) operationRef.current = null;
@@ -379,8 +419,7 @@ export function BenchApp() {
                 disabled={running}
                 onChange={(event) => {
                   setBackend(event.target.value as RequestedBackend);
-                  setReport(null);
-                  setOutput('');
+                  clearPreviousRun();
                 }}
               >
                 <option value="auto">Auto · gated</option>
@@ -402,7 +441,7 @@ export function BenchApp() {
                   disabled={running}
                   onChange={(event) => {
                     setContextSize(Number(event.target.value));
-                    setReport(null);
+                    clearPreviousRun();
                   }}
                 />
                 <small>Max {contextLimit.toLocaleString()}</small>
@@ -418,7 +457,7 @@ export function BenchApp() {
                   disabled={running}
                   onChange={(event) => {
                     setMaxTokens(Number(event.target.value));
-                    setReport(null);
+                    clearPreviousRun();
                   }}
                 />
                 <small>8–512</small>
@@ -426,21 +465,22 @@ export function BenchApp() {
             </div>
 
             <details className="bench-experimental">
-              <summary>Experimental batch shape</summary>
-              <p>Blank uses the pinned runtime defaults. Values are exported as experiments, not release recommendations.</p>
+              <summary>Experimental runtime tuning</summary>
+              <p>Blank batch fields use the pinned runtime defaults. Every option is exported as an experiment, not a release recommendation.</p>
               <div className="bench-field-pair">
                 <label className="bench-field">
                   <span>nBatch</span>
                   <input
                     type="number"
                     min="1"
+                    max={MAX_GLUE_INT}
                     step="1"
                     placeholder="runtime default"
                     value={nBatchInput}
                     disabled={running}
                     onChange={(event) => {
                       setNBatchInput(event.target.value);
-                      setReport(null);
+                      clearPreviousRun();
                     }}
                   />
                 </label>
@@ -449,23 +489,75 @@ export function BenchApp() {
                   <input
                     type="number"
                     min="1"
+                    max={MAX_GLUE_INT}
                     step="1"
                     placeholder="runtime default"
                     value={nUbatchInput}
                     disabled={running}
                     onChange={(event) => {
                       setNUbatchInput(event.target.value);
-                      setReport(null);
+                      clearPreviousRun();
                     }}
                   />
                 </label>
               </div>
+              <div className="bench-field-pair">
+                <label className="bench-field">
+                  <span>Flash Attention request</span>
+                  <select
+                    value={flashMode}
+                    disabled={running}
+                    onChange={(event) => {
+                      setFlashMode(event.target.value as BenchmarkFlashMode);
+                      clearPreviousRun();
+                    }}
+                  >
+                    <option value="off">Off · default</option>
+                    <option value="auto">Auto · observed on</option>
+                  </select>
+                </label>
+                <label className="bench-field">
+                  <span>KV cache K / V</span>
+                  <select
+                    value={kvCacheType}
+                    disabled={running}
+                    onChange={(event) => {
+                      setKvCacheType(event.target.value as BenchmarkKvCacheType);
+                      clearPreviousRun();
+                    }}
+                  >
+                    <option value="f16">f16 / f16 · default</option>
+                    <option value="q8_0">q8_0 / q8_0 · exp.</option>
+                    <option value="q4_0">q4_0 / q4_0 · exp.</option>
+                  </select>
+                </label>
+              </div>
+              <label className="bench-field">
+                <span>WASM runtime</span>
+                <select
+                  value={wasmFlavor}
+                  disabled={running}
+                  onChange={(event) => {
+                    setWasmFlavor(event.target.value as BenchmarkWasmFlavor);
+                    clearPreviousRun();
+                  }}
+                >
+                  <option value="auto">Auto · browser-selected</option>
+                  <option value="jspi">JSPI · explicit</option>
+                  <option value="compat">Compat / Asyncify · explicit</option>
+                </select>
+                <small>Current auto path: {capabilities?.runtime.wasmFlavor ?? 'inspecting'}</small>
+              </label>
               <dl className="bench-locks">
-                <div><dt>flashAttention</dt><dd>false · locked, unvalidated</dd></div>
-                <div><dt>KV cache K / V</dt><dd>f16 / f16 · locked release default</dd></div>
+                <div><dt>Chat release</dt><dd>Flash off · KV f16/f16</dd></div>
+                <div><dt>Evidence rule</dt><dd>native log must match or the run fails</dd></div>
               </dl>
             </details>
 
+            {configurationTuningError && (
+              <p className="bench-inline-warning" role="alert">{configurationTuningError}</p>
+            )}
+
             <div className={`bench-gate ${deviceGate?.allowed ? 'pass' : 'fail'}`}>
               <strong>{deviceGate?.allowed ? 'Device preflight passed' : 'Device preflight blocked'}</strong>
               <span>
@@ -510,7 +602,8 @@ export function BenchApp() {
             <span className="bench-progress-track"><i style={{ width: `${progressPercent}%` }} /></span>
           </section>
 
-          {error && <div className="bench-error" role="alert"><strong>Run rejected</strong><span>{error}</span></div>}
+          {bootstrapError && <div className="bench-error" role="alert"><strong>Runtime inspection failed</strong><span>{bootstrapError}</span></div>}
+          {runError && <div className="bench-error" role="alert"><strong>Run rejected</strong><span>{runError}</span></div>}
 
           <section className="bench-core-section" aria-label="Benchmark measurements">
             <div className="bench-section-heading">
@@ -567,6 +660,16 @@ export function BenchApp() {
                 <div><dt>wllama</dt><dd title={report?.runtime.wllamaRevision}>{report?.runtime.wllamaRevision.slice(0, 12) ?? '—'}</dd></div>
                 <div><dt>llama.cpp</dt><dd title={report?.runtime.llamaCppRevision}>{report?.runtime.llamaCppRevision.slice(0, 12) ?? '—'}</dd></div>
                 <div><dt>Weights</dt><dd title={report?.runtime.manifestRevision ?? undefined}>{report?.runtime.manifestRevision?.slice(0, 12) ?? '—'}</dd></div>
+                <div><dt>WASM</dt><dd>{report?.runtime.wasmFlavor.toUpperCase() ?? '—'}</dd></div>
+              </dl>
+            </article>
+            <article className="bench-data-card bench-tuning-card">
+              <p className="eyebrow ink">Native tuning</p>
+              <dl>
+                <div><dt>Applied</dt><dd>{report ? (report.load.tuning.applied ? 'yes' : 'no') : '—'}</dd></div>
+                <div><dt>Flash</dt><dd>{report ? `${report.load.tuning.requested.flashMode} → ${String(report.load.tuning.observed.flashAttention)}` : '—'}</dd></div>
+                <div><dt>K/V</dt><dd>{report ? `${report.load.tuning.observed.cacheTypeK}/${report.load.tuning.observed.cacheTypeV}` : '—'}</dd></div>
+                <div><dt>KV bytes</dt><dd>{report?.load.tuning.observed.kvBufferBytes === null || report?.load.tuning.observed.kvBufferBytes === undefined ? '—' : formatBytes(report.load.tuning.observed.kvBufferBytes)}</dd></div>
               </dl>
             </article>
           </section>
@@ -579,7 +682,7 @@ export function BenchApp() {
             </div>
             <div>
               <p className="eyebrow ink">Generated text · excluded from export</p>
-              <p>{output || 'Run the assay to inspect the generated specimen.'}</p>
+              <p data-testid="bench-generated-output">{output || 'Run the assay to inspect the generated specimen.'}</p>
             </div>
           </section>
         </main>

src/bench/report.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, it } from 'vitest';
 import type {
+  BackendReport,
   EngineCapabilities,
   LoadModelResult,
   ManifestModelV2,
@@ -13,6 +14,21 @@ import {
   type BenchmarkObservation,
 } from './report';
 
+const JSPI_WASM_SHA256 = 'd'.repeat(64);
+const COMPAT_WASM_SHA256 = 'c'.repeat(64);
+const COMPAT_WORKER_SHA256 = 'b'.repeat(64);
+
+const dflash27bPairing = {
+  support: 'conditional',
+  compatibilityStatus: 'native-greedy-parity-pass',
+  runtimeAvailable: false,
+  blockSize: 16,
+  maskTokenId: 248_070,
+  targetLayerIdsZeroBased: [1, 16, 31, 46, 61],
+  targetTapsOneBased: [2, 17, 32, 47, 62],
+  reason: 'Native parity passes; browser activation remains gated.',
+} as const;
+
 const model = {
   id: '8b',
   displayName: 'Bonsai 8B',
@@ -25,8 +41,16 @@ const manifest = {
     id: 'WaveCut/Bonsai-web-GGUF',
     revision: '112ea7a1a6229bde132b176b9a72477a7ecfde64',
   },
-  dflash: { status: 'conditional' },
-} as ModelManifestV2;
+  dflash: {
+    status: 'conditional',
+    pairings: {
+      '1_7b': { support: 'unsupported', reason: 'No compatible 1.7B drafter is locked.' },
+      '4b': { support: 'unsupported', reason: 'The 27B drafter is incompatible with Bonsai 4B.' },
+      '8b': { support: 'unsupported', reason: 'The 27B drafter is incompatible with Bonsai 8B.' },
+      '27b': dflash27bPairing,
+    },
+  },
+} as unknown as ModelManifestV2;
 
 const capabilities = {
   crossOriginIsolated: true,
@@ -55,18 +79,76 @@ const capabilities = {
     patchSetSha256: 'f'.repeat(64),
     moduleSha256: 'e'.repeat(64),
     wasmFlavor: 'jspi',
-    wasmSha256: 'd'.repeat(64),
+    wasmSha256: JSPI_WASM_SHA256,
     compatWorkerSha256: null,
     tokenEmbeddingOnWebGPU: true,
     tensorPlacementOverrides: false,
   },
 } satisfies EngineCapabilities;
 
+const backendReport = {
+  backends: ['WebGPU'],
+  nGraphSplits: 1,
+  opsOnCpu: 0,
+  layersGpu: { offloaded: 49, total: 49 },
+  flashAttention: false,
+  cacheTypeK: 'f16',
+  cacheTypeV: 'f16',
+  webgpuKvBufferBytes: 56 * 1024 ** 2,
+} satisfies BackendReport;
+
 const loadResult = {
+  modelId: '8b',
   backend: 'webgpu',
-  context: { batchSize: 2_048, microBatchSize: 512 },
-  tuning: { flashAttention: false, cacheTypeK: 'f16', cacheTypeV: 'f16' },
-} as LoadModelResult;
+  gate: {
+    allowed: true,
+    requestedBackend: 'auto',
+    selectedBackend: 'webgpu',
+    reasons: [],
+    warnings: [],
+    requiredBytes: 1_000,
+    availableStorageBytes: 9_000,
+  },
+  shardUrls: ['https://models.test/bonsai-8b.gguf'],
+  context: {
+    size: 4_096,
+    trainingSize: 32_768,
+    layerCount: 49,
+    vocabularySize: 151_936,
+    batchSize: 2_048,
+    microBatchSize: 512,
+  },
+  tuning: {
+    scope: 'release-defaults',
+    requested: {
+      nBatch: null,
+      nUbatch: null,
+      flashMode: 'off',
+      cacheTypeK: 'f16',
+      cacheTypeV: 'f16',
+      wasmFlavor: 'auto',
+    },
+    observed: {
+      nBatch: 2_048,
+      nUbatch: 512,
+      flashAttention: false,
+      cacheTypeK: 'f16',
+      cacheTypeV: 'f16',
+      kvBufferBytes: 56 * 1024 ** 2,
+      wasmFlavor: 'jspi',
+      wasmSha256: JSPI_WASM_SHA256,
+      compatWorkerSha256: null,
+    },
+    applied: true,
+  },
+  chatTemplate: {
+    bytes: 1_024,
+    hasThinkMarker: true,
+    hasToolCallMarker: true,
+    hasToolResponseMarker: true,
+  },
+  backendReport,
+} satisfies LoadModelResult;
 
 function observation(overrides: Partial<BenchmarkObservation> = {}): BenchmarkObservation {
   return {
@@ -78,8 +160,6 @@ function observation(overrides: Partial<BenchmarkObservation> = {}): BenchmarkOb
     requestedBackend: 'auto',
     contextSize: 4_096,
     maxTokens: 64,
-    nBatch: null,
-    nUbatch: null,
     coldLoadMs: 1_234.56,
     warmLoadMs: 234.54,
     coldCachedBytes: 250,
@@ -95,12 +175,7 @@ function observation(overrides: Partial<BenchmarkObservation> = {}): BenchmarkOb
     generationElapsedMs: 900.04,
     timeToFirstTokenMs: 123.45,
     streamedTokenEvents: 5,
-    backendReport: {
-      backends: ['WebGPU'],
-      nGraphSplits: 1,
-      opsOnCpu: 0,
-      layersGpu: { offloaded: 49, total: 49 },
-    },
+    backendReport,
     ...overrides,
   };
 }
@@ -118,14 +193,27 @@ describe('benchmark report', () => {
 
     expect(report.load).toMatchObject({
       tuning: {
-        requestedNBatch: null,
-        requestedNUbatch: null,
-        nBatch: 2_048,
-        nUbatch: 512,
-        flashAttention: false,
-        cacheTypeK: 'f16',
-        cacheTypeV: 'f16',
-        status: 'release-defaults',
+        scope: 'release-defaults',
+        requested: {
+          nBatch: null,
+          nUbatch: null,
+          flashMode: 'off',
+          cacheTypeK: 'f16',
+          cacheTypeV: 'f16',
+          wasmFlavor: 'auto',
+        },
+        observed: {
+          nBatch: 2_048,
+          nUbatch: 512,
+          flashAttention: false,
+          cacheTypeK: 'f16',
+          cacheTypeV: 'f16',
+          kvBufferBytes: 56 * 1024 ** 2,
+          wasmFlavor: 'jspi',
+          wasmSha256: JSPI_WASM_SHA256,
+          compatWorkerSha256: null,
+        },
+        applied: true,
       },
       cold: { durationMs: 1_234.6, cacheState: 'partial' },
       warm: { durationMs: 234.5 },
@@ -149,7 +237,7 @@ describe('benchmark report', () => {
       patchSetSha256: 'f'.repeat(64),
       moduleSha256: 'e'.repeat(64),
       wasmFlavor: 'jspi',
-      wasmSha256: 'd'.repeat(64),
+      wasmSha256: JSPI_WASM_SHA256,
       compatWorkerSha256: null,
     });
     expect(report.environment).toEqual({
@@ -184,24 +272,110 @@ describe('benchmark report', () => {
       dflash: {
         enabled: false,
         runtimeAvailable: false,
-        status: 'conditional',
+        globalStatus: 'conditional',
+        pairing: {
+          support: 'unsupported',
+          reason: 'The 27B drafter is incompatible with Bonsai 8B.',
+        },
       },
     });
   });
 
-  it('records experimental batch shape by the runtime field names without enabling flash attention', () => {
-    const report = buildBenchmarkReport(observation({ nBatch: 512, nUbatch: 128 }));
+  it('records requested and observed experimental tuning without relabelling it as release policy', () => {
+    const experimentalLoadResult: LoadModelResult = {
+      ...loadResult,
+      context: {
+        ...loadResult.context,
+        batchSize: 512,
+        microBatchSize: 128,
+      },
+      tuning: {
+        scope: 'benchmark',
+        requested: {
+          nBatch: 512,
+          nUbatch: 128,
+          flashMode: 'auto',
+          cacheTypeK: 'q8_0',
+          cacheTypeV: 'q8_0',
+          wasmFlavor: 'compat',
+        },
+        observed: {
+          nBatch: 512,
+          nUbatch: 128,
+          flashAttention: true,
+          cacheTypeK: 'q8_0',
+          cacheTypeV: 'q8_0',
+          kvBufferBytes: 28 * 1024 ** 2,
+          wasmFlavor: 'compat',
+          wasmSha256: COMPAT_WASM_SHA256,
+          compatWorkerSha256: COMPAT_WORKER_SHA256,
+        },
+        applied: true,
+      },
+    };
+    const report = buildBenchmarkReport(observation({ loadResult: experimentalLoadResult }));
 
     expect(report.load.tuning).toEqual({
-      requestedNBatch: 512,
-      requestedNUbatch: 128,
-      nBatch: 2_048,
-      nUbatch: 512,
-      flashAttention: false,
-      cacheTypeK: 'f16',
-      cacheTypeV: 'f16',
-      status: 'experimental-batch-shape',
+      scope: 'benchmark',
+      requested: {
+        nBatch: 512,
+        nUbatch: 128,
+        flashMode: 'auto',
+        cacheTypeK: 'q8_0',
+        cacheTypeV: 'q8_0',
+        wasmFlavor: 'compat',
+      },
+      observed: {
+        nBatch: 512,
+        nUbatch: 128,
+        flashAttention: true,
+        cacheTypeK: 'q8_0',
+        cacheTypeV: 'q8_0',
+        kvBufferBytes: 28 * 1024 ** 2,
+        wasmFlavor: 'compat',
+        wasmSha256: COMPAT_WASM_SHA256,
+        compatWorkerSha256: COMPAT_WORKER_SHA256,
+      },
+      applied: true,
+    });
+    expect(report.runtime).toMatchObject({
+      wasmFlavor: 'compat',
+      wasmSha256: COMPAT_WASM_SHA256,
+      compatWorkerSha256: COMPAT_WORKER_SHA256,
+    });
+  });
+
+  it('exports the selected model DFlash pairing for unsupported 1.7B and conditional 27B', () => {
+    const reportFor = (modelId: ManifestModelV2['id']) => buildBenchmarkReport(observation({
+      model: {
+        ...model,
+        id: modelId,
+        displayName: `Bonsai ${modelId}`,
+      } as ManifestModelV2,
+      loadResult: { ...loadResult, modelId },
+    }));
+
+    expect(reportFor('1_7b').releasePolicy.dflash).toEqual({
+      enabled: false,
+      runtimeAvailable: false,
+      globalStatus: 'conditional',
+      pairing: {
+        support: 'unsupported',
+        reason: 'No compatible 1.7B drafter is locked.',
+      },
+    });
+
+    const report27b = reportFor('27b');
+    expect(report27b.releasePolicy.dflash).toEqual({
+      enabled: false,
+      runtimeAvailable: false,
+      globalStatus: 'conditional',
+      pairing: dflash27bPairing,
     });
+    const exportedPairing = report27b.releasePolicy.dflash.pairing;
+    expect(exportedPairing.support).toBe('conditional');
+    if (exportedPairing.support !== 'conditional') throw new Error('Expected the conditional 27B pairing.');
+    expect(exportedPairing.targetLayerIdsZeroBased).not.toBe(dflash27bPairing.targetLayerIdsZeroBased);
   });
 
   it('uses streamed events when final engine usage is present but incomplete', () => {

src/bench/report.ts

@@ -22,8 +22,6 @@ export interface BenchmarkObservation {
   requestedBackend: RequestedBackend;
   contextSize: number;
   maxTokens: number;
-  nBatch: number | null;
-  nUbatch: number | null;
   coldLoadMs: number;
   warmLoadMs: number;
   coldCachedBytes: number | null;
@@ -37,7 +35,7 @@ export interface BenchmarkObservation {
 }
 
 export interface BenchmarkReport {
-  schemaVersion: 1;
+  schemaVersion: 2;
   kind: 'bonsai-browser-benchmark';
   startedAt: string;
   completedAt: string;
@@ -49,16 +47,7 @@ export interface BenchmarkReport {
     contextSize: number;
   };
   load: {
-    tuning: {
-      requestedNBatch: number | null;
-      requestedNUbatch: number | null;
-      nBatch: number;
-      nUbatch: number;
-      flashAttention: boolean;
-      cacheTypeK: LoadModelResult['tuning']['cacheTypeK'];
-      cacheTypeV: LoadModelResult['tuning']['cacheTypeV'];
-      status: 'release-defaults' | 'experimental-batch-shape';
-    };
+    tuning: LoadModelResult['tuning'];
     cold: {
       durationMs: number | null;
       cachedBytesAtStart: number | null;
@@ -133,7 +122,8 @@ export interface BenchmarkReport {
     dflash: {
       enabled: false;
       runtimeAvailable: false;
-      status: ModelManifestV2['dflash']['status'];
+      globalStatus: ModelManifestV2['dflash']['status'];
+      pairing: ModelManifestV2['dflash']['pairings'][ManifestModelV2['id']];
     };
   };
 }
@@ -179,9 +169,10 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
   const device = loadResult.backend === 'webgpu'
     ? capabilities.webgpu.name || 'WebGPU adapter'
     : `CPU · ${capabilities.hardwareConcurrency} logical cores`;
+  const dflashPairing = observation.manifest.dflash.pairings[observation.model.id];
 
   return {
-    schemaVersion: 1,
+    schemaVersion: 2,
     kind: 'bonsai-browser-benchmark',
     startedAt: observation.startedAt,
     completedAt: observation.completedAt,
@@ -194,16 +185,10 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
     },
     load: {
       tuning: {
-        requestedNBatch: observation.nBatch,
-        requestedNUbatch: observation.nUbatch,
-        nBatch: loadResult.context.batchSize,
-        nUbatch: loadResult.context.microBatchSize,
-        flashAttention: loadResult.tuning.flashAttention,
-        cacheTypeK: loadResult.tuning.cacheTypeK,
-        cacheTypeV: loadResult.tuning.cacheTypeV,
-        status: observation.nBatch === null && observation.nUbatch === null
-          ? 'release-defaults'
-          : 'experimental-batch-shape',
+        scope: loadResult.tuning.scope,
+        requested: { ...loadResult.tuning.requested },
+        observed: { ...loadResult.tuning.observed },
+        applied: loadResult.tuning.applied,
       },
       cold: {
         durationMs: roundedNonNegative(observation.coldLoadMs, 1),
@@ -255,9 +240,9 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
       llamaCppRevision: capabilities.runtime.llamaCppRevision,
       patchSetSha256: capabilities.runtime.patchSetSha256,
       moduleSha256: capabilities.runtime.moduleSha256,
-      wasmFlavor: capabilities.runtime.wasmFlavor,
-      wasmSha256: capabilities.runtime.wasmSha256,
-      compatWorkerSha256: capabilities.runtime.compatWorkerSha256,
+      wasmFlavor: loadResult.tuning.observed.wasmFlavor,
+      wasmSha256: loadResult.tuning.observed.wasmSha256,
+      compatWorkerSha256: loadResult.tuning.observed.compatWorkerSha256,
       manifestRepository: observation.manifest.repository.id,
       manifestRevision: observation.manifest.repository.revision,
       crossOriginIsolated: capabilities.crossOriginIsolated,
@@ -288,7 +273,14 @@ export function buildBenchmarkReport(observation: BenchmarkObservation): Benchma
       dflash: {
         enabled: false,
         runtimeAvailable: false,
-        status: observation.manifest.dflash.status,
+        globalStatus: observation.manifest.dflash.status,
+        pairing: dflashPairing.support === 'conditional'
+          ? {
+              ...dflashPairing,
+              targetLayerIdsZeroBased: [...dflashPairing.targetLayerIdsZeroBased],
+              targetTapsOneBased: [...dflashPairing.targetTapsOneBased],
+            }
+          : { ...dflashPairing },
       },
     },
   };

src/engine/device-gate.test.ts

@@ -171,6 +171,10 @@ describe('assertBackendPolicy', () => {
     nGraphSplits: 1,
     opsOnCpu: 0,
     layersGpu: { offloaded: 65, total: 65 },
+    flashAttention: false,
+    cacheTypeK: 'f16',
+    cacheTypeV: 'f16',
+    webgpuKvBufferBytes: 56 * 1024 ** 2,
   };
 
   it('accepts a proven single WebGPU graph', () => {

src/engine/native-log.test.ts

@@ -12,6 +12,9 @@ llama_prepare_model_devices: using device WebGPU (WebGPU) (unknown id)
 load_tensors: offloaded 65/65 layers to GPU
 load_tensors: CPU model buffer size = 72.00 MiB
 load_tensors: WebGPU model buffer size = 3616.77 MiB
+llama_context: flash_attn = disabled
+llama_kv_cache: WebGPU KV buffer size = 56.00 MiB
+llama_kv_cache: size = 56.00 MiB, K (f16): 28.00 MiB, V (f16): 28.00 MiB
 sched_reserve: WebGPU compute buffer size = 301.00 MiB
 sched_reserve: CPU compute buffer size = 13.01 MiB
 sched_reserve: graph splits = 1
@@ -22,6 +25,10 @@ sched_reserve: graph splits = 1
       nGraphSplits: 1,
       opsOnCpu: 0,
       layersGpu: { offloaded: 65, total: 65 },
+      flashAttention: false,
+      cacheTypeK: 'f16',
+      cacheTypeV: 'f16',
+      webgpuKvBufferBytes: 56 * 1024 ** 2,
     });
   });
 
@@ -41,6 +48,34 @@ sched_reserve: graph splits = 1
     expect(parseNativeLog('load_tensors: offloaded 29/29 layers to GPU').nGraphSplits).toBeNull();
   });
 
+  it('reports only resolved Flash Attention state and lets later evidence win', () => {
+    expect(parseNativeLog('llama_context: flash_attn = auto').flashAttention).toBeNull();
+    expect(parseNativeLog([
+      'llama_context: flash_attn = auto',
+      'resolve_fused_ops: Flash Attention enabled',
+    ]).flashAttention).toBe(true);
+    expect(parseNativeLog([
+      'llama_context: flash_attn = enabled',
+      'resolve_fused_ops: Flash Attention not supported, set to disabled',
+    ]).flashAttention).toBe(false);
+  });
+
+  it('retains cache tuning evidence and sums binary WebGPU KV buffer units', () => {
+    const collector = new NativeLogCollector();
+    collector.append(['llama_context: flash_attn = disabled']);
+    collector.append(['llama_kv_cache: K (Q8_0): 1.00 MiB, V (Q4_0): 1.00 MiB']);
+    collector.append(['llama_kv_cache: WebGPU KV buffer size = 512 KiB']);
+    collector.append(['llama_kv_cache: WebGPU KV buffer size = 2.50 MiB']);
+    collector.append(['llama_kv_cache: WebGPU KV buffer size = 1.00 GiB']);
+
+    expect(collector.report()).toMatchObject({
+      flashAttention: false,
+      cacheTypeK: 'q8_0',
+      cacheTypeV: 'q4_0',
+      webgpuKvBufferBytes: (512 * 1024) + (2.5 * 1024 ** 2) + (1024 ** 3),
+    });
+  });
+
   it('parses the native llama.cpp WebGPU device-lost callback', () => {
     expect(parseNativeDeviceLostSignal(
       'ggml_webgpu: Device lost! Reason: 2, Message: GPU process crashed',
@@ -61,6 +96,10 @@ sched_reserve: graph splits = 1
       nGraphSplits: null,
       opsOnCpu: 0,
       layersGpu: null,
+      flashAttention: null,
+      cacheTypeK: null,
+      cacheTypeV: null,
+      webgpuKvBufferBytes: null,
     });
     expect(collector.deviceLostSignal()).toMatchObject({ reason: 1, message: 'unknown' });
 

src/engine/native-log.ts

@@ -8,6 +8,10 @@ export interface BackendReport {
   nGraphSplits: number | null;
   opsOnCpu: number;
   layersGpu: GpuLayerReport | null;
+  flashAttention: boolean | null;
+  cacheTypeK: string | null;
+  cacheTypeV: string | null;
+  webgpuKvBufferBytes: number | null;
 }
 
 export interface NativeDeviceLostSignal {
@@ -22,8 +26,15 @@ export interface NativeDeviceLostWatch {
 }
 
 const BACKEND_NAMES = ['WebGPU', 'CPU', 'Metal', 'CUDA', 'Vulkan'] as const;
-const EVIDENCE = /using device|model buffer size|KV buffer size|compute buffer size|output buffer size|offloaded \d+\/\d+ layers|graph splits\s*=|graph_compute\(|backend\s*=\s*CPU|assigned to (?:device )?CPU|(?:ops?|nodes?).*(?:on|to) CPU/i;
+const EVIDENCE = /using device|model buffer size|KV buffer size|compute buffer size|output buffer size|offloaded \d+\/\d+ layers|graph splits\s*=|graph_compute\(|backend\s*=\s*CPU|assigned to (?:device )?CPU|(?:ops?|nodes?).*(?:on|to) CPU|flash_attn\s*=|flash attention.*(?:enabled|disabled)|\bK\s*\([^)]+\).*\bV\s*\([^)]+\)/i;
 const WEBGPU_DEVICE_LOST = /\bdevice\s+lost!?(?:\s+reason:\s*(-?\d+),\s*message:\s*(.*))?/i;
+const WEBGPU_KV_BUFFER = /\bWebGPU\s+KV buffer size\s*=\s*(\d+(?:\.\d+)?)\s*(KiB|MiB|GiB)\b/i;
+
+const BINARY_UNIT_BYTES = {
+  kib: 1024,
+  mib: 1024 ** 2,
+  gib: 1024 ** 3,
+} as const;
 
 function canonicalBackend(value: string): string {
   const match = BACKEND_NAMES.find((backend) => value.toLowerCase().startsWith(backend.toLowerCase()));
@@ -76,6 +87,10 @@ export function parseNativeLog(input: string | readonly string[]): BackendReport
   const graphSplits: number[] = [];
   let opsOnCpu = 0;
   let layersGpu: GpuLayerReport | null = null;
+  let flashAttention: boolean | null = null;
+  let cacheTypeK: string | null = null;
+  let cacheTypeV: string | null = null;
+  let webgpuKvBufferBytes: number | null = null;
 
   for (const line of lines) {
     const deviceMatch = line.match(/using device\s+([^\s(]+)\s+\(([^)]+)\)/i);
@@ -102,6 +117,30 @@ export function parseNativeLog(input: string | readonly string[]): BackendReport
     if (graphMatch) {
       graphSplits.push(Number(graphMatch[1]));
     }
+
+    const contextFlashMatch = line.match(/\bflash_attn\s*=\s*(enabled|disabled)\b/i);
+    const resolvedFlashMatch = line.match(/\bFlash Attention\b.*\b(enabled|disabled)\b/i);
+    const flashState = (contextFlashMatch ?? resolvedFlashMatch)?.[1];
+    if (flashState) {
+      flashAttention = flashState.toLowerCase() === 'enabled';
+    }
+
+    const cacheTypesMatch = line.match(/\bK\s*\(\s*([^)]+?)\s*\).*\bV\s*\(\s*([^)]+?)\s*\)/i);
+    const cacheK = cacheTypesMatch?.[1];
+    const cacheV = cacheTypesMatch?.[2];
+    if (cacheK && cacheV) {
+      cacheTypeK = cacheK.trim().toLowerCase();
+      cacheTypeV = cacheV.trim().toLowerCase();
+    }
+
+    const kvBufferMatch = line.match(WEBGPU_KV_BUFFER);
+    const kvBufferSize = kvBufferMatch?.[1];
+    const kvBufferUnit = kvBufferMatch?.[2];
+    if (kvBufferSize && kvBufferUnit) {
+      const unitBytes = BINARY_UNIT_BYTES[kvBufferUnit.toLowerCase() as keyof typeof BINARY_UNIT_BYTES];
+      const bytes = Math.round(Number(kvBufferSize) * unitBytes);
+      webgpuKvBufferBytes = (webgpuKvBufferBytes ?? 0) + bytes;
+    }
     opsOnCpu += cpuOpsInLine(line);
   }
 
@@ -110,6 +149,10 @@ export function parseNativeLog(input: string | readonly string[]): BackendReport
     nGraphSplits: graphSplits.length > 0 ? Math.max(...graphSplits) : null,
     opsOnCpu,
     layersGpu,
+    flashAttention,
+    cacheTypeK,
+    cacheTypeV,
+    webgpuKvBufferBytes,
   };
 }
 

src/engine/protocol.ts

@@ -64,6 +64,18 @@ export interface FutureTensorPlacement {
   }>;
 }
 
+export type BenchmarkFlashMode = 'off' | 'auto';
+export type BenchmarkKvCacheType = 'f16' | 'q8_0' | 'q4_0';
+export type BenchmarkWasmFlavor = 'auto' | 'jspi' | 'compat';
+
+export interface BenchmarkLoadTuning {
+  nBatch?: number;
+  nUbatch?: number;
+  flashMode: BenchmarkFlashMode;
+  kvCacheType: BenchmarkKvCacheType;
+  wasmFlavor: BenchmarkWasmFlavor;
+}
+
 export interface EngineChatToolCall {
   id: string;
   type: 'function';
@@ -106,11 +118,7 @@ export type LoadModelParams = ManifestSource & {
   backend: RequestedBackend;
   contextSize?: number;
   threads?: number;
-  nBatch?: number;
-  nUbatch?: number;
-  flashAttention?: boolean;
-  cacheTypeK?: 'f32' | 'f16' | 'q8_0' | 'q5_1' | 'q5_0' | 'q4_1' | 'q4_0';
-  cacheTypeV?: 'f32' | 'f16' | 'q8_0' | 'q5_1' | 'q5_0' | 'q4_1' | 'q4_0';
+  benchmarkTuning?: BenchmarkLoadTuning;
   tensorPlacement?: FutureTensorPlacement;
 };
 
@@ -141,9 +149,27 @@ export interface LoadModelResult {
     microBatchSize: number;
   };
   tuning: {
-    flashAttention: boolean;
-    cacheTypeK: NonNullable<LoadModelParams['cacheTypeK']>;
-    cacheTypeV: NonNullable<LoadModelParams['cacheTypeV']>;
+    scope: 'release-defaults' | 'benchmark';
+    requested: {
+      nBatch: number | null;
+      nUbatch: number | null;
+      flashMode: BenchmarkFlashMode;
+      cacheTypeK: BenchmarkKvCacheType;
+      cacheTypeV: BenchmarkKvCacheType;
+      wasmFlavor: BenchmarkWasmFlavor;
+    };
+    observed: {
+      nBatch: number;
+      nUbatch: number;
+      flashAttention: boolean | null;
+      cacheTypeK: string | null;
+      cacheTypeV: string | null;
+      kvBufferBytes: number | null;
+      wasmFlavor: 'jspi' | 'compat';
+      wasmSha256: string;
+      compatWorkerSha256: string | null;
+    };
+    applied: boolean;
   };
   chatTemplate: {
     bytes: number;

src/engine/runtime-device-lost.test.ts

@@ -66,6 +66,10 @@ describe('BrowserEngineRuntime WebGPU device loss', () => {
       nGraphSplits: null,
       opsOnCpu: 0,
       layersGpu: null,
+      flashAttention: null,
+      cacheTypeK: null,
+      cacheTypeV: null,
+      webgpuKvBufferBytes: null,
     });
     expect(exit).not.toHaveBeenCalled();
     expect(internals.loaded).not.toBeNull();

src/engine/runtime-download.test.ts

@@ -4,7 +4,7 @@ import { verifyBlobSha256 } from './blob-integrity';
 import { EngineRuntimeError } from './errors';
 import type { ManifestModelV2 } from './manifest';
 import { isShardDownloadFailureDetails, type EngineEvent } from './protocol';
-import { BrowserEngineRuntime } from './runtime';
+import { BrowserEngineRuntime, resolveLoadTuning, selectWasmFlavor } from './runtime';
 
 type DownloadOptions = {
   signal?: AbortSignal;
@@ -313,29 +313,169 @@ describe('BrowserEngineRuntime per-shard retry', () => {
   });
 });
 
-describe('BrowserEngineRuntime benchmark load controls', () => {
-  const invalidLoad = (extra: { nBatch?: number; nUbatch?: number }) => new BrowserEngineRuntime().loadModel(
-    'request-invalid-batch',
-    {
-      manifest: { schemaVersion: 2 } as never,
-      modelId: '1_7b',
-      backend: 'webgpu',
-      ...extra,
-    },
-    new AbortController().signal,
-    () => undefined,
-  );
+describe('resolveLoadTuning', () => {
+  const benchmarkDefaults = {
+    flashMode: 'off',
+    kvCacheType: 'f16',
+    wasmFlavor: 'auto',
+  } as const;
+
+  function tuningError(
+    input: unknown,
+    backend: 'auto' | 'webgpu' | 'wasm' = 'webgpu',
+  ): EngineRuntimeError {
+    try {
+      resolveLoadTuning(input, backend);
+    } catch (error) {
+      if (error instanceof EngineRuntimeError) return error;
+      throw error;
+    }
+    throw new Error('Expected benchmark tuning to be rejected.');
+  }
+
+  it('uses immutable release defaults when benchmark tuning is omitted', () => {
+    expect(resolveLoadTuning(undefined, 'auto')).toEqual({
+      scope: 'release-defaults',
+      nBatch: null,
+      nUbatch: null,
+      flashMode: 'off',
+      kvCacheType: 'f16',
+      wasmFlavor: 'auto',
+    });
+    expect(resolveLoadTuning(undefined, 'wasm')).toEqual({
+      scope: 'release-defaults',
+      nBatch: null,
+      nUbatch: null,
+      flashMode: 'off',
+      kvCacheType: 'f16',
+      wasmFlavor: 'auto',
+    });
+  });
+
+  it('accepts an explicit benchmark batch shape without changing release defaults', () => {
+    expect(resolveLoadTuning({
+      ...benchmarkDefaults,
+      nBatch: 512,
+      nUbatch: 128,
+    }, 'auto')).toEqual({
+      scope: 'benchmark',
+      nBatch: 512,
+      nUbatch: 128,
+      flashMode: 'off',
+      kvCacheType: 'f16',
+      wasmFlavor: 'auto',
+    });
+  });
 
-  it('rejects invalid n_batch before touching the manifest or model cache', async () => {
-    await expect(invalidLoad({ nBatch: 0 })).rejects.toMatchObject({ code: 'INVALID_BATCH_SIZE' });
+  it('rejects a microbatch larger than the requested batch', () => {
+    expect(tuningError({
+      ...benchmarkDefaults,
+      nBatch: 128,
+      nUbatch: 256,
+    })).toMatchObject({
+      code: 'INVALID_BENCHMARK_TUNING',
+      message: 'Benchmark n_ubatch must not exceed n_batch.',
+    });
   });
 
-  it('rejects invalid n_ubatch before touching the manifest or model cache', async () => {
-    await expect(invalidLoad({ nUbatch: 1.5 })).rejects.toMatchObject({ code: 'INVALID_UBATCH_SIZE' });
+  it('accepts Flash Attention and quantized KV only on the explicit WebGPU backend', () => {
+    expect(resolveLoadTuning({
+      ...benchmarkDefaults,
+      flashMode: 'auto',
+    }, 'webgpu')).toMatchObject({
+      scope: 'benchmark',
+      flashMode: 'auto',
+      kvCacheType: 'f16',
+      wasmFlavor: 'auto',
+    });
+    for (const kvCacheType of ['q8_0', 'q4_0'] as const) {
+      expect(resolveLoadTuning({
+        ...benchmarkDefaults,
+        flashMode: 'auto',
+        kvCacheType,
+      }, 'webgpu')).toMatchObject({
+        scope: 'benchmark',
+        flashMode: 'auto',
+        kvCacheType,
+        wasmFlavor: 'auto',
+      });
+    }
+
+    expect(tuningError({
+      ...benchmarkDefaults,
+      flashMode: 'auto',
+    }, 'auto')).toMatchObject({
+      code: 'INVALID_BENCHMARK_TUNING',
+    });
+    expect(tuningError({
+      ...benchmarkDefaults,
+      flashMode: 'auto',
+      kvCacheType: 'q8_0',
+    }, 'wasm')).toMatchObject({
+      code: 'INVALID_BENCHMARK_TUNING',
+    });
   });
 
-  it('rejects batch values that overflow the signed glue integer', async () => {
-    await expect(invalidLoad({ nBatch: 2_147_483_648 })).rejects.toMatchObject({ code: 'INVALID_BATCH_SIZE' });
-    await expect(invalidLoad({ nUbatch: Number.MAX_SAFE_INTEGER })).rejects.toMatchObject({ code: 'INVALID_UBATCH_SIZE' });
+  it('rejects malformed and incomplete benchmark tuning objects', () => {
+    for (const input of [null, [], 'off', 42]) {
+      expect(tuningError(input)).toMatchObject({ code: 'INVALID_BENCHMARK_TUNING' });
+    }
+    for (const input of [
+      { kvCacheType: 'f16' },
+      { flashMode: 'off' },
+      { flashMode: 'off', kvCacheType: 'f16' },
+      { ...benchmarkDefaults, flashMode: 'enabled' },
+      { ...benchmarkDefaults, kvCacheType: 'q5_0' },
+      { ...benchmarkDefaults, wasmFlavor: 'future' },
+    ]) {
+      expect(tuningError(input)).toMatchObject({ code: 'INVALID_BENCHMARK_TUNING' });
+    }
+  });
+
+  it('rejects quantized V cache unless Flash Attention is in auto mode', () => {
+    for (const kvCacheType of ['q8_0', 'q4_0'] as const) {
+      expect(tuningError({ ...benchmarkDefaults, kvCacheType })).toMatchObject({
+        code: 'INVALID_BENCHMARK_TUNING',
+        message: expect.stringMatching(/requires Flash Attention/i),
+      });
+    }
+  });
+
+  it('rejects invalid and overflowing signed glue batch values', () => {
+    for (const nBatch of [0, -1, 1.5, Number.NaN, Number.POSITIVE_INFINITY, 2_147_483_648]) {
+      expect(tuningError({ ...benchmarkDefaults, nBatch })).toMatchObject({
+        code: 'INVALID_BATCH_SIZE',
+      });
+    }
+    for (const nUbatch of [0, -1, 1.5, Number.NaN, Number.POSITIVE_INFINITY, Number.MAX_SAFE_INTEGER]) {
+      expect(tuningError({ ...benchmarkDefaults, nUbatch })).toMatchObject({
+        code: 'INVALID_UBATCH_SIZE',
+      });
+    }
+  });
+});
+
+describe('selectWasmFlavor', () => {
+  it('uses the browser-native flavor for auto', () => {
+    expect(selectWasmFlavor('auto', false)).toBe('jspi');
+    expect(selectWasmFlavor('auto', true)).toBe('compat');
+  });
+
+  it('forces compat even when JSPI is available', () => {
+    expect(selectWasmFlavor('compat', false)).toBe('compat');
+    expect(selectWasmFlavor('compat', true)).toBe('compat');
+  });
+
+  it('selects JSPI when the browser supports it', () => {
+    expect(selectWasmFlavor('jspi', false)).toBe('jspi');
+  });
+
+  it('fails loud when JSPI is requested but unavailable', () => {
+    expect(() => selectWasmFlavor('jspi', true)).toThrowError(
+      expect.objectContaining({
+        code: 'BENCHMARK_WASM_FLAVOR_UNAVAILABLE',
+        message: expect.stringMatching(/requires the compatibility runtime/i),
+      }),
+    );
   });
 });

src/engine/runtime.ts

@@ -31,6 +31,9 @@ import {
 import { NativeLogCollector, type BackendReport } from './native-log';
 import { ToolCallAccumulator } from './tool-call-accumulator';
 import type {
+  BenchmarkFlashMode,
+  BenchmarkKvCacheType,
+  BenchmarkWasmFlavor,
   EngineCapabilities,
   EngineEvent,
   GenerateParams,
@@ -50,6 +53,95 @@ const WLLAMA_COMPAT_WORKER_PATH = '/wasm/wllama-compat.js';
 const MAX_GLUE_INT = 2_147_483_647;
 const DEVICE_LOST_EXIT_GRACE_MS = 100;
 
+export interface ResolvedLoadTuning {
+  scope: 'release-defaults' | 'benchmark';
+  nBatch: number | null;
+  nUbatch: number | null;
+  flashMode: BenchmarkFlashMode;
+  kvCacheType: BenchmarkKvCacheType;
+  wasmFlavor: BenchmarkWasmFlavor;
+}
+
+function validateBatchValue(value: unknown, label: 'n_batch' | 'n_ubatch'): number | null {
+  if (value === undefined) return null;
+  if (!Number.isSafeInteger(value) || (value as number) <= 0 || (value as number) > MAX_GLUE_INT) {
+    throw new EngineRuntimeError(
+      label === 'n_batch' ? 'INVALID_BATCH_SIZE' : 'INVALID_UBATCH_SIZE',
+      `${label} must be a positive signed 32-bit integer (maximum ${MAX_GLUE_INT}).`,
+    );
+  }
+  return value as number;
+}
+
+export function resolveLoadTuning(
+  input: unknown,
+  backend: LoadModelParams['backend'],
+): ResolvedLoadTuning {
+  if (input === undefined) {
+    return {
+      scope: 'release-defaults',
+      nBatch: null,
+      nUbatch: null,
+      flashMode: 'off',
+      kvCacheType: 'f16',
+      wasmFlavor: 'auto',
+    };
+  }
+  if (typeof input !== 'object' || input === null || Array.isArray(input)) {
+    throw new EngineRuntimeError('INVALID_BENCHMARK_TUNING', 'Benchmark tuning must be an object.');
+  }
+  const tuning = input as Record<string, unknown>;
+  const flashMode = tuning.flashMode;
+  const kvCacheType = tuning.kvCacheType;
+  const wasmFlavor = tuning.wasmFlavor;
+  if (flashMode !== 'off' && flashMode !== 'auto') {
+    throw new EngineRuntimeError(
+      'INVALID_BENCHMARK_TUNING',
+      'Benchmark flash mode must be off or auto.',
+    );
+  }
+  if (kvCacheType !== 'f16' && kvCacheType !== 'q8_0' && kvCacheType !== 'q4_0') {
+    throw new EngineRuntimeError(
+      'INVALID_BENCHMARK_TUNING',
+      'Benchmark KV cache type must be f16, q8_0, or q4_0.',
+    );
+  }
+  if (wasmFlavor !== 'auto' && wasmFlavor !== 'jspi' && wasmFlavor !== 'compat') {
+    throw new EngineRuntimeError(
+      'INVALID_BENCHMARK_TUNING',
+      'Benchmark WASM flavor must be auto, jspi, or compat.',
+    );
+  }
+  if (kvCacheType !== 'f16' && flashMode !== 'auto') {
+    throw new EngineRuntimeError(
+      'INVALID_BENCHMARK_TUNING',
+      'Quantized V cache requires Flash Attention in auto mode.',
+    );
+  }
+  if ((flashMode === 'auto' || kvCacheType !== 'f16') && backend !== 'webgpu') {
+    throw new EngineRuntimeError(
+      'INVALID_BENCHMARK_TUNING',
+      'Flash Attention and quantized KV experiments require the explicit WebGPU backend.',
+    );
+  }
+  const nBatch = validateBatchValue(tuning.nBatch, 'n_batch');
+  const nUbatch = validateBatchValue(tuning.nUbatch, 'n_ubatch');
+  if (nBatch !== null && nUbatch !== null && nUbatch > nBatch) {
+    throw new EngineRuntimeError(
+      'INVALID_BENCHMARK_TUNING',
+      'Benchmark n_ubatch must not exceed n_batch.',
+    );
+  }
+  return {
+    scope: 'benchmark',
+    nBatch,
+    nUbatch,
+    flashMode,
+    kvCacheType,
+    wasmFlavor,
+  };
+}
+
 function runtimeArtifact(path: string): { bytes: number; sha256: string } {
   const artifact = runtimeSource.files.find((file) => file.path === path);
   if (!artifact) throw new Error(`Runtime provenance is missing ${path}`);
@@ -68,6 +160,23 @@ function needsCompatWasm(): boolean {
   }
 }
 
+export function selectWasmFlavor(
+  requested: BenchmarkWasmFlavor,
+  compatRequired = needsCompatWasm(),
+): 'jspi' | 'compat' {
+  if (requested === 'compat') return 'compat';
+  if (requested === 'jspi') {
+    if (compatRequired) {
+      throw new EngineRuntimeError(
+        'BENCHMARK_WASM_FLAVOR_UNAVAILABLE',
+        'JSPI was requested, but this browser requires the compatibility runtime.',
+      );
+    }
+    return 'jspi';
+  }
+  return compatRequired ? 'compat' : 'jspi';
+}
+
 type EventSink = (event: EngineEvent) => void;
 
 interface LoadedModelState {
@@ -215,6 +324,7 @@ function asReasoningDelta(chunk: ChatCompletionChunk): string | undefined {
 
 export class BrowserEngineRuntime {
   private wllama: Wllama | null = null;
+  private wllamaFlavor: 'jspi' | 'compat' | null = null;
   private compatWorkerCode: string | null = null;
   private loaded: LoadedModelState | null = null;
   private readonly nativeLog = new NativeLogCollector();
@@ -253,6 +363,7 @@ export class BrowserEngineRuntime {
     const invalidatedWllama = this.wllama;
     this.loaded = null;
     this.wllama = null;
+    this.wllamaFlavor = null;
     if (invalidatedWllama) {
       let timeout: ReturnType<typeof setTimeout> | undefined;
       try {
@@ -287,10 +398,16 @@ export class BrowserEngineRuntime {
     );
   }
 
-  private async ensureWllama(): Promise<Wllama> {
-    if (this.wllama) {
+  private async ensureWllama(flavor = selectWasmFlavor('auto')): Promise<Wllama> {
+    if (this.wllama && this.wllamaFlavor === flavor) {
       return this.wllama;
     }
+    if (this.wllama) {
+      await this.wllama.exit().catch(() => undefined);
+      this.wllama = null;
+      this.loaded = null;
+      this.wllamaFlavor = null;
+    }
     installWorkerDocumentShim();
     if (this.compatWorkerCode === null) {
       const response = await fetch(absoluteAssetUrl(WLLAMA_COMPAT_WORKER_PATH));
@@ -306,6 +423,7 @@ export class BrowserEngineRuntime {
       { default: absoluteAssetUrl(WLLAMA_WASM_PATH) },
       {
         parallelDownloads: 3,
+        forceCompat: flavor === 'compat',
         logger: {
           debug: (...values: unknown[]) => this.nativeLog.append(values),
           log: (...values: unknown[]) => this.nativeLog.append(values),
@@ -319,6 +437,7 @@ export class BrowserEngineRuntime {
       worker: { code: this.compatWorkerCode },
     }, 'firefox_safari');
     this.wllama = wllama;
+    this.wllamaFlavor = flavor;
     return wllama;
   }
 
@@ -541,26 +660,8 @@ export class BrowserEngineRuntime {
         params.tensorPlacement,
       );
     }
-    if (params.nBatch !== undefined && (
-      !Number.isSafeInteger(params.nBatch)
-      || params.nBatch <= 0
-      || params.nBatch > MAX_GLUE_INT
-    )) {
-      throw new EngineRuntimeError(
-        'INVALID_BATCH_SIZE',
-        `n_batch must be a positive signed 32-bit integer (maximum ${MAX_GLUE_INT}).`,
-      );
-    }
-    if (params.nUbatch !== undefined && (
-      !Number.isSafeInteger(params.nUbatch)
-      || params.nUbatch <= 0
-      || params.nUbatch > MAX_GLUE_INT
-    )) {
-      throw new EngineRuntimeError(
-        'INVALID_UBATCH_SIZE',
-        `n_ubatch must be a positive signed 32-bit integer (maximum ${MAX_GLUE_INT}).`,
-      );
-    }
+    const tuning = resolveLoadTuning(params.benchmarkTuning, params.backend);
+    const wasmFlavor = selectWasmFlavor(tuning.wasmFlavor);
     await this.unload();
     this.nativeLog.clear();
     emitProgress(sink, requestId, {
@@ -585,7 +686,7 @@ export class BrowserEngineRuntime {
     });
     const model = findManifestModel(manifest, params.modelId);
     const urls = orderedShardUrls(manifest, model);
-    const wllama = await this.ensureWllama();
+    const wllama = await this.ensureWllama(wasmFlavor);
     const cached = await this.inspectCachedShards(wllama, model, urls, signal);
     const [adapter, storage] = await Promise.all([
       inspectWebGpuAdapter(),
@@ -630,13 +731,13 @@ export class BrowserEngineRuntime {
         offload_token_embedding: gate.selectedBackend === 'webgpu',
         n_ctx: contextSize,
         n_threads: threads,
-        n_batch: params.nBatch,
-        n_ubatch: params.nUbatch,
+        n_batch: tuning.nBatch ?? undefined,
+        n_ubatch: tuning.nUbatch ?? undefined,
         seed: 42,
-        flash_attn: params.flashAttention ?? false,
+        flash_attn: tuning.flashMode === 'auto',
         warmup: false,
-        cache_type_k: params.cacheTypeK ?? 'f16',
-        cache_type_v: params.cacheTypeV ?? 'f16',
+        cache_type_k: tuning.kvCacheType,
+        cache_type_v: tuning.kvCacheType,
       }));
       await this.assertWebGpuAlive('load', gate.selectedBackend, model);
       throwIfAborted(signal);
@@ -652,6 +753,54 @@ export class BrowserEngineRuntime {
       }
       const context = wllama.getLoadedContextInfo();
       const template = wllama.getChatTemplate() ?? '';
+      const wasmArtifact = runtimeArtifact(
+        wasmFlavor === 'compat' ? 'public/wasm/wllama-compat.wasm' : 'public/wasm/wllama.wasm',
+      );
+      const compatWorkerSha256 = wasmFlavor === 'compat'
+        ? runtimeArtifact('public/wasm/wllama-compat.js').sha256
+        : null;
+      const tuningApplied = (tuning.nBatch === null || context.n_batch === tuning.nBatch)
+        && (tuning.nUbatch === null || context.n_ubatch === tuning.nUbatch)
+        && report.flashAttention === (tuning.flashMode === 'auto')
+        && report.cacheTypeK === tuning.kvCacheType
+        && report.cacheTypeV === tuning.kvCacheType
+        && (gate.selectedBackend !== 'webgpu'
+          || (report.webgpuKvBufferBytes !== null && report.webgpuKvBufferBytes > 0))
+        && (tuning.wasmFlavor === 'auto' || tuning.wasmFlavor === wasmFlavor);
+      const tuningResult: LoadModelResult['tuning'] = {
+        scope: tuning.scope,
+        requested: {
+          nBatch: tuning.nBatch,
+          nUbatch: tuning.nUbatch,
+          flashMode: tuning.flashMode,
+          cacheTypeK: tuning.kvCacheType,
+          cacheTypeV: tuning.kvCacheType,
+          wasmFlavor: tuning.wasmFlavor,
+        },
+        observed: {
+          nBatch: context.n_batch,
+          nUbatch: context.n_ubatch,
+          flashAttention: report.flashAttention,
+          cacheTypeK: report.cacheTypeK,
+          cacheTypeV: report.cacheTypeV,
+          kvBufferBytes: report.webgpuKvBufferBytes,
+          wasmFlavor,
+          wasmSha256: wasmArtifact.sha256,
+          compatWorkerSha256,
+        },
+        applied: tuningApplied,
+      };
+      if (!tuningApplied) {
+        throw new EngineRuntimeError(
+          tuning.scope === 'benchmark'
+            ? 'BENCHMARK_TUNING_NOT_APPLIED'
+            : 'RUNTIME_POLICY_NOT_APPLIED',
+          tuning.scope === 'benchmark'
+            ? 'The native runtime did not apply the requested benchmark tuning exactly.'
+            : 'The native runtime did not preserve the fixed release tuning policy.',
+          tuningResult,
+        );
+      }
       this.loaded = { manifest, model, backend: gate.selectedBackend };
       emitProgress(sink, requestId, {
         phase: 'load',
@@ -674,11 +823,7 @@ export class BrowserEngineRuntime {
           batchSize: context.n_batch,
           microBatchSize: context.n_ubatch,
         },
-        tuning: {
-          flashAttention: params.flashAttention ?? false,
-          cacheTypeK: params.cacheTypeK ?? 'f16',
-          cacheTypeV: params.cacheTypeV ?? 'f16',
-        },
+        tuning: tuningResult,
         chatTemplate: {
           bytes: new TextEncoder().encode(template).byteLength,
           hasThinkMarker: template.includes('<think>'),
@@ -691,6 +836,8 @@ export class BrowserEngineRuntime {
       if (error instanceof EngineRuntimeError && error.code === 'WEBGPU_DEVICE_LOST') throw error;
       await this.assertWebGpuAlive('load', gate.selectedBackend, model);
       await wllama.exit().catch(() => undefined);
+      this.wllama = null;
+      this.wllamaFlavor = null;
       this.loaded = null;
       throw error;
     }

vendor/wllama-bonsai/PATCH.diff

@@ -119,10 +119,40 @@ index 2b3cca7..e91b395 100644
    n_threads?: number;
    embeddings?: boolean;
 diff --git a/src/wllama.ts b/src/wllama.ts
-index ce74b2b..8057028 100644
+index ce74b2b..1796ca1 100644
 --- a/src/wllama.ts
 +++ b/src/wllama.ts
-@@ -478,6 +478,27 @@ export class Wllama {
+@@ -55,6 +55,13 @@ export interface WllamaLogger {
+ 
+ // TODO: bring back useCache
+ export interface WllamaConfig {
++  /**
++   * Always use the configured compat worker JS and WASM pair, even when the
++   * browser supports the default JSPI runtime.
++   *
++   * Default: forceCompat = false
++   */
++  forceCompat?: boolean;
+   /**
+    * If true, suppress all log messages from native CPP code
+    */
+@@ -238,13 +245,13 @@ export class Wllama {
+   /**
+    * Set compatibility options for Wllama.
+    * @param compat Set to null to disable compatibility, or 'default' to use the default compat resources from CDN.
+-   * @param mode 'safari' by default; If set to 'firefox_safari', the compat mode will **also** be enabled on Firefox, which will significantly degrade the performance but allow using WebGPU on Firefox.
++   * @param mode 'safari' by default; If set to 'firefox_safari', the compat mode will **also** be enabled on Firefox, which will significantly degrade the performance but allow using WebGPU on Firefox. Ignored when forceCompat is true.
+    */
+   setCompat(
+     compat: WllamaCompat | null | 'default',
+     mode: 'safari' | 'firefox_safari' = 'safari'
+   ) {
+-    if (mode === 'safari') {
++    if (mode === 'safari' && !this.config.forceCompat) {
+       if (isFirefox()) {
+         this.compat = null;
+         return;
+@@ -478,6 +485,27 @@ export class Wllama {
      if (this.proxy) {
        throw new WllamaError('Module is already initialized', 'load_error');
      }
@@ -150,7 +180,7 @@ index ce74b2b..8057028 100644
      // detect if we can use multi-thread and webgpu
      const supportMultiThread = await isSupportMultiThread();
      const hwConccurency = Math.floor((navigator.hardwareConcurrency || 1) / 2);
-@@ -518,7 +539,8 @@ export class Wllama {
+@@ -518,7 +546,8 @@ export class Wllama {
        // if async read is not supported, use mmap; refer to README-dev.md for more details
        use_mmap: !canUseAsyncFileRead(workerResources.compat),
        use_mlock: false,
@@ -160,7 +190,7 @@ index ce74b2b..8057028 100644
        n_ctx: params.n_ctx ?? 1024,
        n_threads: this.useMultiThread ? nbThreads : 1,
        n_ctx_auto: false, // not supported for now
-@@ -529,6 +551,7 @@ export class Wllama {
+@@ -529,6 +558,7 @@ export class Wllama {
        embeddings: params.embeddings,
        offload_kqv: params.offload_kqv,
        n_batch: params.n_batch,
@@ -168,7 +198,7 @@ index ce74b2b..8057028 100644
        pooling_type: params.pooling_type as string,
        rope_scaling_type: params.rope_scaling_type as string,
        rope_freq_base: params.rope_freq_base,
-@@ -579,6 +602,10 @@ export class Wllama {
+@@ -579,6 +609,10 @@ export class Wllama {
        skip_chat_parsing: params.skip_chat_parsing,
        prefill_assistant: params.prefill_assistant,
      });
@@ -179,3 +209,29 @@ index ce74b2b..8057028 100644
      const loadedCtxInfo: LoadedContextInfo & GlueMsgLoadRes = {
        ...loadResult,
        metadata: {},
+@@ -1094,7 +1128,24 @@ export class Wllama {
+       wasmPath: absoluteUrl(this.pathConfig['default']),
+       compat: false,
+     };
+-    if (needCompat()) {
++    const forceCompat = this.config.forceCompat === true;
++    if (forceCompat || needCompat()) {
++      const compatWorker = this.compat?.worker;
++      const hasCompatWorker =
++        (typeof compatWorker === 'string' && compatWorker.trim().length > 0) ||
++        (typeof compatWorker === 'object' &&
++          compatWorker !== null &&
++          typeof compatWorker.code === 'string' &&
++          compatWorker.code.trim().length > 0);
++      const compatWasm = this.compat?.wasm;
++      const hasCompatWasm =
++        typeof compatWasm === 'string' && compatWasm.trim().length > 0;
++      if (forceCompat && (!hasCompatWorker || !hasCompatWasm)) {
++        throw new WllamaError(
++          'forceCompat requires both compat worker JS and compat WASM resources. Configure them with setCompat() before loading.',
++          'load_error'
++        );
++      }
+       if (!this.compat) {
+         this.logger().warn(
+           'Not using compat mode' +

vendor/wllama-bonsai/SOURCE.json

@@ -4,12 +4,12 @@
   "wllamaRevision": "912c18b75d4358c1405a64646b8dbe43a205943b",
   "llamaCppRevision": "00fa7cb284cbf133fc426733bd64238a3588a33e",
   "license": "MIT",
-  "patch": "Expose offload_token_embedding and n_ubatch; place token_embd.weight on WebGPU whenever GPU layers are requested; fail loudly when the requested WebGPU buffer is unavailable.",
+  "patch": "Expose offload_token_embedding and n_ubatch; place token_embd.weight on WebGPU whenever GPU layers are requested; fail loudly when the requested WebGPU buffer is unavailable; add an opt-in forceCompat selector for controlled benchmark A/B runs.",
   "patchSet": {
     "format": "git-diff-binary",
     "path": "vendor/wllama-bonsai/PATCH.diff",
-    "bytes": 6896,
-    "sha256": "62efebdb3071ba49b87c835c1b45716257fe5ea6cb7eec6ee82f6a2f33bc581c",
+    "bytes": 9350,
+    "sha256": "ad8786a295eaeb75ee03752d7f84d30f14ea373407844d5c589f7ab46800bb4b",
     "files": [
       "CMakeLists.txt",
       "cpp/glue.hpp",
@@ -28,8 +28,8 @@
   "files": [
     {
       "path": "vendor/wllama-bonsai/esm/index.js",
-      "bytes": 364498,
-      "sha256": "67eb16e67297b05e8fc7424d9b273d1586ad8097f8551216b55fa4b77d5e5c51"
+      "bytes": 365397,
+      "sha256": "24bcef8aea8e27fb7b7e2d9e6ea94ba8ced7bfbffd09a0675821b9eb1b4a4c9f"
     },
     {
       "path": "public/wasm/wllama.wasm",

vendor/wllama-bonsai/esm/index.cjs

@@ -2650,10 +2650,10 @@ var Wllama = class {
   /**
    * Set compatibility options for Wllama.
    * @param compat Set to null to disable compatibility, or 'default' to use the default compat resources from CDN.
-   * @param mode 'safari' by default; If set to 'firefox_safari', the compat mode will **also** be enabled on Firefox, which will significantly degrade the performance but allow using WebGPU on Firefox.
+   * @param mode 'safari' by default; If set to 'firefox_safari', the compat mode will **also** be enabled on Firefox, which will significantly degrade the performance but allow using WebGPU on Firefox. Ignored when forceCompat is true.
    */
   setCompat(compat, mode = "safari") {
-    if (mode === "safari") {
+    if (mode === "safari" && !this.config.forceCompat) {
       if (isFirefox()) {
         this.compat = null;
         return;
@@ -3395,11 +3395,23 @@ var Wllama = class {
     });
   }
   getWorkerResources() {
+    var _a, _b;
     const workerResources = {
       wasmPath: absoluteUrl(this.pathConfig["default"]),
       compat: false
     };
-    if (needCompat()) {
+    const forceCompat = this.config.forceCompat === true;
+    if (forceCompat || needCompat()) {
+      const compatWorker = (_a = this.compat) == null ? void 0 : _a.worker;
+      const hasCompatWorker = typeof compatWorker === "string" && compatWorker.trim().length > 0 || typeof compatWorker === "object" && compatWorker !== null && typeof compatWorker.code === "string" && compatWorker.code.trim().length > 0;
+      const compatWasm = (_b = this.compat) == null ? void 0 : _b.wasm;
+      const hasCompatWasm = typeof compatWasm === "string" && compatWasm.trim().length > 0;
+      if (forceCompat && (!hasCompatWorker || !hasCompatWasm)) {
+        throw new WllamaError(
+          "forceCompat requires both compat worker JS and compat WASM resources. Configure them with setCompat() before loading.",
+          "load_error"
+        );
+      }
       if (!this.compat) {
         this.logger().warn(
           "Not using compat mode" + (isFirefox() ? " (expected on Firefox - WebGPU will be disabled)" : "")

vendor/wllama-bonsai/esm/index.js

@@ -2615,10 +2615,10 @@ var Wllama = class {
   /**
    * Set compatibility options for Wllama.
    * @param compat Set to null to disable compatibility, or 'default' to use the default compat resources from CDN.
-   * @param mode 'safari' by default; If set to 'firefox_safari', the compat mode will **also** be enabled on Firefox, which will significantly degrade the performance but allow using WebGPU on Firefox.
+   * @param mode 'safari' by default; If set to 'firefox_safari', the compat mode will **also** be enabled on Firefox, which will significantly degrade the performance but allow using WebGPU on Firefox. Ignored when forceCompat is true.
    */
   setCompat(compat, mode = "safari") {
-    if (mode === "safari") {
+    if (mode === "safari" && !this.config.forceCompat) {
       if (isFirefox()) {
         this.compat = null;
         return;
@@ -3360,11 +3360,23 @@ var Wllama = class {
     });
   }
   getWorkerResources() {
+    var _a, _b;
     const workerResources = {
       wasmPath: absoluteUrl(this.pathConfig["default"]),
       compat: false
     };
-    if (needCompat()) {
+    const forceCompat = this.config.forceCompat === true;
+    if (forceCompat || needCompat()) {
+      const compatWorker = (_a = this.compat) == null ? void 0 : _a.worker;
+      const hasCompatWorker = typeof compatWorker === "string" && compatWorker.trim().length > 0 || typeof compatWorker === "object" && compatWorker !== null && typeof compatWorker.code === "string" && compatWorker.code.trim().length > 0;
+      const compatWasm = (_b = this.compat) == null ? void 0 : _b.wasm;
+      const hasCompatWasm = typeof compatWasm === "string" && compatWasm.trim().length > 0;
+      if (forceCompat && (!hasCompatWorker || !hasCompatWasm)) {
+        throw new WllamaError(
+          "forceCompat requires both compat worker JS and compat WASM resources. Configure them with setCompat() before loading.",
+          "load_error"
+        );
+      }
       if (!this.compat) {
         this.logger().warn(
           "Not using compat mode" + (isFirefox() ? " (expected on Firefox - WebGPU will be disabled)" : "")

vendor/wllama-bonsai/esm/wllama.d.ts

@@ -13,6 +13,13 @@ export interface WllamaLogger {
     error: typeof console.error;
 }
 export interface WllamaConfig {
+    /**
+     * Always use the configured compat worker JS and WASM pair, even when the
+     * browser supports the default JSPI runtime.
+     *
+     * Default: forceCompat = false
+     */
+    forceCompat?: boolean;
     /**
      * If true, suppress all log messages from native CPP code
      */
@@ -176,7 +183,7 @@ export declare class Wllama {
     /**
      * Set compatibility options for Wllama.
      * @param compat Set to null to disable compatibility, or 'default' to use the default compat resources from CDN.
-     * @param mode 'safari' by default; If set to 'firefox_safari', the compat mode will **also** be enabled on Firefox, which will significantly degrade the performance but allow using WebGPU on Firefox.
+     * @param mode 'safari' by default; If set to 'firefox_safari', the compat mode will **also** be enabled on Firefox, which will significantly degrade the performance but allow using WebGPU on Firefox. Ignored when forceCompat is true.
      */
     setCompat(compat: WllamaCompat | null | 'default', mode?: 'safari' | 'firefox_safari'): void;
     /**