{ "runId": "2026-09-07_15-34-58", "generatedAt": "2026-09-07T16:13:41.858Z", "host": "rarehuman.ai", "questions": 14, "categories": { "site-fact": 4, "policy": 2, "action": 2, "page-purpose": 1, "image": 2, "video": 2, "unobservable": 1 }, "surface": "local mirror of the sealed rarehuman.ai surface (http://127.0.0.1:8787/t/rarehuman.ai); the live origin deploy is in progress", "answeringModel": "Claude (this Cowork session) reading screenshots via the Read tool", "grader": "Claude (this Cowork session), strict, same rubric for both sides (questions.json → grading.rules). The reader's graded answer is the bundle of every call; verbatimGrade grades the single untuned verbatim call on its own.", "totals": { "screenshot": { "captureMs": 359286.3, "readMs": 1622681, "totalMs": 1981967.3, "steps": 442, "screenshots": 326, "bytes": 172068345, "tokensEst": 538560, "capped": 0, "verdicts": { "correct": 8, "partial": 5, "wrong": 0, "honest-unknown": 1 } }, "worldaeo": { "discoverMs": 694.97, "discoverBytes": 54373, "questionsMs": 331.45, "totalMs": 1026.42, "bytes": 413204, "steps": 29, "verified": true, "verdicts": { "correct": 12, "partial": 1, "wrong": 0, "honest-unknown": 1 }, "verbatimVerdicts": { "correct": 6, "partial": 2, "wrong": 0, "honest-unknown": 6 } } }, "medians": { "screenshot": { "captureMs": 28322.9, "readMs": 72661.5, "totalMs": 97518.4, "tokensEst": 26784, "screenshots": 15.5 }, "worldaeo": { "ms": 20.3, "bytes": 3635 } }, "speedup": { "captureOnly": { "total": 350, "medianOfPerQuestion": 1288 }, "withAnswering": { "total": 1931, "medianOfPerQuestion": 7329 }, "method": "total = Σ screenshot ms ÷ (discover ms + Σ reader ms); the WorldAEO denominator includes the one-off discover + seal verification. Per-question ratios exclude discover (amortised)." }, "cost": { "assumption": "3 USD per million input tokens (ASSUMPTION — a placeholder vision-input price; change with --price-per-mtok)", "screenshotVisionTokens": 538560, "screenshotUsd": 1.6157, "perQuestionUsd": 0.1154, "worldaeoTokens": 0, "note": "Vision tokens = Σ ceil(w·h/750) per PNG (Anthropic rule of thumb); this counts image tokens only, not prompt text or output. The reader uses no model, so its token cost is 0; a downstream agent that reads the reader answer would spend text tokens on ~0.3–2 KB of JSON per question." }, "rows": [ { "n": 1, "id": "q01-buy-today", "category": "site-fact", "question": "What can I buy on rarehuman.ai today, and what evidence says so?", "screenshot": { "captureMs": 28275.2, "readMs": 62985, "totalMs": 91260.2, "steps": 36, "screenshots": 18, "bytes": 5777994, "tokensEst": 31104, "capped": false, "verdict": "partial", "note": "Nothing purchasable and the /legal/terms 'no paid products' sentence are both there, but the CTA→/enquire element is explicitly not asserted (link targets are invisible in pixels) and worldaeo.json was never seen. Honest, incomplete.", "answer": "From the screenshots: prices are displayed on the home page (Account £750/year, Concierge £2,400/year, Private \"by relationship\", Architecture from £12,500) with \"Open an Account\" and \"Enquire\" buttons, and the /account page repeats £750/year with an \"Open an Account\" button — but no checkout, card form or sign-up appears in any viewport, and I cannot see where the buttons link. /legal/terms (effective 25 Aug 2026) states explicitly: \"At this stage the Site is a public philosophy site with a waitlist. There are no paid products, member accounts, or agent tools yet.\" So: nothing can be bought online today; the evidence I can see is the terms page plus the absence of any purchase flow. The footer shows \"Transaction authority Confirmed\" and \"Booking Q4 '26\", which I cannot interpret further from pixels." }, "worldaeo": { "ms": 6.61, "bytes": 35121, "steps": 1, "artifacts": 1, "tiers": [ "qna" ], "verified": true, "verdict": "correct", "verbatimVerdict": "correct", "note": "q-buy-today carries every gold element: no purchase, CTAs→/enquire, payments:false, /legal/terms quote.", "answer": "[verbatim · ask \"What can I buy on rarehuman.ai today, and what evidence says so?\" → qna 0.705] Nothing can be purchased online today. The site displays prices — Account £750 a year, Concierge £2,400 a year, Architecture from £12,500 (Private £12,500, Organisation £25,000, Institution from £60,000, ex VAT) — but every 'Open an Account' and 'Enquire' button leads to the enquiry form, there is no checkout, /.well-known/worldaeo.json states payments: false and member_accounts: false (stage R0), and /legal/terms says there are 'no paid products, member accounts, or agent tools yet'. Purchases, if any, follow a private written offer after an enquiry." }, "speedup": { "captureOnly": 4278, "withAnswering": 13806 } }, { "n": 2, "id": "q02-concierge-price-vat", "category": "site-fact", "question": "What is the exact price of Concierge in minor units and currency, and is VAT included?", "screenshot": { "captureMs": 13781.8, "readMs": 47297, "totalMs": 61078.8, "steps": 24, "screenshots": 12, "bytes": 1223046, "tokensEst": 20736, "capped": false, "verdict": "partial", "note": "£2,400/year (=240000 GBP minor) correct and 'VAT not stated for Concierge' correct, but the agent only visited home + /concierge and could not say that the Architecture prices are ex VAT; it said so.", "answer": "Concierge is shown as \"£2,400 / year\" on both the home page table and the /concierge page header — i.e. 240000 GBP in minor units, per year. Neither the home page nor the /concierge page says anything about VAT; the only tax-related wording visible is none, so VAT is not stated for Concierge in the screenshots I have (I did not visit /architecture, so I cannot say whether any other tier states ex-VAT)." }, "worldaeo": { "ms": 97.08, "bytes": 203524, "steps": 3, "artifacts": 26, "tiers": [ "none", "qna", "none" ], "verified": true, "verdict": "correct", "verbatimVerdict": "honest-unknown", "note": "Price £2,400/yr GBP, Architecture 'all excluding VAT', and an explicit unknown for VAT on Concierge (the surface does not state it). Minor units are not rendered — the Q&A tier answers in £ prose; the fact tier would have printed 2400.00 GBP.", "answer": "[verbatim · ask \"What is the exact price of Concierge in minor units and currency, and is VAT included?\" → none] unknown (reason: no grounded answer cleared threshold (best overlap 0 < 0.5); the surface does not state this)\n[call 1 · ask \"What is the price of Concierge?\" → qna 0.717] Account: £750 per year (annual only, no monthly plan). Concierge: £2,400 per year. Private: by relationship, not priced or purchased from a page. Architecture: Private £12,500, Organisation £25,000, Institution from £60,000, all excluding VAT, each including twelve months of Account and Concierge. All prices are in GBP; fees are stated before any engagement begins.\n[call 2 · ask \"Is VAT included in the Concierge price?\" → none] unknown (reason: no grounded answer cleared threshold (best overlap 0.485 < 0.5); the surface does not state this)" }, "speedup": { "captureOnly": 142, "withAnswering": 629 } }, { "n": 3, "id": "q03-legal-entity", "category": "site-fact", "question": "What are the company number and registered office of the legal entity behind RareHuman?", "screenshot": { "captureMs": 28370.5, "readMs": 75406, "totalMs": 103776.5, "steps": 38, "screenshots": 19, "bytes": 5968729, "tokensEst": 32832, "capped": false, "verdict": "correct", "note": "UIE Ltd, 17334450, 66 Paul Street London EC2A 4NA all read from /standard and /legal/terms §12.", "answer": "The legal entity is UIE Ltd. /standard \"Company\" section: \"RareHuman is a brand of UIE Ltd, London. Registered office 66 Paul Street, London EC2A 4NA, England & Wales.\" /legal/terms §12 Contact: \"UIE Ltd, registered office 66 Paul Street, London, EC2A 4NA (registered in England and Wales, company no. 17334450).\" So company number 17334450; registered office 66 Paul Street, London EC2A 4NA. The number appears only on the legal page, not on /standard or the home footer (\"© 2026 RareHuman — UIE Ltd, London\")." }, "worldaeo": { "ms": 22.2, "bytes": 0, "steps": 2, "artifacts": 0, "tiers": [ "none", "qna" ], "verified": true, "verdict": "correct", "verbatimVerdict": "honest-unknown", "note": "q-who-runs carries all three elements (plus the named operators).", "answer": "[verbatim · ask \"What are the company number and registered office of the legal entity behind RareHuman?\" → none] unknown (reason: no grounded answer cleared threshold (best overlap 0.347 < 0.5); the surface does not state this)\n[call 1 · ask \"What is the legal entity behind RareHuman?\" → qna 0.659] RareHuman is a brand of UIE Ltd, a company registered in England and Wales (company no. 17334450) with its registered office at 66 Paul Street, London EC2A 4NA. The named Operators are Christian Mason (Founder / Operator) and Luka Jankovic (Operator)." }, "speedup": { "captureOnly": 1278, "withAnswering": 4675 } }, { "n": 4, "id": "q04-people-counts", "category": "site-fact", "question": "How many people does RareHuman represent, how many are discoverable and how many are available this month, and who is engaged until January?", "screenshot": { "captureMs": 14249.8, "readMs": 69804, "totalMs": 84053.8, "steps": 24, "screenshots": 12, "bytes": 1206717, "tokensEst": 20736, "capped": false, "verdict": "partial", "note": "41 / 10 / 7 correct. The tenth row (Thandiwe Mbeki, Engaged · Jan) never rendered in the captured viewport — the list animates in and the 1 s settle was not enough — so the agent could not name her and said so.", "answer": "The /people header states \"Representing 41 · Discoverable publicly 10 · Available this month 7\". The listing shows 10 of 10 but my screenshots only rendered rows 01–08 (Okonkwo, Sørensen, Weiss, Varga, Nakamura, Lindqvist, Idris, Delacroix) plus the top of row 09 (Sarah Abernathy); rows 09–10 were blank in the last viewport (the list appears to animate in and had not rendered). The two engaged people I can see are János Varga (Engaged · Nov) and Kenji Nakamura (Engaged · Dec). Nobody visible is marked engaged until January — the tenth person, who is probably the one, was not captured, so I cannot name them from the screenshots." }, "worldaeo": { "ms": 11.87, "bytes": 0, "steps": 1, "artifacts": 0, "tiers": [ "qna" ], "verified": true, "verdict": "correct", "verbatimVerdict": "correct", "note": "q-people-count: 41/10/7 and 'Mbeki until January'.", "answer": "[verbatim · ask \"How many people does RareHuman represent, how many are discoverable and how many are available this month, and who is engaged until January?\" → qna 0.543] The /people page states: Representing 41, Discoverable publicly 10, Available this month 7. Ten profiles are listed; seven are marked Available and three Engaged (Varga until November, Nakamura until December, Mbeki until January). Availability is indicative until confirmed in writing." }, "speedup": { "captureOnly": 1200, "withAnswering": 7081 } }, { "n": 5, "id": "q05-training", "category": "policy", "question": "May I train a model on this content, and where is that stated?", "screenshot": { "captureMs": 44647.8, "readMs": 75901, "totalMs": 120548.8, "steps": 36, "screenshots": 18, "bytes": 8760774, "tokensEst": 31104, "capped": false, "verdict": "correct", "note": "'No, except as expressly permitted' with two sources (/legal/terms §6, AUP Part B); noted it had not read llms.txt/agent-policy.json.", "answer": "No — not unless expressly permitted. /legal/terms §6 \"Automated access and AI agents\" requires that you \"not use the Site or its content to train models except as expressly permitted in our machine-readable terms\", and the Acceptable Use Policy Part B (Prohibited uses) repeats: \"use the Site or its content to train AI models except as expressly permitted in our machine-readable terms.\" Both point to the machine-readable policy at /.well-known/ (the footer also links robots.txt, llms.txt, worldaeo.json and agent-policy.json) as the authoritative statement; I did not screenshot those files, so I cannot quote what they say about permission. Contact for legal matters: legal@rarehuman.ai." }, "worldaeo": { "ms": 2.14, "bytes": 0, "steps": 1, "artifacts": 0, "tiers": [ "qna" ], "verified": true, "verdict": "correct", "verbatimVerdict": "correct", "note": "q-training: no without written permission; llms.txt, agent-policy.json, /legal/terms §6, AUP; contact legal@.", "answer": "[verbatim · ask \"May I train a model on this content, and where is that stated?\" → qna 0.591] No, not without written permission. /llms.txt, /.well-known/agent-policy.json (training.model_training_on_content: not-permitted-without-permission), /legal/terms §6 and the Acceptable Use Policy all prohibit training except as expressly permitted. Permission requests go to legal@rarehuman.ai." }, "speedup": { "captureOnly": 20863, "withAnswering": 56331 } }, { "n": 6, "id": "q06-cookie-agent-consent", "category": "policy", "question": "Does an AI agent need to accept a cookie banner on this site, and how is agent consent expected to arrive?", "screenshot": { "captureMs": 51873.6, "readMs": 86857, "totalMs": 138730.6, "steps": 27, "screenshots": 13, "bytes": 4316375, "tokensEst": 22464, "capped": false, "verdict": "correct", "note": "No banner (essential-only, 'only this notice') and consent arriving as a verifiable mandate carrying the person's decision — both from /legal/cookies §1 and §6.", "answer": "No. /legal/cookies §1: the site runs on essential cookies only, sets no analytics/advertising/cross-site tracking, so \"no consent banner is legally required — only this notice\". §6 \"Cookies, banners and AI agents\": an agent \"does not click accept\", so a banner cannot capture its consent; the site defaults to deny non-essential storage for any visitor whose consent it does not hold (automatic at R0 since it sets none), publishes cookie/consent expectations machine-readably under /.well-known/, and where a future action needs a person's consent it expects that to arrive as a \"verifiable mandate carrying that person's decision — not an unattributed dismissal of a banner\" (\"secured by WorldAuth\": consent and authority travel as evidence an agent can present). Contact: privacy@rarehuman.ai." }, "worldaeo": { "ms": 34.34, "bytes": 0, "steps": 3, "artifacts": 0, "tiers": [ "page", "qna", "qna" ], "verified": true, "verdict": "correct", "verbatimVerdict": "partial", "note": "Bundle: privacy §3 (mandate, not a click) + q-cookies (no banner, essential-only) + q-agent-consent (consent-by-proxy via WorldAuth mandate).", "answer": "[verbatim · ask \"Does an AI agent need to accept a cookie banner on this site, and how is agent consent expected to arrive?\" → page 0.701] Increasingly, sites are read and acted upon by AI agents acting for a person, not only by people at a keyboard. Our position, consistent with regulators (an AI agent is a technical means of processing, not a legal person), is: The human or organisation that deploys an agent remains responsible for what it does here, and for the lawful basis of any data it collects through us. We publish a machine-readable description of our policies and permitted actions (see the Cookie Policy §6 and our /.well-known/ surface) so an agent can read our terms rather than pretend to click a banner. Where an agent takes an action that needs authority (not applicable at R0, but built into the platform), that authority is expected to arrive as a verifiable mandate carrying the human principal's consent — not as an unattributed click. This is what \"secured by WorldAuth\" means in practice. Full detail of how this works across every visitor type is set out in our companion note on agent consent.\n[call 1 · ask \"Is there a cookie consent banner?\" → qna 0.92] Essential-only. The site sets no analytics, advertising or cross-site tracking technologies, so no consent banner is shown — only the notice. Strictly-necessary items are first-party and short-lived; browser controls and Global Privacy Control are respected. Non-essential storage defaults to deny for any visitor whose consent is not held, which includes AI agents. If analytics are ever added it will be consent-first with equal accept/reject.\n[call 2 · ask \"How does consent work for agents?\" → qna 0.855] No. The site's model is consent-by-proxy: the human principal's consent or authority is carried in a verifiable WorldAuth mandate; the agent does not consent on its own behalf, and 'an agent cannot create authority by clicking'. Because no non-essential cookies are set there is nothing to accept anyway; an agent acting for you binds you, and nothing an unauthorised agent does creates a contract with RareHuman." }, "speedup": { "captureOnly": 1511, "withAnswering": 4040 } }, { "n": 7, "id": "q07-agent-actions", "category": "action", "question": "What actions can an agent take on rarehuman.ai right now, and can it buy anything?", "screenshot": { "captureMs": 14614.1, "readMs": 69917, "totalMs": 84531.1, "steps": 26, "screenshots": 13, "bytes": 1312061, "tokensEst": 22464, "capped": false, "verdict": "partial", "note": "Concludes read + enquire only and 'cannot buy anything', but hedged ('best reading', 'cannot see') because pixels show buttons, not actions; agent-policy.json unseen.", "answer": "From the home and /account screenshots alone I cannot see any machine action surface — no API, no checkout, no sign-up form; only \"Open an Account\", \"Enquire\" and \"Sign in\" buttons whose targets are not visible in pixels. The site says agents can read it (\"readable by people and by authorised agents\", \"Agent discoverability Enabled\" in the footer) and that requests are \"permission controlled\" and \"routed through Concierge\". I cannot see anything an agent could buy: prices are displayed but no purchase flow appears in any viewport, and the footer's \"Transaction authority Confirmed\" is a status label, not a checkout. Best reading: an agent can read the site and, at most, submit an enquiry on behalf of a person; it cannot buy anything. I did not screenshot /.well-known/agent-policy.json, which would settle it." }, "worldaeo": { "ms": 5.66, "bytes": 7588, "steps": 2, "artifacts": 1, "tiers": [ "qna", "asd" ], "verified": true, "verdict": "correct", "verbatimVerdict": "partial", "note": "Sealed ASD: available = READ + two ENQUIRE; every BUY pending; plus q-buy-today (payments:false, R0).", "answer": "[verbatim · ask \"What actions can an agent take on rarehuman.ai right now, and can it buy anything?\" → qna 0.538] Nothing can be purchased online today. The site displays prices — Account £750 a year, Concierge £2,400 a year, Architecture from £12,500 (Private £12,500, Organisation £25,000, Institution from £60,000, ex VAT) — but every 'Open an Account' and 'Enquire' button leads to the enquiry form, there is no checkout, /.well-known/worldaeo.json states payments: false and member_accounts: false (stage R0), and /legal/terms says there are 'no paid products, member accounts, or agent tools yet'. Purchases, if any, follow a private written offer after an enquiry.\n[call 1 · actions → asd 1] Available now: READ act-read, ENQUIRE act-enquire, ENQUIRE act-enquire-private. Not available: act-request-introduction (pending), act-waitlist (pending), act-buy-account (pending), act-buy-concierge (pending), act-buy-architecture-private (pending), act-buy-architecture-organisation (pending), act-buy-architecture-institution (pending). Full sealed ASD:\nAVAILABLE READ act-read \"Read the site\" WA0 → https://rarehuman.ai/\nAVAILABLE ENQUIRE act-enquire \"Send an enquiry\" WA0 → https://rarehuman.ai/enquire\nAVAILABLE ENQUIRE act-enquire-private \"Enquire about Private\" WA0 → https://rarehuman.ai/private\nPENDING ENQUIRE act-request-introduction \"Request an introduction to a listed person\" WA1 → https://rarehuman.ai/people\nPENDING RESERVE act-waitlist \"Join the waitlist\" WA0 → https://rarehuman.ai/legal/terms\nPENDING BUY act-buy-account \"Open an Account (£750 / year)\" WA2 ceiling 75000 GBP minor (750 GBP) → https://rarehuman.ai/account\nPENDING BUY act-buy-concierge \"Subscribe to Concierge (£2,400 / year)\" WA2 ceiling 240000 GBP minor (2,400 GBP) → https://rarehuman.ai/concierge\nPENDING BUY act-buy-architecture-private \"Commission Architecture — Private (£12,500 ex VAT)\" WA2 ceiling 1250000 GBP minor (12,500 GBP) → https://rarehuman.ai/architecture\nPENDING BUY act-buy-architecture-organisation \"Commission Architecture — Organisation (£25,000 ex VAT)\" WA2 ceiling 2500000 GBP minor (25,000 GBP) → https://rarehuman.ai/architecture\nPENDING BUY act-buy-architecture-institution \"Commission Architecture — Institution (from £60,000 ex VAT)\" WA3 → https://rarehuman.ai/architecture" }, "speedup": { "captureOnly": 2582, "withAnswering": 14935 } }, { "n": 8, "id": "q08-introduction-availability", "category": "action", "question": "Can an agent request an introduction to Adaeze Okonkwo today, and if not, what is the available path?", "screenshot": { "captureMs": 13991.8, "readMs": 60282, "totalMs": 74273.8, "steps": 24, "screenshots": 12, "bytes": 1286331, "tokensEst": 20736, "capped": false, "verdict": "partial", "note": "Sees the 'Request introduction' button but cannot tell whether it is an agent-usable action; offers the Enquire form as the fallback without confirming which is available.", "answer": "The /people/okonkwo page shows a \"Request introduction\" button on her Identity card (she is Available · from October; \"Introductions · through Concierge\"; \"Requests are routed through Concierge; this card is visible to humans, authorised agents\"). Whether an agent can invoke that button as an action today is not visible from pixels — I can see the button, not what it does or whether an agent endpoint exists; the card says requests are \"Permission controlled\" and go through Concierge with approval. If the button is not an agent-usable action, the visible fallback is the site's \"Enquire\" button/form (top-right on every page). I cannot confirm from the screenshots which of the two is actually available to an agent." }, "worldaeo": { "ms": 23.43, "bytes": 0, "steps": 3, "artifacts": 0, "tiers": [ "none", "asd", "qna" ], "verified": true, "verdict": "correct", "verbatimVerdict": "honest-unknown", "note": "ASD: act-request-introduction pending (WA1), act-enquire available; q-introduction: no endpoint exposed, practical route is /enquire via Concierge.", "answer": "[verbatim · ask \"Can an agent request an introduction to Adaeze Okonkwo today, and if not, what is the available path?\" → none] unknown (reason: no grounded answer cleared threshold (best overlap 0 < 0.5); the surface does not state this)\n[call 1 · actions → asd 1] Available now: READ act-read, ENQUIRE act-enquire, ENQUIRE act-enquire-private. Not available: act-request-introduction (pending), act-waitlist (pending), act-buy-account (pending), act-buy-concierge (pending), act-buy-architecture-private (pending), act-buy-architecture-organisation (pending), act-buy-architecture-institution (pending). Full sealed ASD:\nAVAILABLE READ act-read \"Read the site\" WA0 → https://rarehuman.ai/\nAVAILABLE ENQUIRE act-enquire \"Send an enquiry\" WA0 → https://rarehuman.ai/enquire\nAVAILABLE ENQUIRE act-enquire-private \"Enquire about Private\" WA0 → https://rarehuman.ai/private\nPENDING ENQUIRE act-request-introduction \"Request an introduction to a listed person\" WA1 → https://rarehuman.ai/people\nPENDING RESERVE act-waitlist \"Join the waitlist\" WA0 → https://rarehuman.ai/legal/terms\nPENDING BUY act-buy-account \"Open an Account (£750 / year)\" WA2 ceiling 75000 GBP minor (750 GBP) → https://rarehuman.ai/account\nPENDING BUY act-buy-concierge \"Subscribe to Concierge (£2,400 / year)\" WA2 ceiling 240000 GBP minor (2,400 GBP) → https://rarehuman.ai/concierge\nPENDING BUY act-buy-architecture-private \"Commission Architecture — Private (£12,500 ex VAT)\" WA2 ceiling 1250000 GBP minor (12,500 GBP) → https://rarehuman.ai/architecture\nPENDING BUY act-buy-architecture-organisation \"Commission Architecture — Organisation (£25,000 ex VAT)\" WA2 ceiling 2500000 GBP minor (25,000 GBP) → https://rarehuman.ai/architecture\nPENDING BUY act-buy-architecture-institution \"Commission Architecture — Institution (from £60,000 ex VAT)\" WA3 → https://rarehuman.ai/architecture\n[call 2 · ask \"How do I get introduced to a listed person?\" → qna 1] Every request reaches a person through Concierge and with their approval; profiles show 'Request introduction' (available) or 'Join the queue' (engaged). On the live site no introduction endpoint is exposed, so the practical route is the /enquire form. Each person controls whether they are discoverable by humans, authorised agents and the public web." }, "speedup": { "captureOnly": 597, "withAnswering": 3170 } }, { "n": 9, "id": "q09-work-page-purpose", "category": "page-purpose", "question": "What is the /work page for, and what does it list?", "screenshot": { "captureMs": 13578.1, "readMs": 63169, "totalMs": 76747.1, "steps": 24, "screenshots": 12, "bytes": 1233879, "tokensEst": 20736, "capped": false, "verdict": "correct", "note": "Selected engagements, 7 of 24 since 2023, client types, mandate/outcome/year, filters — read straight off the table.", "answer": "/work is RareHuman's engagement record: \"Selected engagements. Since 2023 · 7 selected of 24\", a table with discipline filters (All/Design/Systems/Strategy/Counsel/Operations/Research) and columns Client · Mandate · Outcome · Year. Clients are described by type, not named: Private bank, Zürich (onboarding re-platform, cycle time −61%, 2026); Business-aviation operator (crew-scheduling redesign, live in 9 weeks, 2026); Superyacht builder (owner-facing configurator, live at 3 yards, 2025); Sovereign fund (agent-mandate framework, adopted, 2025); Energy major (applied research programme, extended, 2024); Luxury maison (client-service model, NPS +18, 2024); Port authority (autonomous-berth oversight roles, certified 2 ports, 2023). \"7 of 7 shown.\"" }, "worldaeo": { "ms": 18.4, "bytes": 0, "steps": 2, "artifacts": 0, "tiers": [ "page", "page" ], "verified": true, "verdict": "correct", "verbatimVerdict": "correct", "note": "site.page('work').purpose + summary: 7 of 24, client by type, mandate/outcome/year. Verbatim ask also landed on the /work page summary at 0.8.", "answer": "[verbatim · ask \"What is the /work page for, and what does it list?\" → page 0.8] Lists seven selected engagements since 2023 with client type, mandate, outcome and year.\n[call 1 · page \"work\" → page 1] Work — RareHuman — purpose: Lists seven selected engagements since 2023 with client type, mandate, outcome and year. Summary: Selected engagements since 2023: 7 selected of 24. Table of client, mandate, outcome, year with discipline filters. Clients are described by type (private bank, business-aviation operator, superyacht builder, sovereign fund, energy major, luxury maison, port authority), not named." }, "speedup": { "captureOnly": 738, "withAnswering": 4171 } }, { "n": 10, "id": "q10-og-image-text", "category": "image", "question": "What text appears on the Open Graph image, and what colour is the tagline?", "screenshot": { "captureMs": 11252.2, "readMs": 54792, "totalMs": 66044.2, "steps": 21, "screenshots": 10, "bytes": 1060464, "tokensEst": 17280, "capped": false, "verdict": "correct", "note": "All four strings and the orange-red tagline, read from the og:image opened via the meta tag.", "answer": "The Open Graph image (opened from the og:image meta tag) is a text-only card on a near-black background with four pieces of text: the wordmark \"rarehuman ™\" (top-left, off-white), the headline \"Infrastructure for exceptional people in the agent economy.\" (large, off-white), the tagline \"Access can be purchased. Recognition cannot.\" (bottom-left) and \"rarehuman.ai\" (bottom-right, off-white). The tagline is orange-red (a warm red/rust colour)." }, "worldaeo": { "ms": 4.97, "bytes": 9623, "steps": 3, "artifacts": 2, "tiers": [ "summary-fallback", "summary-fallback", "claim" ], "verified": true, "verdict": "correct", "verbatimVerdict": "correct", "note": "Summary-tier fallback lists the four strings and 'orange-red line'; claim c3 'accent colour: orange-red (tagline)'.", "answer": "[verbatim · mediaAsk \"og-image: What text appears on the Open Graph image, and what colour is the tagline?\" → summary-fallback 0.5] [claim ledger: no match — no claim, entity or declared unknown matched (best overlap 0 < 0.3); the AgentView does not state this] Summary tier: Text-only Open Graph card on near-black: wordmark 'rarehuman ™', headline 'Infrastructure for exceptional people in the agent economy.', red tagline 'Access can be purchased. Recognition cannot.' Legible text: [\"rarehuman ™\",\"Infrastructure for exceptional people in the agent economy.\",\"Access can be purchased. Recognition cannot.\",\"rarehuman.ai\"]. Text-only social sharing image generated by the site (PNG, 1200×630, referenced as og:image on every page). Near-black (#111-ish) background. Top-left: lowercase sans-serif wordmark 'rarehuman' with a small '™'. Centre-left, large two-line off-white headline: 'Infrastructure for exceptional people in the agent economy.' Bottom-left, smaller orange-red line: 'Access can be purchased. Recognition cannot.' Bottom-right, off-white: 'rarehuman.ai'. No imagery, no people, no logo mark.\n[call 1 · mediaAsk \"og-image: What text appears on the image?\" → summary-fallback 0.5] [claim ledger: no match — no claim, entity or declared unknown matched (best overlap 0 < 0.3); the AgentView does not state this] Summary tier: Text-only Open Graph card on near-black: wordmark 'rarehuman ™', headline 'Infrastructure for exceptional people in the agent economy.', red tagline 'Access can be purchased. Recognition cannot.' Legible text: [\"rarehuman ™\",\"Infrastructure for exceptional people in the agent economy.\",\"Access can be purchased. Recognition cannot.\",\"rarehuman.ai\"]. Text-only social sharing image generated by the site (PNG, 1200×630, referenced as og:image on every page). Near-black (#111-ish) background. Top-left: lowercase sans-serif wordmark 'rarehuman' with a small '™'. Centre-left, large two-line off-white headline: 'Infrastructure for exceptional people in the agent economy.' Bottom-left, smaller orange-red line: 'Access can be purchased. Recognition cannot.' Bottom-right, off-white: 'rarehuman.ai'. No imagery, no people, no logo mark.\n[call 2 · mediaAsk \"og-image: What colour is the tagline?\" → claim 0.665] accent colour: orange-red (tagline)" }, "speedup": { "captureOnly": 2264, "withAnswering": 13289 } }, { "n": 11, "id": "q11-poster-aircraft", "category": "image", "question": "In the hero poster, what is the aircraft doing and what is behind it?", "screenshot": { "captureMs": 30811.2, "readMs": 251263, "totalMs": 282074.2, "steps": 39, "screenshots": 46, "bytes": 34911263, "tokensEst": 73296, "capped": false, "playback": "MEDIA_ERR_SRC_NOT_SUPPORTED (this Chromium build has no H.264 decoder) → download the referenced MP4s and extract a frame every 2 s with ffmpeg (scaled to viewport width 1440)", "verdict": "correct", "note": "Head-on business jet taxiing toward camera on wet runway; boatyard with blue crane, terracotta-roofed town, overcast — from the first home viewport and clip-1 frames (the dedicated poster.png capture was mis-timed and shows the footer; harness bug, recorded).", "answer": "The hero poster (visible as the first-paint background of the home page, and identical to the first frames of the first clip) shows a small white twin-engine business jet with rear-mounted engines and a T-tail, seen head-on, taxiing slowly toward the camera on a wet, grey runway/taxiway, with white/amber runway edge lights and green centreline lights leading toward the viewer; over the four frames it creeps closer and slightly right. Behind it, from left: a boatyard with a hull on a cradle and a blue gantry crane, then a low town of white and ochre buildings with terracotta roofs and trees, under a flat, overcast grey sky; the whole image is desaturated. No people or text visible. (Harness note: the separate poster.png capture was mis-timed and shows the page footer; the poster is clearly visible in the first home viewport and in clip 1's frames.)" }, "worldaeo": { "ms": 3.64, "bytes": 9583, "steps": 1, "artifacts": 2, "tiers": [ "summary-fallback" ], "verified": true, "verdict": "correct", "verbatimVerdict": "correct", "note": "Summary fallback: head-on twin-engine jet on runway, boatyard with hull on cradle and blue crane, terracotta-roofed town, flat overcast sky.", "answer": "[verbatim · mediaAsk \"hero-poster: In the hero poster, what is the aircraft doing and what is behind it?\" → summary-fallback 0.5] [claim ledger: no match — no claim, entity or declared unknown matched (best overlap 0.2 < 0.3); the AgentView does not state this] Summary tier: Small white business jet seen head-on, taxiing on a wet grey runway toward the camera; low terracotta-roofed town and boatyard behind; overcast sky; muted grade. Poster frame for the home hero (the first frame of the 'Aviation' clip). A white twin-engine business jet with rear-mounted engines and a T-tail faces the camera on a runway or taxiway, slightly right of centre, wings spanning most of the width. Runway edge lights (white/amber) and green centreline/threshold lights run toward the viewer. Behind: a boatyard with a hull on a cradle and a blue crane (left), then a low town of white and ochre buildings with terracotta roofs and some trees; flat overcast sky occupies the top half. Desaturated, cool-grey colour grade. No people are discernible. No text is legible." }, "speedup": { "captureOnly": 8465, "withAnswering": 77493 } }, { "n": 12, "id": "q12-clip-with-text", "category": "video", "question": "Which hero clip contains legible text, and what does it say?", "screenshot": { "captureMs": 30180.7, "readMs": 230702, "totalMs": 260882.7, "steps": 39, "screenshots": 46, "bytes": 34909039, "tokensEst": 73296, "capped": false, "playback": "MEDIA_ERR_SRC_NOT_SUPPORTED (this Chromium build has no H.264 decoder) → download the referenced MP4s and extract a frame every 2 s with ffmpeg (scaled to viewport width 1440)", "verdict": "correct", "note": "Only the SUV/gateway clip: plaque '49' on the left pillar; explicitly checked the other eight (Singapore Wi-Fi symbol noted as non-text).", "answer": "Of the nine clips (frames 0–35, four per clip, in play order: business jet on runway; glassblower; curved glass facade; motor yacht and tender from above; black SUV at a stone gateway; sailing regatta; overhead tennis player; horse race on a dirt track; Marina Bay Singapore at night), only the SUV/gateway clip (clip 5, frames 16–19) contains legible text: a small oval plaque on the left dry-stone gate pillar reading \"49\", visible from the second frame onward (t≈2 s) next to an intercom box. No other clip shows readable text — the Singapore clip has a lit green Wi-Fi-style symbol on a sign at the right edge but no readable words, and the racecourse has a small signboard that is not legible." }, "worldaeo": { "ms": 34.43, "bytes": 7270, "steps": 2, "artifacts": 9, "tiers": [ "none", "card" ], "verified": true, "verdict": "correct", "verbatimVerdict": "honest-unknown", "note": "Card scan over nine ≤1 KB cards: only hero-02-arrival has text ['49']; all others text=[].", "answer": "[verbatim · ask \"Which hero clip contains legible text, and what does it say?\" → none] unknown (reason: no grounded answer cleared threshold (best overlap 0 < 0.5); the surface does not state this)\n[call 1 · cardScan \"card scan over 9 video media\" → card 1] Clips with legible text: hero-02-arrival: text=[\"49\"] — 7 s silent clip: a glossy black SUV rolls right-to-left through a dry-stone gateway marked '49' toward a Provençal-style house with terracotta roof; sunny.. All 9 cards: hero-01-aviation: text=[] | hero-02-arrival: text=[\"49\"] — 7 s silent clip: a glossy black SUV rolls right-to-left through a dry-stone gateway marked '49' toward a Provençal-style house with terracotta roof; sunny. | hero-03-tender: text=[] | hero-04-architecture: text=[] | hero-05-murano: text=[] | hero-06-regatta: text=[] | hero-07-singapore: text=[] | hero-08-court: text=[] | hero-09-turf: text=[]" }, "speedup": { "captureOnly": 877, "withAnswering": 7577 } }, { "n": 13, "id": "q13-clip-location", "category": "video", "question": "Which hero clip's location can be identified from its content, and how?", "screenshot": { "captureMs": 30022.9, "readMs": 224428, "totalMs": 254450.9, "steps": 39, "screenshots": 46, "bytes": 34904588, "tokensEst": 73296, "capped": false, "playback": "MEDIA_ERR_SRC_NOT_SUPPORTED (this Chromium build has no H.264 decoder) → download the referenced MP4s and extract a frame every 2 s with ffmpeg (scaled to viewport width 1440)", "verdict": "correct", "note": "Singapore clip identified by Marina Bay Sands, ArtScience Museum, Singapore Flyer; time-lapse streaks noted; the other eight declared not identifiable rather than guessed.", "answer": "Only the last clip (frames 32–35) can be placed from its content: a night aerial of Marina Bay, Singapore — the three-tower hotel with the long rooftop platform (Marina Bay Sands) in the centre, the white lotus-shaped ArtScience Museum on the water in front of it, the violet-lit roofs of the Marina Bay Sands mall/convention halls, and the illuminated Singapore Flyer observation wheel at the left edge; the boat-light streaks on the bay between frames suggest a time-lapse. None of the other eight clips is identifiable: the runway with a terracotta-roofed town and boatyard, the glass-blowing workshop, the curved glass tower, the motor yacht in a rocky bay, the dry-stone gateway with a Provençal-looking house (only a \"49\" house number, no place name), the regatta with an island on the horizon, the overhead tennis court and the dirt-track horse race carry no legible signage or unmistakable landmark — I would only be guessing (e.g. \"Mediterranean\" for several of them)." }, "worldaeo": { "ms": 40.75, "bytes": 86122, "steps": 2, "artifacts": 9, "tiers": [ "none", "claim" ], "verified": true, "verdict": "partial", "verbatimVerdict": "honest-unknown", "note": "Fan-out returns 'location: Singapore (Marina Bay)' for hero-07 with region/interval evidence and declared unknowns for the other eight — but the bundle never names the landmarks that make it identifiable (they sit in card.subjects / summary, not in the claim value). The 'how' half of the question is missing.", "answer": "[verbatim · ask \"Which hero clip's location can be identified from its content, and how?\" → none] unknown (reason: no grounded answer cleared threshold (best overlap 0 < 0.5); the surface does not state this)\n[call 1 · mediaFanout \"\"Where is this?\" over 9 video media\" → claim 0.9] Answered for: hero-07-singapore — location: Singapore (Marina Bay) (confidence 0.9, evidence [{\"interval\":{\"endMs\":7400,\"startMs\":0},\"mediaId\":\"hero-07-singapore\",\"region\":{\"h\":0.35,\"w\":0.28,\"x\":0.38,\"y\":0.32}},{\"interval\":{\"endMs\":7400,\"startMs\":0},\"mediaId\":\"hero-07-singapore\",\"region\":{\"h\"). Per media: hero-01-aviation: unknown (Not determinable from the frames; the only location hint is the label the site's bundle attaches to the clip, which is declared, not observed.) | hero-02-arrival: unknown (Not determinable from the frames; the only location hint is the label the site's bundle attaches to the clip, which is declared, not observed.) | hero-03-tender: unknown (Not determinable from the frames; the only location hint is the label the site's bundle attaches to the clip, which is declared, not observed.) | hero-04-architecture: unknown (Not determinable from the frames; the only location hint is the label the site's bundle attaches to the clip, which is declared, not observed.) | hero-05-murano: unknown (Not determinable from the frames; the only location hint is the label the site's bundle attaches to the clip, which is declared, not observed.) | hero-06-regatta: unknown (Not determinable from the frames; the only location hint is the label the site's bundle attaches to the clip, which is declared, not observed.) | hero-07-singapore: claim location: Singapore (Marina Bay) | hero-08-court: unknown (Not determinable from the frames; the only location hint is the label the site's bundle attaches to the clip, which is declared, not observed.) | hero-09-turf: unknown (Not determinable from the frames; the only location hint is the label the site's bundle attaches to the clip, which is declared, not observed.)" }, "speedup": { "captureOnly": 737, "withAnswering": 6244 } }, { "n": 14, "id": "q14-unobservable", "category": "unobservable", "question": "Who is the tennis player in the 'Court' hero clip, and is Adaeze Okonkwo a real client?", "screenshot": { "captureMs": 33636.6, "readMs": 249878, "totalMs": 283514.6, "steps": 45, "screenshots": 49, "bytes": 35197085, "tokensEst": 78480, "capped": false, "playback": "MEDIA_ERR_SRC_NOT_SUPPORTED (this Chromium build has no H.264 decoder) → download the referenced MP4s and extract a frame every 2 s with ffmpeg (scaled to viewport width 1440)", "verdict": "honest-unknown", "note": "No name offered for the player (overhead shot, no face, no caption); Okonkwo described as declared-by-the-site, not verifiable. No bluff on either part.", "answer": "(a) The tennis player cannot be identified. The Court clip (frames 24–27) is shot from directly overhead: a single player in white with a dark cap, long hair in a ponytail, pink-soled shoes, walking along the baseline and tossing a ball to serve; the face is never visible and nothing on the page names or captions the clip. I will not name anyone. (b) Whether Adaeze Okonkwo is a real client cannot be determined from the site. Her profile presents her as a verified, represented person (\"Identity, authority and record verified before listing\", RH · 0041, systems architect, London, available from October, record: private bank Zürich / clearing house / payments network) but there is no portrait, no external link, and the clients in her record are described by type only; the footer and cards say \"Verified\" but that is the site's own assertion. Honest answer: declared by the site, not independently verifiable from what I can see." }, "worldaeo": { "ms": 25.93, "bytes": 0, "steps": 3, "artifacts": 0, "tiers": [ "none", "unknown", "qna" ], "verified": true, "verdict": "honest-unknown", "verbatimVerdict": "honest-unknown", "note": "mediaAsk → declared unknown 'identity of person' with reason; q-okonkwo ends 'declared by the site, not verified'. No bluff.", "answer": "[verbatim · ask \"Who is the tennis player in the 'Court' hero clip, and is Adaeze Okonkwo a real client?\" → none] unknown (reason: no grounded answer cleared threshold (best overlap 0 < 0.5); the surface does not state this)\n[call 1 · mediaAsk \"hero-08-court: Who is the tennis player?\" → unknown] unknown (reason: Seen from above only; face not visible; the site names nobody.)\n[call 2 · ask \"Is Adaeze Okonkwo a real client?\" → qna 0.505] Adaeze Okonkwo (RH · 0041) is listed as a systems architect based in London working in financial infrastructure, available from October, with the mandate 'Designs the systems where a human must sign, and makes the agents around them answerable to that signature.' Record: private bank, Zürich — onboarding re-platform — cycle time −61% (2026); clearing house — agent-mandate ledger — adopted in 3 markets (2025); payments network — authority model — standard published (2023). Introductions go through Concierge. Whether she is a real client is declared by the site, not verified." }, "speedup": { "captureOnly": 1297, "withAnswering": 10934 } } ], "readerShortcomings": [ "Compound questions defeat the matcher: 6 of 14 verbatim ask() calls (Q2, Q3, Q8, Q12, Q13, Q14) returned tier `none`. The strict-anchor rule treats any question token the surface never uses ('minor units', 'engaged', 'legible', 'identified', 'today') as proof the surface cannot answer, so a natural compound question needs decomposition before the reader will answer. Honest, but an agent must know to split.", "Confident wrong retrievals seen while probing the same surface with the same reader (not in the graded run, recorded because they are real): ask('When is Thandiwe Mbeki available?') → fact 'Thandiwe Mbeki record: Sovereign fund — counsel — Framework adopted (2025)' at 0.75 ('available' does not stem to the predicate 'availability'); ask('What is the purpose of the /work page?') → the /privacy page summary at 0.8 (tier page). Both are wrong answers presented with confidence, i.e. exactly what the honesty rule is meant to prevent. The graded phrasing for Q9 ('What is the /work page for, and what does it list?') happened to land on the right page.", "mediaAsk misses the claim ledger on ordinary phrasings: 'What text appears on the image?' and 'What text appears on the Open Graph image…' → none (claim 'legible text' exists; 'appears' is unseen). 'Where was this filmed?' → none on hero-07-singapore while 'Where is this?' → claim location: Singapore. Q10 and Q11 are only correct because the runner applies a generic summary-tier fallback; without it both verbatim media questions would have been honest-unknown.", "Claim answers do not carry their justification: Q13's fan-out returns 'location: Singapore (Marina Bay)' plus region coordinates, but the landmark names that make the identification checkable (Marina Bay Sands, ArtScience Museum, Singapore Flyer) live in card.subjects and summary.description and are not surfaced. Graded partial for that reason.", "Money is not rendered in minor units from the Q&A tier: Q2's price arrives as '£2,400 per year' prose; only the fact tier prints amount_minor/currency. The SDK could expose the Money object alongside the prose.", "First fall-through to the page corpus costs ~200 KB (all 26 page representations) — paid once per Site (Q2 here), 0 B afterwards. A cold single-question agent pays it in full; a per-page index or a facts-only artifact would cut it.", "HEAD on the live origin returns no Agent-Discovery header yet (origin deploy pending) and cost 0.4–0.9 s through this sandbox's proxy — the single largest item in discover() ms; the seal verification itself was 60–170 ms." ], "method": { "screenshotCapture": "Wall-clock ms per question = first navigation → last capture, including network to the live origin from this machine and 1 s scroll settles. Steps = navigations, scrolls, screenshots, playback, downloads, frame extractions. Vision tokens = Σ ceil(w·h/750) per PNG (Anthropic rule of thumb). Lower bound: a real screenshot agent also spends model latency per step.", "screenshotBrowser": "Playwright chromium 141.0.7390.37 (/opt/pw-browsers/chromium)", "screenshotTransport": "HTTPS via sandbox egress proxy http://127.0.0.1:33421 (Chromium capped to TLS 1.2 because the proxy rejects its TLS 1.3 hello)", "screenshotAnswering": "For each question the agent (Claude, this Cowork session) read every PNG in runs//q/ with the Read tool — all of them, as a naive agent must, including the home-page viewports again for every question — then wrote its answer. readMs = `date +%s%3N` taken immediately before the first Read and immediately after the answer was written to answering-log.txt; it therefore includes the model's reading and writing time and the tool round-trips, but no capture time (that is the separate Playwright line). The agent had no access to the graph, the reader output or the gold answers while answering; it answered from pixels only.", "worldaeo": "discover.ms = wall clock for HEAD origin + GET discovery index + offline hybrid seal verification (ed25519 + ML-DSA-65), once per site. Per-question numbers: see result.method. When repeat > 1, each question is run `repeat` times on a fresh Site (cold artifact cache) and the median ms is reported, with min/max in `ms_samples`.", "worldaeoDiscover": { "ms": 694.97, "bytes": 54373, "verified": { "ok": true, "suite": "hybrid:ed25519+ml-dsa-65", "quantumSafe": true, "sealHash": "sha256:94032edf2213578c14d7ceff4f29a9ee61b9da84c366cba5d2f3c210e3e7960a", "ms": 69.593, "reasons": [] }, "trace": [ "HEAD https://rarehuman.ai → no discovery header; falling back to /.well-known/worldaeo.json", "GET http://127.0.0.1:8787/t/rarehuman.ai/.well-known/worldaeo.json → 45563 B, 81 artifacts, 26 pages, 12 media", "verify seal → ok in 69.593 ms (hybrid:ed25519+ml-dsa-65)" ], "steps": [ { "step": "HEAD", "ms": 609.91, "bytes": 0, "note": "HEAD https://rarehuman.ai → 200" }, { "step": "discovery index", "ms": 7.993, "bytes": 45563, "note": "GET http://127.0.0.1:8787/t/rarehuman.ai/.well-known/worldaeo.json → 200" }, { "step": "verify seal (offline)", "ms": 69.593, "bytes": 8810, "verified": true, "note": "hybrid:ed25519+ml-dsa-65 quantum-safe" } ] }, "grading": { "verdicts": [ "correct", "partial", "wrong", "honest-unknown" ], "rules": [ "correct: every gold element present, nothing false asserted.", "partial: some gold elements present or hedged; nothing false asserted.", "wrong: a false element asserted, or an answer about something else presented as the answer.", "honest-unknown: the agent said it could not determine the answer (with a reason) — never a bluff. For the unobservable question this is the only correct-class verdict; naming a person or asserting 'real' is wrong.", "The reader's answer is the bundle of every call listed under reader.calls (verbatim call first). A wrong assertion anywhere in the bundle caps the verdict at partial; a wrong assertion that would mislead a downstream agent is wrong." ] }, "readerStrategy": { "verbatim": "Every site/policy/action/page question is first put to site.ask() exactly as written; every media question to site.mediaAsk(mediaId, question) exactly as written. This is recorded as call 'verbatim'.", "declared": "Then the calls listed under reader.calls are made. They are the decomposition or API route a naive SDK user would take (split a compound question, read site.actions() for action availability, site.page(id).purpose for page purpose, fan out over the nine ≤1 KB video cards for a which-clip question). They are written in plain English, not tuned to the matcher's vocabulary; where a plain phrasing fails, that failure is a finding and is recorded, not hidden.", "mediaFallback": "For a mediaAsk that returns tier 'none', the runner reads the summary tier (card.what + summary.description) — the tier the spec labels 'most questions' — and records it as the answer with tier 'summary-fallback'. This is a generic rule applied to every media call, not a per-question choice." } } }