1"use strict";(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[3871],{34526:(e,s,t)=>{t.d(s,{A:()=>m});var a=t(37876),i=t(97581),r=t(91108),n=t(42679),l=t(55421),d=t(92631);let o=e=>"specialist"===e?"bg-primary":"openai"===e?"bg-info":"bg-secondary",c=({title:e,values:s,formatValue:t,winner:n,scaleMax:c,models:m})=>{let x=m.map(e=>s[e.id]),p=Math.max(...x),u=Math.min(...x),b=Math.max(c??p,p),h="max"===n?p:u;return(0,a.jsxs)("div",{className:"rounded-[14px] border border-border bg-surface-paper p-4",children:[(0,a.jsx)("div",{className:"mb-4",children:(0,a.jsx)(l.A,{level:"p",variant:"bodyMd",className:"leading-tight text-ink",children:e})}),(0,a.jsx)("div",{className:"space-y-3.5",children:m.map(n=>{let c=s[n.id],m=Math.min(Math.max(c/b*100,8),100),x="specialist"===n.id;return(0,a.jsxs)("div",{className:"grid grid-cols-[1rem_0.25rem_minmax(0,1fr)_auto] items-center gap-x-2 gap-y-2",children:[(0,a.jsx)("span",{className:"flex h-4 w-4 shrink-0 items-center justify-center",children:c===h&&(0,a.jsx)(i.Xs,{size:14,className:"text-ink-muted","aria-label":"Winning model"})}),(0,a.jsx)("span",{className:(0,d.cn)("h-4 w-1 shrink-0 rounded-pill",o(n.id)),"aria-hidden":"true"}),(0,a.jsx)(l.A,{level:"p",variant:"small",className:(0,d.cn)("break-words leading-snug text-ink-secondary",x&&"text-ink"),fontWeight:x?"medium":"normal",children:n.name}),(0,a.jsx)("span",{className:"shrink-0 font-ibm-mono text-xs text-ink",children:t(c)}),(0,a.jsx)("div",{className:"col-start-2 col-end-5 h-2 overflow-hidden rounded-pill bg-surface-light",children:(0,a.jsx)(r.P.div,{initial:!1,animate:{width:`${m}%`},transition:{duration:.55,ease:[.22,1,.36,1]},className:(0,d.cn)("h-full rounded-pill",o(n.id))})})]},`${e}-${n.id}`)})})]})},m=({benchmark:e,headingId:s,className:t})=>(0,a.jsxs)(n.A,{tone:"paper",radius:"card",bordered:!0,className:(0,d.cn)("w-full overflow-hidden rounded-[18px] border-border bg-surface-paper p-4 sm:p-5",t),children:[(0,a.jsxs)("div",{className:"mb-5 border-b border-border-divider pb-5",children:[e.label&&(0,a.jsx)(l.A,{id:s,level:"h3",variant:"bodyMd",className:"mb-2 break-words font-ibm-mono text-primary",children:e.label}),(0,a.jsx)(l.A,{level:"p",variant:"bodyMd",className:"max-w-3xl leading-snug text-ink",children:e.summary})]}),(0,a.jsx)("div",{className:"grid gap-4 sm:grid-cols-2",children:e.metrics.map(s=>(0,a.jsx)(c,{...s,models:e.models},s.title))})]})},66865:(e,s,t)=>{t.d(s,{A:()=>o});var a=t(37876),i=t(92631);let r={teal:{bg:"bg-tag-teal",text:"text-tag-teal",border:"border-tag-teal"},emerald:{bg:"bg-tag-emerald",text:"text-tag-emerald",border:"border-tag-emerald"},violet:{bg:"bg-tag-violet",text:"text-tag-violet",border:"border-tag-violet"},sky:{bg:"bg-tag-sky",text:"text-tag-sky",border:"border-tag-sky"},orange:{bg:"bg-tag-orange",text:"text-tag-orange",border:"border-tag-orange"}},n=[{name:"GLM 5.1 base",detail:"Open model baseline",hue:"teal",accuracy:56,tokenEfficiency:54,latency:58,latencyRaw:"2.4s"},{name:"GLM 5.1 FTv1",detail:"First tuned checkpoint",hue:"emerald",accuracy:78,tokenEfficiency:76,latency:68,latencyRaw:"1.7s"},{name:"GLM 5.1 FTv2",detail:"Best specialist checkpoint",hue:"violet",accuracy:86,tokenEfficiency:84,latency:78,latencyRaw:"1.1s"},{name:"Claude 4.8",detail:"Frontier API model",hue:"sky",accuracy:79,tokenEfficiency:42,latency:50,latencyRaw:"2.9s"},{name:"GPT 5.5",detail:"Frontier API model",hue:"orange",accuracy:82,tokenEfficiency:38,latency:55,latencyRaw:"2.6s"}],l=[{key:"accuracy",label:"Accuracy",caption:"Pass@4 score on held-out financial analysis tasks",valueLabel:e=>`${e.accuracy}%`},{key:"tokenEfficiency",label:"Token efficiency",caption:"Normalized useful work per token",valueLabel:e=>`${e.tokenEfficiency}%`},{key:"latency",label:"Latency",caption:"Normalized responsiveness, raw latency shown",valueLabel:e=>e.latencyRaw}],d=({metric:e})=>{let s=[...n].sort((s,t)=>t[e.key]-s[e.key]);return(0,a.jsxs)("div",{className:"rounded-[12px] border border-border bg-surface-paper p-3 shadow-lb-xs",children:[(0,a.jsxs)("div",{className:"mb-3",children:[(0,a.jsx)("p",{className:"font-ibm text-xs font-semibold text-ink",children:e.label}),(0,a.jsx)("p",{className:"mt-1 font-ibm text-[10px] leading-snug text-ink-muted",children:e.caption})]}),(0,a.jsx)("div",{className:"space-y-2",children:s.map(s=>{let t=r[s.hue],n=s[e.key];return(0,a.jsxs)("div",{className:"group rounded-[9px] border border-border-divider bg-surface px-2.5 py-2 transition-all duration-base ease-lb hover:-translate-y-
1px hover:border-border-strong hover:bg-action-hover hover:shadow-lb-xs",children:[(0,a.jsxs)("div",{className:"mb-1.5 flex items-center justify-between gap-2",children:[(0,a.jsxs)("div",{className:"flex min-w-0 items-center gap-2",children:[(0,a.jsx)("span",{className:(0,i.cn)("h-2.5 w-2.5 shrink-0 rounded-pill",t.bg),"aria-hidden":"true"}),(0,a.jsx)("p",{className:"truncate font-ibm text-[11px] font-semibold text-ink",children:s.name})]}),(0,a.jsx)("span",{className:(0,i.cn)("font-ibm-mono text-[10px] font-semibold",t.text),children:e.valueLabel(s)})]}),(0,a.jsx)("div",{className:"h-2 overflow-hidden rounded-pill bg-action-selected",children:(0,a.jsx)("span",{className:(0,i.cn)("block h-full rounded-pill transition-all duration-slow ease-lb",t.bg),style:{width:`${n}%`},"aria-hidden":"true"})})]},`${e.key}-${s.name}`)})})]})},o=({className:e})=>(0,a.jsxs)("div",{className:(0,i.cn)("relative min-w-0 max-w-full overflow-hidden rounded-[18px] border border-border bg-surface-paper shadow-lb-xl",e),children:[(0,a.jsx)("div",{className:"border-b border-border bg-soft-primary px-3 py-3 text-soft-primary-text sm:px-4",children:(0,a.jsxs)("div",{className:"flex flex-wrap items-start justify-between gap-3",children:[(0,a.jsxs)("div",{className:"min-w-0",children:[(0,a.jsxs)("div",{className:"flex flex-wrap items-center gap-2",children:[(0,a.jsx)("p",{className:"truncate font-ibm text-sm font-semibold",children:"specialist-model-evals"}),(0,a.jsx)("span",{className:"rounded-pill bg-success px-2 py-0.5 font-ibm-mono text-[9px] uppercase tracking-[0.12em] text-success-contrast",children:"ranked"})]}),(0,a.jsx)("p",{className:"mt-1 font-ibm-mono text-[10px]",children:"Accuracy, efficiency, and latency across model candidates"})]}),(0,a.jsxs)("div",{className:"rounded-[8px] border border-border-divider bg-soft-success px-3 py-2 text-right text-soft-success-text",children:[(0,a.jsx)("p",{className:"font-ibm-mono text-[9px] uppercase tracking-[0.14em]",children:"Leader"}),(0,a.jsx)("p",{className:"mt-1 font-ibm text-sm font-semibold",children:"GLM 5.1 FTv2"})]})]})}),(0,a.jsxs)("div",{className:"bg-surface p-3 sm:p-4",children:[(0,a.jsx)("div",{className:"mb-3 flex flex-wrap items-center gap-2",children:n.map(e=>{let s=r[e.hue];return(0,a.jsxs)("div",{className:"flex items-center gap-1.5 rounded-pill border border-border bg-surface-paper px-2 py-1",children:[(0,a.jsx)("span",{className:(0,i.cn)("h-2 w-2 rounded-pill",s.bg),"aria-hidden":"true"}),(0,a.jsx)("span",{className:"break-words font-ibm-mono text-[9px] text-ink-secondary",children:e.name})]},e.name)})}),(0,a.jsx)("div",{className:"grid gap-3 xl:grid-cols-3",children:l.map(e=>(0,a.jsx)(d,{metric:e},e.key))})]})]})},83423:(e,s,t)=>{t.d(s,{A:()=>c});var a=t(37876),i=t(14232),r=t(3351),n=t(27664),l=t(92631);let d={title:"0554-crude-pipeline-lbo-waterfall",promptTitle:"Pipeline LBO Returns Model with Multi-Tier Waterfall",promptIntro:"You are an associate at a private equity fund preparing an investment committee update for a crude oil pipeline project acquired through an LBO. The task is to build an analyst-ready workbook that connects operating performance, the pre-computed debt schedule, exit valuation, and sponsor-management waterfall.",workbookInstructions:["Financing Assumptions","Operating Assumptions","Debt Schedule","Model","Returns"],files:[{name:"inputs.xlsx",path:"/workspace/files/problem/inputs.xlsx"},{name:"source.xlsx",path:"/workspace/files/problem/source.xlsx"}],gradingResults:[{id:"str01",name:"exact_five_visible_sheets",outputScore:"100%",goldScore:"100%"},{id:"str02",name:"sheet_order",outputScore:"100%",goldScore:"100%"},{id:"str03",name:"financing_sheet",outputScore:"100%",goldScore:"100%"},{id:"str04",name:"operating_sheet",outputScore:"100%",goldScore:"100%"},{id:"str05",name:"debt_sheet",outputScore:"100%",goldScore:"100%"},{id:"str08",name:"operating_year_headers",outputScore:"0%",goldScore:"100%",isMissed:!0},{id:"str09",name:"debt_year_headers",outputScore:"0%",goldScore:"100%",isMissed:!0}]},o=["Attachments","Grading","Issues","QA","Forms","Synthesizers","Solver"],c=({className:e,data:s=d,defaultDrawerOpen:t=!1})=>{let[c,m]=(0,i.useState)(t);return(0,a.jsxs)("div",{className:(0,l.cn)("relative min-w-0 max-w-full overflow-hidden rounded-[18px] border border-border bg-surface-paper shadow-lb-xl",e),children:[(0,a.jsx)("div",{className:"flex items-center justify-between border-b border-border bg-soft-primary px-3 py-3 text-soft-primary-text sm:px-4",children:(0,a.jsx)("div",{className:"min-w-0",children:(0,a.jsx)("p",{className:"truncate font-ibm text-sm font-semibold",children:s.title})})}),(0,a.jsxs)("div",{className:"bg-surface px-3 py-4 sm:px-4",children:[(0,a.jsxs)("div",{className:"mb-2 flex items-center justify-between",children:[(0,a.jsx)("p",{className:"font-ibm text-xs font-medium text-ink",children:"Task runs"}),(0,a.jsx)("p",{className:"font-ibm text-[10px] text-ink-muted",children:"Show all 10 runs"})]}),(0,a.jsx)("button",{type:"button","aria-expanded":c,onClick:()=>m(!0),className:"group mb-4 flex w-full cursor-pointer items-center rounded-[10px] border border-border bg-surface-paper px-3 py-2 text-left transition-all duration-base ease-lb hover:-translate-y-
1px hover:border-border-strong hover:bg-action-hover hover:shadow-lb-xs focus-visible:outline-none focus-visible:shadow-lb-focus",children:(0,a.jsxs)("div",{className:"flex min-w-0 items-center gap-2",children:[(0,a.jsx)("span",{className:"h-2 w-2 shrink-0 rounded-pill bg-warning","aria-hidden":"true"}),(0,a.jsxs)("div",{className:"min-w-0",children:[(0,a.jsxs)("div",{className:"flex flex-wrap items-center gap-2",children:[(0,a.jsx)("span",{className:"font-ibm text-xs font-semibold text-ink",children:"Attempt 5"}),(0,a.jsx)("span",{className:"rounded-pill bg-success px-2 py-0.5 font-ibm-mono text-[9px] uppercase tracking-[0.12em] text-success-contrast",children:"completed"}),(0,a.jsx)("span",{className:"rounded-pill bg-action-selected px-2 py-0.5 font-ibm-mono text-[9px] text-ink-secondary",children:"Baseline score 29%"})]}),(0,a.jsx)("p",{className:"mt-0.5 truncate font-ibm-mono text-[10px] text-ink-muted",children:"GLM 5.1 baseline - 2026-06-03 18:34:11 UTC"})]})]})}),(0,a.jsxs)("div",{className:"grid min-h-[310px] min-w-0 max-w-full overflow-hidden rounded-[12px] border border-border bg-surface-paper lg:grid-cols-[0.95fr_1fr]",children:[(0,a.jsxs)("div",{className:"min-w-0 border-b border-border-divider p-4 lg:border-r lg:border-b-0",children:[(0,a.jsxs)("div",{className:"mb-3 flex items-center justify-between",children:[(0,a.jsx)("span",{className:"font-ibm text-xs font-medium text-ink-secondary",children:"Prompt"}),(0,a.jsx)("span",{className:"font-ibm-mono text-[10px] text-ink-muted",children:"6798 chars"})]}),(0,a.jsxs)("div",{className:"space-y-3 font-ibm text-[11px] leading-relaxed text-ink-secondary",children:[(0,a.jsxs)("p",{className:"font-medium text-ink",children:["# ",s.promptTitle]}),(0,a.jsx)("p",{children:"## Context"}),(0,a.jsx)("p",{children:s.promptIntro}),(0,a.jsxs)("p",{children:["Save the completed workbook to"," ",(0,a.jsx)("span",{className:"break-all font-ibm-mono text-ink",children:"/workspace/output/model.xlsx"}),"."]}),(0,a.jsxs)("div",{children:[(0,a.jsx)("p",{className:"mb-2",children:"## Workbook to build"}),(0,a.jsx)("ol",{className:"space-y-1",children:s.workbookInstructions.map((e,s)=>(0,a.jsxs)("li",{children:[s+1,". ",e]},e))})]})]})]}),(0,a.jsxs)("div",{className:"min-w-0 p-4",children:[(0,a.jsx)("div",{className:"mb-4 min-w-0 overflow-x-auto border-b border-border-divider pb-3",children:(0,a.jsx)("div",{className:"flex w-max min-w-full gap-3",children:o.map((e,s)=>(0,a.jsx)("span",{className:(0,l.cn)("shrink-0 font-ibm text-[10px] font-medium",0===s?"text-ink":"text-ink-muted"),children:e},e))})}),(0,a.jsxs)("div",{className:"mb-3",children:[(0,a.jsxs)("div",{className:"flex items-center gap-2",children:[(0,a.jsx)("p",{className:"font-ibm text-xs font-semibold text-ink",children:"Task files"}),(0,a.jsx)("span",{className:"font-ibm text-[10px] text-ink-muted",children:"Optional"})]}),(0,a.jsx)("p",{className:"mt-1 font-ibm text-[10px] leading-relaxed text-ink-muted",children:"Files accessible to the model at the container mount path."})]}),(0,a.jsx)("div",{className:"space-y-2",children:s.files.map(e=>(0,a.jsxs)("div",{className:"flex items-center justify-between rounded-[8px] border border-border bg-surface px-3 py-2",children:[(0,a.jsxs)("div",{className:"min-w-0",children:[(0,a.jsx)("p",{className:"font-ibm text-xs font-medium text-ink",children:e.name}),(0,a.jsx)("p",{className:"truncate font-ibm-mono text-[10px] text-ink-muted",children:e.path})]}),(0,a.jsx)("button",{type:"button","aria-label":`Download ${e.name}`,className:"flex h-7 w-7 shrink-0 items-center justify-center rounded-pill text-ink-muted transition-colors duration-base ease-lb hover:bg-action-hover hover:text-ink focus-visible:outline-none focus-visible:shadow-lb-focus",children:(0,a.jsx)(r.f5,{size:16,"aria-hidden":"true"})})]},e.name))})]})]})]}),c&&(0,a.jsxs)("div",{className:"absolute inset-0 z-10",children:[(0,a.jsx)("button",{type:"button","aria-label":"Close attempt details",onClick:()=>m(!1),className:"absolute inset-0 cursor-default bg-grey-900/60"}),(0,a.jsxs)("aside",{"aria-label":"Attempt grading results",className:"absolute top-0 right-0 flex h-full w-full max-w-[360px] flex-col bg-surface-paper shadow-lb-xl sm:max-w-[46%]",children:[(0,a.jsxs)("div",{className:"border-b border-border-divider px-4 py-4",children:[(0,a.jsxs)("div",{className:"mb-3 flex items-start justify-between gap-3",children:[(0,a.jsxs)("div",{className:"min-w-0",children:[(0,a.jsxs)("div",{className:"flex flex-wrap items-center gap-2",children:[(0,a.jsx)("p",{className:"font-ibm text-sm font-semibold text-ink",children:"Attempt 5"}),(0,a.jsx)("span",{className:"rounded-pill bg-success px-2 py-0.5 font-ibm-mono text-[9px] uppercase tracking-[0.12em] text-success-contrast",children:"completed"}),(0,a.jsx)("span",{className:"rounded-pill bg-action-selected px-2 py-0.5 font-ibm-mono text-[9px] text-ink-secondary",children:"Baseline score 29%"})]}),(0,a.jsx)("p",{className:"mt-1 truncate font-ibm-mono text-[10px] text-ink-muted",children:"GLM 5.1 baseline - 6/3/2026, 11:24:57 AM"})]}),(0,a.jsx)("button",{type:"button","aria-label":"Close attempt details",onClick:()=>m(!1),className:"flex h-7 w-7 shrink-0 items-center justify-center rounded-pill text-ink-muted transition-colors duration-base ease-lb hover:bg-action-hover hover:text-ink focus-visible:outline-none focus-visible:shadow-lb-focus",children:(0,a.jsx)(n.bm,{size:16,"aria-hidden":"true"})})]}),(0,a.jsxs)("div",{className:"flex flex-wrap gap-x-4 gap-y-1 font-ibm text-[11px] text-ink-secondary",children:[(0,a.jsxs)("span",{children:["Baseline score ",(0,a.jsx)("strong",{className:"text-error",children:"29%"})]}),(0,a.jsxs)("span",{children:["Total time ",(0,a.jsx)("strong",{className:"text-ink",children:"535.9s"})]}),(0,a.jsxs)("span",{children:["Turns ",(0,a.jsx)("strong",{className:"text-ink",children:"1"})]})]})]}),(0,a.jsxs)("div",{className:"min-h-0 flex-1 overflow-y-auto px-4 py-4",children:[(0,a.jsxs)("div",{className:"mb-4 flex items-start justify-between",children:[(0,a.jsxs)("div",{children:[(0,a.jsx)("p",{className:"font-ibm text-sm font-semibold text-ink",children:"Grading results"}),(0,a.jsx)("p",{className:"mt-1 font-ibm text-base font-semibold text-error",children:"29%"})]}),(0,a.jsxs)("div",{className:"text-right font-ibm text-[10px] text-ink-muted",children:[(0,a.jsx)("p",{children:"Graded 6/3/2026, 11:34:46 AM"}),(0,a.jsx)("p",{className:"mt-3 text-ink-secondary",children:"Expand all - Collapse all"})]})]}),(0,a.jsx)("div",{className:"space-y-2",children:s.gradingResults.map((e,s)=>(0,a.jsxs)("div",{className:"grid grid-cols-[auto_minmax(0,1fr)] items-center gap-2 rounded-[8px] border border-border-divider bg-surface px-2.5 py-2 sm:grid-cols-[auto_minmax(0,1fr)_auto_auto]",children:[(0,a.jsx)("span",{className:"font-ibm text-xs text-ink-muted","aria-hidden":"true",children:">"}),(0,a.jsxs)("p",{className:"truncate font-ibm text-[11px] text-ink",children:["#",s+1," ",e.id,"_",e.name]}),(0,a.jsxs)("span",{className:(0,l.cn)("col-start-2 rounded-[6px] px-2 py-1 font-ibm-mono text-[9px] sm:col-start-auto",e.isMissed?"bg-error-50 text-error":"bg-success-50 text-success"),children:["Model output ",e.outputScore]}),(0,a.jsxs)("span",{className:"col-start-2 rounded-[6px] bg-success-50 px-2 py-1 font-ibm-mono text-[9px] text-success sm:col-start-auto",children:["Gold standard ",e.goldScore]})]},e.id))})]})]})]})]})}},88370:(e,s,t)=>{t.d(s,{A:()=>g});var a=t(37876),i=t(14232),r=t(93426),n=t(27664),l=t(92631);let d=[{label:"Base",duration:"8h 20m",detail:"Baseline evaluation across 120 held-out tasks and 4 attempts each"},{label:"Training",duration:"11h 45m",detail:"GRPO run on the H100 cluster"},{label:"Tuned",duration:"8h 55m",detail:"Tuned evaluation against the same held-out suite"}],o=[{title:"Training data",items:[{label:"Source",value:"4,300 financial analysis tasks"}
1,{label:"Task family",value:"DCF, LBO, acquisition, projection"}]},{title:"Model",items:[{label:"Base model",value:"GLM 5.1"},{label:"Training method",value:"GRPO"}]},{title:"Infrastructure",items:[{label:"Compute",value:"GKE, H100 cluster"},{label:"Endpoints",value:"Baseline and tuned"}]}],c=[{name:"0195-acme-industrials-dcf-irr",baseline:["54%","58%","52%","56%"],tuned:["84%","87%","82%","86%"],reward:"0.55 -> 0.85",delta:"+0.30",rubrics:[{id:"dcf01",title:"revenue_driver_link",description:"Revenue forecast links to Acme Industrials volume and pricing drivers."},{id:"dcf02",title:"ebitda_margin_bridge",description:"EBITDA margin forecast reconciles historical margins to forward assumptions."},{id:"dcf03",title:"working_capital_schedule",description:"Working capital schedule uses receivables, inventory, and payables assumptions."},{id:"dcf04",title:"capex_depreciation_rollforward",description:"Capex and depreciation roll forward correctly through the forecast period."},{id:"dcf05",title:"unlevered_fcf",description:"Unlevered free cash flow includes tax, capex, and working-capital changes."},{id:"dcf06",title:"terminal_value",description:"Terminal value uses the requested exit multiple and terminal EBITDA."},{id:"dcf07",title:"irr_moic_outputs",description:"IRR and MOIC outputs reconcile to entry equity and exit equity value."},{id:"dcf08",title:"sensitivity_table",description:"Sensitivity table flexes exit multiple and EBITDA margin assumptions."}]},{name:"0248-pumptech-acquisition-rollup",baseline:["61%","56%","58%","63%"],tuned:["88%","91%","87%","90%"],reward:"0.60 -> 0.89",delta:"+0.29",rubrics:[{id:"acq01",title:"target_company_rollup",description:"All acquired PumpTech subsidiaries are included in the consolidated rollup."},{id:"acq02",title:"purchase_price_allocation",description:"Purchase prices, debt assumptions, and transaction fees map to each target."},{id:"acq03",title:"synergy_schedule",description:"Cost and revenue synergies phase in according to the acquisition timeline."},{id:"acq04",title:"pro_forma_revenue",description:"Pro forma revenue combines standalone and acquired-company revenue correctly."},{id:"acq05",title:"pro_forma_ebitda",description:"Pro forma EBITDA includes target EBITDA and synergy adjustments."},{id:"acq06",title:"leverage_covenants",description:"Net leverage and covenant metrics reflect combined debt and EBITDA."},{id:"acq07",title:"returns_waterfall",description:"Sponsor returns waterfall uses consolidated exit value and debt paydown."},{id:"acq08",title:"deal_comparison",description:"The workbook compares standalone and acquisition-case returns."}]},{name:"0515-hotel-renovation-noi-yield",baseline:["48%","52%","55%","50%"],tuned:["80%","84%","82%","86%"],reward:"0.51 -> 0.83",delta:"+0.32",rubrics:[{id:"hotel01",title:"room_count_and_adr",description:"Room count, ADR, and occupancy assumptions drive room revenue."},{id:"hotel02",title:"occupancy_ramp",description:"Occupancy ramp reflects renovation and stabilization timing."},{id:"hotel03",title:"ancillary_revenue",description:"Food, beverage, parking, and resort-fee revenue use stated percentages."},{id:"hotel04",title:"department_margin",description:"Department margins separate room, F&B, and other operating income."},{id:"hotel05",title:"fixed_costs",description:"Fixed costs include property taxes, insurance, and franchise fees."},{id:"hotel06",title:"renovation_capex",description:"Renovation capex is timed before stabilized cash-flow periods."},{id:"hotel07",title:"noi_yield",description:"NOI yield uses stabilized NOI divided by all-in basis."},{id:"hotel08",title:"exit_value",description:"Exit value uses the stated cap rate on stabilized NOI."}]},{name:"0843-hartwell-plant-expansion-dcf",baseline:["50%","53%","57%","55%"],tuned:["81%","83%","85%","84%"],reward:"0.54 -> 0.83",delta:"+0.29",rubrics:[{id:"mfg01",title:"unit_volume_forecast",description:"Manufacturing unit volume forecast follows plant utilization assumptions."},{id:"mfg02",title:"price_mix_effect",description:"Revenue captures ASP growth and product-mix effects."},{id:"mfg03",title:"cogs_material_labor",description:"COGS separates raw material, labor, and overhead cost drivers."},{id:"mfg04",title:"gross_margin_bridge",description:"Gross margin bridge reflects scale, mix, and input-cost inflation."},{id:"mfg05",title:"maintenance_capex",description:"Maintenance capex is tied to revenue and asset base assumptions."},{id:"mfg06",title:"tax_and_nol",description:"Tax calculation applies NOL usage before cash tax payments."},{id:"mfg07",title:"discounted_fcf",description:"Forecast free cash flows are discounted at the requested WACC."},{id:"mfg08",title:"terminal_sensitivity",description:"Terminal-value sensitivity flexes WACC and growth-rate assumptions."}]}
1,{name:"0120-smasco-revenue-forecast",baseline:["52%","49%","55%","51%"],tuned:["82%","85%","88%","84%"],reward:"0.52 -> 0.85",delta:"+0.33",rubrics:[{id:"proj01",title:"historical_revenue_import",description:"Historical revenue imports cleanly from the source statements."},{id:"proj02",title:"segment_growth_assumptions",description:"Segment-level growth rates drive projected revenue by business line."},{id:"proj03",title:"opex_schedule",description:"Operating expenses use fixed, variable, and inflation-linked assumptions."},{id:"proj04",title:"gross_profit_projection",description:"Gross profit ties to revenue and segment margin assumptions."},{id:"proj05",title:"working_capital_turns",description:"Working-capital turns calculate receivables, inventory, and payables."},{id:"proj06",title:"cash_flow_statement",description:"Cash flow statement reconciles net income, working capital, and capex."},{id:"proj07",title:"debt_schedule",description:"Debt schedule computes interest expense and mandatory amortization."},{id:"proj08",title:"balance_sheet_check",description:"Projected balance sheet balances in every forecast year."}]},{name:"0631-proactivate-lbo-exit-returns",baseline:["57%","60%","54%","58%"],tuned:["86%","89%","84%","87%"],reward:"0.57 -> 0.87",delta:"+0.30",rubrics:[{id:"str01",title:"exact_five_visible_sheets",description:"Workbook has exactly five visible sheets."},{id:"str02",title:"sheet_order",description:"Workbook follows the requested sheet order from source tabs to Model to Returns."},{id:"str03",title:"financing_sheet",description:"Financing Assumptions sheet exists."},{id:"str04",title:"operating_sheet",description:"Operating Assumptions sheet exists."},{id:"str05",title:"debt_sheet",description:"Debt Schedule sheet exists."},{id:"str10",title:"model_year_headers",description:"Model has 2013-2017 year headers."},{id:"str11",title:"returns_scenario_headers",description:"Returns has 7.0x, 8.0x, 9.0x, and 10.0x scenario columns."},{id:"ret12",title:"gross_irr",description:"Returns gross IRR by exit scenario matches expected value."}]}],m=({children:e})=>(0,a.jsx)("span",{className:"rounded-pill bg-success px-2 py-0.5 font-ibm-mono text-[9px] uppercase tracking-[0.12em] text-success-contrast",children:e}),x=({value:e,phase:s,onSelect:t})=>{let i="x"===e;return(0,a.jsx)("button",{type:"button",onClick:t,className:(0,l.cn)("rounded-[6px] border px-2 py-1 font-ibm-mono text-[10px] font-semibold transition-all duration-base ease-lb hover:-translate-y-px hover:border-border-strong hover:bg-action-hover hover:shadow-lb-xs focus-visible:outline-none focus-visible:shadow-lb-focus",i?"border-error-100 bg-error-50 text-error":"Tuned"===s?"border-success bg-soft-success text-soft-success-text":"border-border bg-surface-paper text-ink"),children:e})},p=()=>(0,a.jsx)("div",{className:"grid grid-cols-1 gap-1.5 min-[420px]:grid-cols-3","aria-label":"Completed run stages",children:d.map((e,s)=>(0,a.jsxs)("div",{className:"group relative min-w-0 rounded-[8px] border border-success bg-soft-success px-2.5 py-2 text-soft-success-text transition-all duration-base ease-lb hover:-translate-y-px hover:border-success-dark hover:shadow-lb-xs",children:[(0,a.jsxs)("div",{className:"min-w-0",children:[(0,a.jsx)("p",{className:"truncate font-ibm text-[11px] font-semibold",children:e.label}),(0,a.jsx)("p",{className:"font-ibm-mono text-[9px]",children:e.duration})]}),s<d.length-1?(0,a.jsx)("span",{className:"absolute top-1/2 right-[-10px] z-10 hidden h-5 w-5 -translate-y-1/2 items-center justify-center rounded-pill border border-success bg-surface-paper text-success shadow-lb-xs xl:flex","aria-hidden":"true",children:(0,a.jsx)(r.Qp,{size:12})}):null,(0,a.jsxs)("div",{className:"pointer-events-none absolute top-full left-2 z-30 mt-1 hidden w-max max-w-[220px] rounded-[8px] border border-border bg-surface-paper px-2 py-1.5 font-ibm text-[10px] leading-snug text-ink-secondary shadow-lb-md group-hover:block",children:[(0,a.jsx)("span",{className:"font-semibold text-ink",children:e.label}),(0,a.jsxs)("span",{children:[" completed in ",e.duration]}),e.detail&&(0,a.jsxs)("span",{children:[" - ",e.detail]})]})]},e.label))}),u=()=>(0,a.jsx)("div",{className:"overflow-hidden rounded-[10px] border border-border-divider bg-surface-paper",children:(0,a.jsxs)("table",{className:"w-full border-collapse text-left",children:[(0,a.jsx)("thead",{className:"bg-action-selected",children:(0,a.jsx)("tr",{children:["Metric","Value"].map(e=>(0,a.jsx)("th",{scope:"col",className:"px-3 py-2 font-ibm-mono text-[10px] font-semibold uppercase tracking-[0.12em]",children:e},e))})}),(0,a.jsx)("tbody",{className:"divide-y divide-border-divider",children:o.flatMap(e=>e.items.map(s=>(0,a.jsxs)("tr",{children:[(0,a.jsx)("td",{className:"px-3 py-2.5 font-ibm-mono text-[10px] uppercase tracking-[0.1em]",children:s.label}),(0,a.jsx)("td",{className:"px-3 py-2.5 font-ibm text-xs font-semibold",children:s.value})]},`${e.title}-${s.label}`)))})]})}),b=()=>(0,a.jsxs)("div",{className:"space-y-3 rounded-[12px] border border-border bg-surface-paper p-3 shadow-lb-xs",children:[(0,a.jsx)("p",{className:"font-ibm text-xs font-semibold text-ink",children:"Training run summary"}),(0,a.jsx)(p,{}),(0,a.jsx)("div",{className:"space-y-3",children:(0,a.jsx)(u,{})})]}),h=({attempt:e,onClose:s})=>{let t,i="x"===e.value?"0%":e.value,r=Number.parseInt(i.replace("%",""),10),d="Tuned"===e.phase&&r>=80?"text-success":"text-ink",o=(t=Math.max(0,Math.min(100,Number.parseInt(e.value.replace("%",""),10)))*e.rubrics.length,e.rubrics.map((e,s)=>{let a;return{...e,modelOutput:(a=Math.max(0,Math.min(100,t)),t-=a,`${Math.round(a)}%`),goldStandard:"100%"}})),c="Baseline"===e.phase?"GLM 5.1 baseline":"Fine-tuned GLM 5.1 v7";return(0,a.jsxs)("div",{className:"absolute inset-0 z-20",children:[(0,a.jsx)("button",{type:"button","aria-label":"Close attempt results",onClick:s,className:"absolute inset-0 cursor-default bg-grey-900/60"}),(0,a.jsxs)("aside",{"aria-label":"Attempt results",className:"absolute top-0 right-0 flex h-full w-full max-w-[560px] flex-col bg-surface-paper shadow-lb-xl",children:[(0,a.jsxs)("div",{className:"border-b border-border-divider px-4 py-4",children:[(0,a.jsxs)("div",{className:"mb-3 flex items-start justify-between gap-3",children:[(0,a.jsxs)("div",{className:"min-w-0",children:[(0,a.jsxs)("div",{className:"flex flex-wrap items-center gap-2",children:[(0,a.jsxs)("p",{className:"font-ibm text-sm font-semibold text-ink",children:["Attempt ",e.attemptNumber]}),(0,a.jsx)(m,{children:"completed"}),(0,a.jsx)("p",{className:"truncate font-ibm text-xs text-ink-secondary",children:c})]}),(0,a.jsxs)("p",{className:"mt-1 truncate font-ibm-mono text-[10px] text-ink-muted",children:[e.phase," eval - ",e.taskName]})]}),(0,a.jsx)("button",{type:"button","aria-label":"Close attempt results",onClick:s,className:"flex h-7 w-7 shrink-0 items-center justify-center rounded-pill text-ink-muted transition-colors duration-base ease-lb hover:bg-action-hover hover:text-ink focus-visible:outline-none focus-visible:shadow-lb-focus",children:(0,a.jsx)(n.bm,{size:16,"aria-hidden":"true"})})]}),(0,a.jsxs)("div",{className:"flex flex-wrap gap-x-4 gap-y-1 font-ibm text-[11px] text-ink-secondary",children:[(0,a.jsxs)("span",{children:["Score"," ",(0,a.jsx)("strong",{className:d,children:i})]}),(0,a.jsxs)("span",{children:["Total time ",(0,a.jsx)("strong",{className:"text-ink",children:"380.8s"})]}),(0,a.jsxs)("span",{children:["Turns ",(0,a.jsx)("strong",{className:"text-ink",children:"21"})]})]})]}),(0,a.jsxs)("div",{className:"min-h-0 flex-1 overflow-y-auto px-4 py-4",children:[(0,a.jsxs)("div",{className:"mb-4 flex items-start justify-between gap-3",children:[(0,a.jsxs)("div",{children:[(0,a.jsx)("p",{className:"font-ibm text-sm font-semibold text-ink",children:"Rubric results"}),(0,a.jsx)("p",{className:(0,l.cn)("mt-1 font-ibm text-base font-semibold",d),children:i})]}),(0,a.jsxs)("div",{className:"text-right font-ibm text-[10px] text-ink-muted",children:[(0,a.jsx)("p",{children:"Graded 6/12/2026, 11:19:48 AM"}),(0,a.jsx)("p",{className:"mt-3 text-ink-secondary",children:"Expand all - Collapse all"})]})]}),(0,a.jsx)("div",{className:"space-y-2",children:o.map((e,s)=>{let t="0%"!==e.modelOutput;return(0,a.jsxs)("div",{className:"overflow-hidden rounded-[8px] border border-border-divider bg-surface",children:[(0,a.jsxs)("div",{className:"grid grid-cols-[auto_minmax(0,1fr)] items-center gap-2 px-3 py-2 sm:grid-cols-[auto_minmax(0,1fr)_auto]",children:[(0,a.jsx)("span",{className:"font-ibm text-xs text-ink-muted","aria-hidden":"true",children:">"}),(0,a.jsxs)("p",{className:"truncate font-ibm text-[11px] font-semibold text-ink",children:["#",s+1," ",e.id,"_",e.title]}),(0,a.jsxs)("div",{className:"col-start-2 flex flex-wrap items-center gap-1.5 sm:col-start-auto",children:[(0,a.jsxs)("span",{className:(0,l.cn)("rounded-[6px] px-2 py-1 font-ibm-mono text-[9px]",t?"bg-success-50 text-success":"bg-error-50 text-error"),children:["Model output ",e.modelOutput]}),(0,a.jsxs)("span",{className:"rounded-[6px] bg-success-50 px-2 py-1 font-ibm-mono text-[9px] text-success",children:["Gold standard ",e.goldStandard]})]})]}),(0,a.jsx)("p",{className:"border-t border-border-divider bg-surface-light px-3 py-2 font-ibm text-[11px] leading-relaxed text-ink-secondary sm:px-8",children:e.description})]},e.id)})})]})]})]})},f=({onSelectAttempt:e})=>(0,a.jsxs)("div",{className:"space-y-3 rounded-[12px] border border-border bg-surface-paper p-3 shadow-lb-xs",children:[(0,a.jsxs)("div",{className:"flex flex-wrap items-center justify-between gap-2",children:[(0,a.jsx)("p",{className:"font-ibm text-xs font-semibold text-ink",children:"Evaluations"}),(0,a.jsx)("p",{className:"font-ibm-mono text-[9px] uppercase tracking-[0.14em] text-ink-muted",children:"4 attempts per task"})]}),(0,a.jsxs)("div",{className:"overflow-hidden rounded-[10px] border border-border-divider bg-surface-paper",children:[(0,a.jsx)("div",{className:"grid gap-2 border-b border-border-divider bg-action-selected px-3 py-2 xl:grid-cols-[minmax(190px,1fr)_1.1fr_1.1fr_5.5rem]",children:["Problem","Baseline","Tuned","Delta"].map(e=>(0,a.jsx)("p",{className:"font-ibm-mono text-[10px] uppercase tracking-[0.12em] text-ink-secondary",children:e},e))}),c.map(s=>{let t=s.delta.startsWith("+"),i="--"===s.delta;return(0,a.jsxs)("div",{className:"grid gap-2 border-b border-border-divider px-3 py-3 transition-colors duration-base ease-lb last:border-b-0 hover:bg-action-hover xl:grid-cols-[minmax(190px,1fr)_1.1fr_1.1fr_5.5rem]",children:[(0,a.jsx)("p",{className:"truncate font-ibm text-xs font-semibold text-ink",children:s.name}),(0,a.jsx)("div",{className:"flex flex-wrap gap-2",children:s.baseline.map((t,i)=>
1(0,a.jsx)(x,{value:t,phase:"Baseline",onSelect:()=>e({taskName:s.name,phase:"Baseline",value:t,attemptNumber:i+1,rubrics:s.rubrics})},`${s.name}-baseline-${i}`))}),(0,a.jsx)("div",{className:"flex flex-wrap gap-2",children:s.tuned.map((t,i)=>(0,a.jsx)(x,{value:t,phase:"Tuned",onSelect:()=>e({taskName:s.name,phase:"Tuned",value:t,attemptNumber:i+1,rubrics:s.rubrics})},`${s.name}-tuned-${i}`))}),(0,a.jsxs)("div",{className:"text-left",children:[(0,a.jsx)("span",{className:(0,l.cn)("inline-flex rounded-pill px-2 py-0.5 font-ibm-mono text-[10px] font-semibold",i?"bg-action-selected text-ink-muted":t?"bg-soft-success text-soft-success-text":"bg-soft-error text-soft-error-text"),children:s.delta}),(0,a.jsx)("p",{className:"mt-1 font-ibm-mono text-[10px] text-ink-secondary",children:s.reward})]})]},s.name)})]})]}),g=({className:e})=>{let[s,t]=(0,i.useState)(null);return(0,a.jsxs)("div",{className:(0,l.cn)("relative min-w-0 max-w-full overflow-hidden rounded-[18px] border border-border bg-surface-paper shadow-lb-xl",e),children:[(0,a.jsx)("div",{className:"border-b border-border bg-soft-primary px-3 py-3 text-soft-primary-text sm:px-4",children:(0,a.jsxs)("div",{className:"flex flex-wrap items-start justify-between gap-3",children:[(0,a.jsxs)("div",{className:"min-w-0",children:[(0,a.jsxs)("div",{className:"flex flex-wrap items-center gap-2",children:[(0,a.jsx)("p",{className:"truncate font-ibm text-sm font-semibold",children:"financial-agent-glm-5.1-v7-fits6h"}),(0,a.jsx)(m,{children:"completed"})]}),(0,a.jsx)("p",{className:"mt-1 font-ibm-mono text-[10px]",children:"GRPO run - GLM 5.1"})]}),(0,a.jsxs)("div",{className:"w-full rounded-[8px] border border-border-divider bg-soft-success px-3 py-2 text-left text-soft-success-text sm:w-auto sm:text-right",children:[(0,a.jsx)("p",{className:"font-ibm-mono text-[9px] uppercase tracking-[0.14em]",children:"Run result"}),(0,a.jsxs)("p",{className:"mt-1 font-ibm text-sm font-semibold",children:["86% Pass@4 score"," ",(0,a.jsx)("span",{className:"font-ibm-mono text-[10px]",children:"(+30%)"})]})]})]})}),(0,a.jsxs)("div",{className:"grid gap-3 bg-surface p-3 sm:p-4 xl:grid-cols-[minmax(220px,1fr)_minmax(0,3fr)] xl:items-start",children:[(0,a.jsx)("div",{className:"min-w-0",children:(0,a.jsx)(b,{})}),(0,a.jsx)("div",{className:"min-w-0",children:(0,a.jsx)(f,{onSelectAttempt:t})})]}),s&&(0,a.jsx)(h,{attempt:s,onClose:()=>t(null)})]})}}}]);
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.