/* global window */ /* ============================================================ P&ID Arena, scripted demo data ------------------------------------------------------------ All races below are MOCK. Each model result is a pre-written answer with a wall-clock duration and a final latency reading. // replace with live SSE feeds ============================================================ */ // Corpus presets shown in the header segmented control. const CORPORA = [ { id: 1, label: "1 sheet", sub: "Debutanizer", sheets: 1, detail: "NGL Plant · Debutanizer · R279032044P01" }, { id: 10, label: "10 sheets", sub: "Units 1–3", sheets: 10, detail: "NGL Plant · Units 1–3 · 10 sheets parsed" }, { id: 20, label: "20 sheets", sub: "Full set", sheets: 20, detail: "NGL Plant · Full P&ID set · 20 sheets parsed" }, ]; // The four contestants. Order matters: ours is first. // `gui` is the competitor's own web app — the panel links the SAME // query there so the audience can see we're not faking the race. // (Live deep-linking depends on each vendor; treat as illustrative.) const MODELS = [ { key: "intui", name: "IntuigenceAI", version: "P&ID Agent v1.4", ours: true, gui: "https://app.intuigence.ai/arena" }, { key: "gpt", name: "ChatGPT", version: "GPT-5", gui: "https://chatgpt.com/?q=" }, { key: "claude", name: "Claude", version: "Opus 4.6", gui: "https://claude.ai/new?q=" }, { key: "gemini", name: "Gemini", version: "2.5 Pro", gui: "https://gemini.google.com/app" }, ]; // Question bank. `verified: true` shows the ground-truth checkmark. const QUESTIONS = [ { id: "pic1108", text: "What is the prefix for tag PIC1108, and how did you find it?", category: "Tags", verified: true, truth: "Q69, per Note 1 on sheet R279032044P01.", }, { id: "valves", text: "How many control valves are on the debutanizer overhead line, and what are their tags?", category: "Valves", verified: true, truth: "2, FV1102 and PV1108, both on sheet R279032044P01.", }, { id: "relief", text: "Is there a relief path from the overhead drum to flare? Cite the device.", category: "Safety", verified: true, truth: "Yes, PSV1140 discharges to the flare header (sheet R279032044P03).", }, { id: "connect", text: "What does the bottoms pump P-110A feed downstream?", category: "Connectivity", verified: true, truth: "Reboiler E-114, then back to the column (sheet R279032044P02).", }, ]; // ── The hero script: PIC1108 across 1 / 10 / 20 sheets ────────── // `ms` = wall-clock streaming time for the demo; `latency` = the // number the on-screen timer freezes at. Ours runs LONGEST on // purpose, honesty sells it. const SCENARIOS = { pic1108: { 1: { intui: { verdict: "correct", latency: 4.8, ms: 4200, evidence: true, answer: "The prefix for PIC1108 is Q69. Note 1 on sheet R279032044P01 assigns the Q69 unit prefix to every instrument in this area. PIC1108 is the debutanizer overhead pressure controller.", note: "Matches ground truth, cited Note 1." }, gpt: { verdict: "correct", latency: 6.6, ms: 5600, answer: "PIC1108 is a Pressure Indicating Controller. The general notes on this drawing give the unit prefix as Q69, so the full tag prefix is Q69.", note: "Correct, the note was on the single loaded sheet." }, claude: { verdict: "partial", latency: 7.4, ms: 6400, answer: "PIC1108 is a pressure indicating controller on loop 1108. The drawing notes appear to define a unit prefix, which looks like Q69, though I'd confirm against the legend.", note: "Right value, hedged, did not commit to the source." }, gemini: { verdict: "partial", latency: 6.9, ms: 6000, answer: "PIC1108 = Pressure Indicating Controller, loop number 1108. A unit prefix (Q69) is referenced in the notes; full context not fully resolved.", note: "Identified the loop; prefix attribution incomplete." }, }, 10: { intui: { verdict: "correct", latency: 9.6, ms: 7200, evidence: true, answer: "The prefix for PIC1108 is Q69. Across the 10 loaded sheets, Note 1 on R279032044P01 is the only place the prefix is defined; the knowledge graph links PIC1108 to that note directly.", note: "Matches ground truth, resolved the note across 10 sheets." }, gpt: { verdict: "partial", latency: 7.1, ms: 5800, answer: "PIC1108 is a pressure controller. The prefix is likely Q69 based on the unit numbering, but several sheets use different prefixes so I cannot be fully certain.", note: "Correct value, low confidence, could not isolate the note." }, claude: { verdict: "partial", latency: 8.6, ms: 6800, answer: "PIC1108 is a pressure indicating controller. I see prefix references across the set but cannot confidently tie PIC1108 to a single one without the legend sheet.", note: "Declined to commit across the larger set." }, gemini: { verdict: "partial", latency: 7.8, ms: 6400, answer: "PIC1108 is loop 1108, a pressure indicating controller. Multiple prefixes appear across the 10 sheets; unable to disambiguate the correct one.", note: "Loop identified; prefix not resolved." }, }, 20: { intui: { verdict: "correct", latency: 14.2, ms: 9000, evidence: true, answer: "The prefix for PIC1108 is Q69. With all 20 sheets parsed into the knowledge graph, PIC1108 resolves to a single instrument node on sheet R279032044P01, whose prefix is governed by Note 1 on that same sheet. PIC1108 is the debutanizer overhead pressure controller.", note: "Matches ground truth, Q69 per Note 1, R279032044P01." }, gpt: { verdict: "incorrect", latency: 7.4, ms: 6200, answer: "The prefix for PIC1108 is P-11. Under standard ISA loop numbering the leading digits denote the unit area, so PIC1108 sits in unit 11 with prefix P-11.", note: "Ground truth is Q69 (Note 1). GPT-5 applied an ISA rule that is not used in these drawings." }, claude: { verdict: "partial", latency: 9.0, ms: 7400, answer: "I cannot determine the prefix for PIC1108 from the provided drawings with confidence. The tag appears on a pressure loop, but I cannot locate a definitive prefix assignment across the 20 sheets.", note: "Correctly declined to guess, but missed Note 1 on R279032044P01." }, gemini: { verdict: "failed", latency: 6.2, ms: 5200, failAt: 0.45, answer: "Reviewing the drawing set, PIC1108 appears on a pressure control loop and the unit prefix is defined in", note: "No answer produced, context limit hit with 20 drawings loaded.", failMsg: "Context limit exceeded, unable to process 20 drawings simultaneously." }, }, }, }; // ── Other bank questions: fully scripted for 20 sheets (the money // scenario); other corpora use the generic asymmetry builder. ── SCENARIOS.valves = { 20: { intui: { verdict: "correct", latency: 12.8, ms: 8600, evidence: true, answer: "There are 2 control valves on the debutanizer overhead line: FV1102 (flow) and PV1108 (pressure). Both sit on sheet R279032044P01 and connect to the overhead line node in the knowledge graph.", note: "Matches ground truth, FV1102 and PV1108, R279032044P01." }, gpt: { verdict: "incorrect", latency: 7.0, ms: 6000, answer: "There are 3 control valves on the overhead line: FV1102, PV1108, and TV1110, handling flow, pressure, and temperature respectively.", note: "Ground truth is 2 valves. TV1110 is on the reboiler line, not the overhead." }, claude: { verdict: "partial", latency: 8.4, ms: 7000, answer: "I can identify at least FV1102 and PV1108 on the overhead line, but I cannot guarantee the count is complete across all 20 sheets.", note: "Both correct valves found, declined to commit to the total." }, gemini: { verdict: "failed", latency: 6.4, ms: 5200, failAt: 0.4, answer: "Scanning the overhead line across the loaded sheets, the control valves present are", note: "No answer produced, context limit hit with 20 drawings loaded.", failMsg: "Context limit exceeded, unable to process 20 drawings simultaneously." }, }, }; SCENARIOS.relief = { 20: { intui: { verdict: "correct", latency: 13.4, ms: 8800, evidence: true, answer: "Yes. PSV1140 on the overhead drum discharges to the flare header on sheet R279032044P03. The knowledge graph traces a continuous path: drum → PSV1140 → flare header.", note: "Matches ground truth, PSV1140 to flare, R279032044P03." }, gpt: { verdict: "partial", latency: 7.2, ms: 6200, answer: "Yes, there is a relief path. A pressure safety valve relieves the drum to the flare system, though I cannot confirm the exact tag from the drawings.", note: "Correct conclusion, could not cite the device (PSV1140)." }, claude: { verdict: "partial", latency: 8.8, ms: 7200, answer: "There is likely a relief path to flare, but I cannot confidently trace it or cite the relief device across the 20 sheets.", note: "Hedged, did not identify PSV1140 or the sheet." }, gemini: { verdict: "failed", latency: 6.0, ms: 5000, failAt: 0.42, answer: "Tracing the overhead drum connections toward the flare header, the relief device", note: "No answer produced, context limit hit with 20 drawings loaded.", failMsg: "Context limit exceeded, unable to process 20 drawings simultaneously." }, }, }; SCENARIOS.connect = { 20: { intui: { verdict: "correct", latency: 13.0, ms: 8600, evidence: true, answer: "Bottoms pump P-110A feeds reboiler E-114, which returns vapor to the column. The path P-110A → E-114 → column is on sheet R279032044P02 and is a connected subgraph.", note: "Matches ground truth, P-110A to E-114 to column, R279032044P02." }, gpt: { verdict: "incorrect", latency: 7.3, ms: 6100, answer: "P-110A is the reflux pump; it feeds the overhead accumulator and returns reflux to the top of the column.", note: "Ground truth: P-110A is the bottoms pump feeding reboiler E-114, not reflux." }, claude: { verdict: "partial", latency: 8.6, ms: 7100, answer: "P-110A appears to be a bottoms pump feeding a downstream exchanger, but I cannot confidently identify the exchanger tag across the set.", note: "Right role, did not resolve E-114 or the sheet." }, gemini: { verdict: "failed", latency: 6.2, ms: 5100, failAt: 0.43, answer: "Following the discharge of P-110A through the loaded drawings, the downstream equipment", note: "No answer produced, context limit hit with 20 drawings loaded.", failMsg: "Context limit exceeded, unable to process 20 drawings simultaneously." }, }, }; // ── Generic builder ───────────────────────────────────────────── // For free-typed questions and bank questions in corpora we didn't // hand-script. No ground truth, so NO correctness verdict, instead // the asymmetry stands on its own: ours attaches structured // evidence, competitors return unverified text. function buildGeneric(corpus, text) { const big = corpus >= 20; return { intui: { verdict: "unverified", latency: big ? 13.6 : corpus >= 10 ? 9.4 : 5.2, ms: big ? 8800 : corpus >= 10 ? 7000 : 4400, evidence: true, answer: "Parsed " + corpus + " sheet" + (corpus > 1 ? "s" : "") + " into the knowledge graph and resolved the relevant tags, lines, and connections. The structured answer below is grounded in the highlighted symbols and cited to a specific sheet.", note: "Grounded in parsed evidence, open the evidence panel to verify." }, gpt: { verdict: "unverified", latency: big ? 7.2 : 6.4, ms: big ? 6200 : 5400, answer: "Here is a plausible answer based on the drawing text. I'm reasoning over the document as flat text, so I can't point to a specific symbol or sheet to back this up.", note: "Plain-text answer, no structured evidence." }, claude: { verdict: "unverified", latency: big ? 8.6 : 7.0, ms: big ? 7000 : 6000, answer: "I can offer an interpretation, but without a structured parse of the drawings I can't verify it against specific symbols or connectivity.", note: "Plain-text answer, no structured evidence." }, gemini: big ? { verdict: "failed", latency: 6.2, ms: 5200, failAt: 0.42, answer: "Working through the loaded drawing set, the relevant elements", note: "No answer produced, context limit hit with 20 drawings loaded.", failMsg: "Context limit exceeded, unable to process 20 drawings simultaneously." } : { verdict: "unverified", latency: 6.8, ms: 5800, answer: "Based on the text I can extract, here is a best-effort response. I'm not able to cite a specific symbol or sheet as evidence.", note: "Plain-text answer, no structured evidence." }, }; } // Resolve a question (by id or free text) + corpus into a race. function resolveRace(corpus, questionId, freeText) { if (questionId && SCENARIOS[questionId] && SCENARIOS[questionId][corpus]) { return { results: SCENARIOS[questionId][corpus], verified: true }; } // Known bank question but un-scripted corpus → generic, but keep the // ground-truth correctness flavor by tagging ours correct. if (questionId && SCENARIOS[questionId]) { const g = buildGeneric(corpus, freeText || ""); g.intui.verdict = "correct"; return { results: g, verified: true }; } return { results: buildGeneric(corpus, freeText || ""), verified: false }; } Object.assign(window, { CORPORA, MODELS, QUESTIONS, SCENARIOS, resolveRace, });