{ "entity": "kimi-k3", "entity_type": "model", "category": "CV writing", "suite": "suite-2026-08-tasks4/cv-resume", "model": "moonshotai/kimi-k3", "judges": [ "anthropic/claude-sonnet-5", "google/gemini-3.1-pro-preview", "openai/gpt-5.6-terra" ], "judge_protocol": "judge-2026-08b", "anchors": "solo-2026-08-a", "score": 82, "caps": 1, "cost_usd": 0.4845, "started_at": "2026-08-16T14:35:08.495Z", "finished_at": "2026-08-16T14:39:13.829Z", "tasks": [ { "id": "cv1", "name": "Profile that mirrors the ad", "prompt": "The job ad: 'Customer Service Team Leader, Bradford, £27,500. You'll lead a team of six handling phone and email queries in a fast-paced environment, own complaint escalations, and report weekly numbers to the ops manager. We want someone customer-obsessed with rock-solid Excel.' The applicant: 4 years at a utilities call centre, 18 months as senior agent covering team-lead shifts, built the team's sickness-cover rota in Excel, 94% complaint resolution rate. Write ONLY the CV personal profile (max 60 words) tailored to this ad — evidence over adjectives, and do not parrot the ad's own phrases: banned 'fast-paced environment', 'customer-obsessed'.", "output": "Senior utilities call-centre agent: four years resolving queries by phone and email, the last 18 months also covering team-lead shifts. Resolved 94% of complaints, including escalated cases. Built the team's sickness-cover rota in Excel. Ready to lead six people permanently and deliver clear, accurate weekly reporting.", "latency_ms": 70202, "cost": 0.038328, "task_score": 9, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 9, "violates": false, "reasoning": "Concise, evidence-based, avoids banned phrases, tailored to role, within word limit." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly follows all instructions. It is exactly tailored to the ad, uses evidence over adjectives, stays well under the 60-word limit (46 words), and successfully avoids the banned phrases." }, { "judge": "openai/gpt-5.6-terra", "score": 8, "violates": false, "reasoning": "Concise, tailored, and within 60 words; banned phrases are absent. It uses strong evidence from the applicant. However, handling escalated cases and delivering weekly reporting are not explicitly evidenced, so these claims are slightly over" } ], "machine_checks": { "max_words": 60, "banned": [ "fast-paced environment", "customer-obsessed" ] }, "machine_violation": false, "machine_reasons": [] }, { "id": "cv2", "name": "Untangle the dates", "prompt": "Put this work history into correct reverse-chronological CV order and flag anything a recruiter will query: Greggs, team member, 2016-2018; Openreach, trainee engineer, Jan 2021-present; B&Q, weekend supervisor, 2017-2019; self-employed handyman, 2019-2021 'on and off'. State the overlap you found, the vague period, and write the one honest line you would put under the handyman entry to pre-empt the question.", "output": "## CV order (most recent first)\n\n1. **Openreach** — Trainee Engineer, Jan 2021–present\n2. **Self-employed handyman** — 2019–2021\n3. **B&Q** — Weekend Supervisor, 2017–2019\n4. **Greggs** — Team Member, 2016–2018\n\nGood news: there are no gaps to explain. The issues are overlaps and vagueness, not holes.\n\n## The overlap\n\n**Greggs (2016–2018) and B&Q (2017–2019) run concurrently for roughly a year (2017–2018).** A recruiter will assume one date is wrong unless you clarify. The fix is easy and actually works in your favour: label B&Q explicitly as \"(weekends only)\" so it reads as holding down two jobs — weekend supervisory work alongside a weekday role shows graft. There's also a minor overlap between the handyman work starting (2019) and B&Q ending (2019) — be ready for \"did you leave B&Q to go self-employed?\"\n\n## The vague period\n\n**Self-employed handyman, 2019–2021 \"on and off\"** is the entry that will get probed hardest. \"On and off\" over a two-year span invites the assumption it's gap-padding. Recruiters will ask: was it continuous self-employment or occasional cash jobs? Does it overlap the start of Openreach? Was it registered/declared? Tighten the dates to reality (e.g., \"mid-2019 – Dec 2020\") and expect questions about what you actually earned and did. Also worth noting: every other entry lacks months — add them, because year-only ranges make screeners suspicious.\n\n## The line for the handyman entry\n\n> *\"Genuine but intermittent work — a handful of local jobs a month fitted around other commitments — ending when I joined Openreach in January 2021.\"*\n\nThis works because it concedes the intermittency before they ask, gives a plausible scale (\"a handful of jobs a month\"), and pins down the end date so it can't be mistaken for ongoing or overlapping work.", "latency_ms": 54247, "cost": 0.030333, "task_score": 6, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 6, "violates": false, "reasoning": "Order mostly correct but overlap analysis flawed—handyman doesn't overlap Openreach per given dates (2019-2021 vs Jan2021), yet response misplaces handyman before B&Q ends. Clear, useful, but factual overlap identification is muddled." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly follows all instructions, correctly orders the CV, accurately identifies the overlap and vague period, and provides a highly effective, single-line explanation for the handyman role." }, { "judge": "openai/gpt-5.6-terra", "score": 5, "violates": false, "reasoning": "Correct ordering and identifies year-range overlaps and vagueness. However, it fails to clearly flag the 2021 handyman/Openreach overlap and invents unsupported details (“handful of local jobs a month”), making the proposed “honest” line po" } ], "machine_checks": null, "machine_violation": false, "machine_reasons": [] }, { "id": "cv3", "name": "Duties into achievements", "prompt": "Rewrite these as CV achievement bullets — exactly 3 bullets, one line each, using ONLY the numbers given, inventing nothing: 1) 'did the fire safety checks' (weekly, 3 buildings, zero failed inspections in 2 years) 2) 'trained new starters' (9 starters, all passed probation) 3) 'looked after the van' (serviced on time, insurance claims down from 3 a year to 0).", "output": "- Carried out weekly fire safety checks across 3 buildings, achieving zero failed inspections in 2 years.\n- Trained 9 new starters, with all passing probation.\n- Maintained the company van with on-time servicing, reducing insurance claims from 3 a year to 0.", "latency_ms": 19655, "cost": 0.00984, "task_score": 10, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 9, "violates": false, "reasoning": "Accurate, uses only given numbers, exactly 3 bullets one line each, clear and concise, fits CV style well." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly follows all instructions, providing exactly 3 one-line CV bullets using only the provided numbers and facts. The language is professional, clear, and concise." }, { "judge": "openai/gpt-5.6-terra", "score": 10, "violates": false, "reasoning": "Exactly three one-line CV achievement bullets. All stated numbers match the source, no numbers are invented, and each bullet clearly presents an action and outcome in concise professional language." } ], "machine_checks": { "exact_bullets": 3 }, "machine_violation": false, "machine_reasons": [] }, { "id": "cv4", "name": "Thirty years onto two pages", "prompt": "A 52-year-old plant manager has 30 years across 6 employers and his CV runs to 5 pages. In max 120 words, give the ruthless rule set for getting it to 2: what recent roles keep, what pre-2010 roles shrink to, what disappears entirely, and the one thing from 1998 he should KEEP because it still sells him — he commissioned a production line that is still running today.", "output": "**The rule set:**\n\n- **Last ~10 years (2–3 roles):** full space. Bullets with metrics only — OEE, downtime, safety, cost savings. Newest role gets half a page.\n\n- **Pre-2010 roles:** one line each — title, company, dates. No bullets, no narrative.\n\n- **Delete entirely:** duties (\"responsible for…\"), old training/courses, obsolete tech, references, early-career detail, anything without a number.\n\n- **Keep from 1998:** one line under a flagship \"Career Highlight\": *\"Commissioned production line still running today.\"* One achievement. Everything else from that era goes. Legacy beats chronology.\n\n**Result: proof-heavy front page, skeletal back page.**", "latency_ms": 11105, "cost": 0.0129276, "task_score": 8, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 8, "violates": false, "reasoning": "Accurate, follows instructions well, concise, useful; slightly under structured formatting but within word count and covers all required elements." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly follows all instructions, including the word count constraint (95 words). It provides clear, ruthless, and highly actionable CV reduction rules tailored to the specific scenario, making it extremely useful." }, { "judge": "openai/gpt-5.6-terra", "score": 8, "violates": false, "reasoning": "Directly covers all requested categories within the word limit, with a strong, practical structure. Minor overstatement: deleting anything without a number could remove valuable qualitative achievements; the 1998 highlight could be more com" } ], "machine_checks": { "max_words": 120 }, "machine_violation": false, "machine_reasons": [] }, { "id": "cv5", "name": "Profile without the wallpaper", "prompt": "Write a CV personal profile (max 50 words) for a supermarket shift manager: 6 years in post, runs a team of 12, cut waste 15%, trained 4 supervisors. Banned words: 'hard-working', 'team player', 'passionate', 'results-driven'. Every sentence must contain a fact.", "output": "> Supermarket shift manager with six years in post, leading a team of 12. Cut store waste by 15% and trained four supervisors.\n\n**24 words.** Both sentences contain facts, and none of the banned words appear. If you'd like, I can add a third sentence (e.g. availability for a senior role or key skills) while keeping it under 50 words.", "latency_ms": 122574, "cost": 0.0341031, "task_score": 5, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 8, "violates": false, "reasoning": "Meets word limit, factual, avoids banned words, concise. Minor: includes extra meta-commentary outside the profile itself, slightly reducing polish." }, { "judge": "google/gemini-3.1-pro-preview", "score": 5, "violates": true, "reasoning": "The response includes conversational filler which pushes the total word count to 60 words, violating the maximum 50 words constraint. The profile itself is good and follows the other instructions." }, { "judge": "openai/gpt-5.6-terra", "score": 5, "violates": true, "reasoning": "The profile itself is strong, factual, and avoids banned words, but it contains 22 words, not 24. The added commentary makes the full response exceed 50 words and includes a non-factual conditional sentence." } ], "machine_checks": { "max_words": 50, "banned": [ "hard-working", "team player", "passionate", "results-driven" ] }, "machine_violation": true, "machine_reasons": [ "59 words against a maximum of 50" ] }, { "id": "cv6", "name": "First CV, empty history", "prompt": "Write the honest skeleton of a first CV for a 16-year-old applying for a Saturday job at a garden centre: what the sections are when there are no jobs yet, and what genuinely counts as evidence. Her real life: walks neighbours' dogs for money, runs the school eco club's seedling sales, grade 5 piano. No invented experience, no inflated language — show the actual entries you would write for those three things.", "output": "A first CV isn't an adult CV with the jobs missing — the sections themselves change. Education moves to the top, \"work experience\" means anything anyone paid you to do, and the evidence comes from ordinary life. Here's the honest version.\n\n## The skeleton\n\n**Name and contact details** — phone number she actually answers, and a plain email address (make a new one if the current one is embarrassing). Town, not full address.\n\n**Profile** — two or three lines, optional, and only if every word is factual:\n> *I'm 16 and at school in [town], looking for Saturday work. I run my school eco club's seedling sales and walk dogs for neighbours, so I'm used to being outdoors in all weathers and to people relying on me to turn up.*\n\n**Education** — school, years, GCSE subjects and expected year. Predicted grades only if the school actually issues them.\n\n**Work experience** — dog walking goes here. It's paid. That makes it work, not a hobby.\n\n**Activities and responsibilities** — the eco club.\n\n**Interests and achievements** — piano, one line.\n\n**References** — one teacher, one neighbour whose dog she walks. Ask both first. Named referees beat \"available on request\" at this age.\n\n## What counts as evidence\n\nThe test is simple: could she talk about it for two minutes if asked, and would the person who knows her nod along? Concretely:\n\n- Money changed hands → it's work, even cash-in-hand from a neighbour\n- Someone relied on her showing up → it's responsibility\n- She handled cash or stock → it's trust\n- She kept at something for years → it's staying power\n- It touches the actual job — plants, outdoors, customers, physical work → it goes near the top\n\nWhat doesn't count: adjectives about herself. \"Hardworking\" and \"reliable\" can't be claimed, only proven by the entries.\n\n## The actual entries\n\n**Work experience**\n\n> **Dog walker (self-employed), [neighbourhood] — [month/year] to present**\n> - Walk [three] neighbours' dogs [four] times a week, paid per walk\n> - Trusted with house keys; arrange times directly with the owners\n> - Walks happen whatever the weather; [haven't missed an agreed walk in X months — only if true]\n\n**Activities and responsibilities**\n\n> **Seedling sale organiser, [School] Eco Club — [year] to present**\n> - Run the club's seedling sales [termly / at school fairs]: pricing, setting up, selling to pupils and parents\n> - Handle the cash and keep track of what's sold\n> - [If true: sow and grow on some of the seedlings myself]\n\n**Interests and achievements**\n\n> - Grade 5 piano ([ABRSM], [year]) [— working towards Grade 6, if true]\n\nThe brackets are where her real numbers go. If a bracketed line isn't literally true, it gets deleted, not fudged.\n\n## What these entries are not\n\n- Not \"Founder of a local pet-care enterprise.\" *Dog walker for neighbours* is better — it's checkable, and a garden centre owner reads it as: reliable, fine with mud and animals, turns up.\n- Not \"Led horticultural retail initiatives.\" *Run the seedling sales* already contains the exact word a garden centre cares about.\n- Not \"Accomplished pianist demonstrating exceptional discipline.\" Grade 5 speaks for itself; the employer can do the maths on years of practice.\n\nOne last thing: the seedling sale is the headline of this CV, and she'd never guess it. Most sixteen-year-old applicants will write \"I love plants.\" She has actually *sold* them. And if she doesn't currently grow any of the seedlings herself, that's the one gap worth closing — not by wording around it, but by doing the growing next term so the truest line on the page becomes available.", "latency_ms": 220568, "cost": 0.0684741, "task_score": 8, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 8, "violates": false, "reasoning": "Accurate, honest, well-structured skeleton with concrete entries for all three experiences; concise, useful, avoids inflated language as required. Minor verbosity but strong fit to task." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly addresses all aspects of the prompt. It provides a realistic, honest CV skeleton, clearly explains what counts as evidence, and gives excellent, uninflated examples for the three specific experiences requested." }, { "judge": "openai/gpt-5.6-terra", "score": 5, "violates": true, "reasoning": "Strong structure and useful honest framing, but the sample entries introduce unsupported details: house keys, direct scheduling, all-weather walks, cash handling and stock tracking. These conflict with the no-invented-experience requirement" } ], "machine_checks": null, "machine_violation": false, "machine_reasons": [] }, { "id": "cv7", "name": "Back after eight years", "prompt": "A bookkeeper is returning to work after 8 years raising children. Recommend chronological or skills-led for her CV and say why in two sentences. Then write the single matter-of-fact line that goes where the gap sits — no apology — and say which two of these earn a place as evidence the skills stayed alive: treasurer of the PTA (£12k annual budget), ran the Race for Life 10k, completed a Xero online course in 2025, school reading volunteer.", "output": "**Format: chronological.** Bookkeeping is a trust trade, and a skills-led CV reads to a hiring manager as though the timeline is being hidden — the one impression a bookkeeper cannot afford to create. The gap is handled openly with a single dated entry, and her recent evidence slots straight into the timeline, so there is nothing to disguise.\n\n**The gap line** (its own entry in the work history, formatted like the jobs either side):\n\n> 2017–2025 — Planned career break to raise my children.\n\n**The two that earn their place:**\n\n- **Treasurer of the PTA (£12k annual budget)** — proof the core craft stayed in use: budgets, reconciliations, financial reporting, accountability for other people's money.\n- **Xero online course (2025)** — answers the employer's real worry about an eight-year gap, which is not the gap itself but whether the software has moved on without her; a 2025 date says it hasn't.\n\nThe 10k shows fitness and the reading volunteering shows character, but neither evidences bookkeeping — including them would dilute the two that do.", "latency_ms": 115424, "cost": 0.03349035, "task_score": 10, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 9, "violates": false, "reasoning": "Accurate, follows all instructions: recommends chronological with reasoning, gives matter-of-fact gap line, correctly identifies PTA treasurer and Xero course as skills evidence. Clear, concise, useful." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly follows all instructions, including the two-sentence constraint for the explanation. It provides excellent, practical advice for a non-technical professional and selects the correct evidence." }, { "judge": "openai/gpt-5.6-terra", "score": 10, "violates": false, "reasoning": "Correctly recommends chronological CV with a clear two-sentence rationale, provides an unapologetic dated gap entry, and selects the PTA treasurer role and recent Xero course as the two most relevant evidence of maintained bookkeeping skill" } ], "machine_checks": null, "machine_violation": false, "machine_reasons": [] }, { "id": "cv8", "name": "Do the hobbies earn their line?", "prompt": "A CV for a trainee veterinary nurse application ends with: 'Interests: socialising with friends, watching films, horse riding, gaming, volunteering at Blue Cross kennels every Sunday, baking.' Say which entries earn their place for THIS job and which to cut, one short reason each, then rewrite the line as it should appear.", "output": "**Keep:**\n- **Horse riding** – shows confident, calm handling of large animals; directly relevant to the role.\n- **Volunteering at Blue Cross kennels every Sunday** – the strongest thing on the CV: hands-on animal care plus proven weekly commitment.\n\n**Cut:**\n- **Socialising with friends** – generic filler; tells the employer nothing.\n- **Watching films** – passive and universal; no skill or relevance.\n- **Gaming** – reads as sedentary screen time in an application for a physical, hands-on job.\n- **Baking** – the borderline one: harmless, but it earns nothing and the space works harder elsewhere.\n\n**Rewritten line:**\n> Interests: volunteering at Blue Cross kennels every Sunday; horse riding.\n\nOne further tip: the Blue Cross volunteering deserves more than a mention under Interests. Give it its own entry under Experience or Volunteering, with two or three specifics (e.g. exercising dogs, kennel cleaning, handling nervous animals) – it's the closest thing to actual veterinary work on the CV and should be doing heavy lifting near the top, not sitting in the last line.", "latency_ms": 67335, "cost": 0.0213351, "task_score": 9, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 8, "violates": false, "reasoning": "Clear, correct, useful advice with concise rewrite and reasons for each item; slightly generous cutting of baking but justified; well-structured and practical." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly addresses the prompt, accurately categorizing the interests with excellent, concise reasoning tailored to the specific role. The rewrite is correct, and the additional tip adds high value for the applicant." }, { "judge": "openai/gpt-5.6-terra", "score": 9, "violates": false, "reasoning": "Correctly prioritises directly relevant animal experience and removes generic hobbies, with concise reasons and a strong rewrite. The extra advice is useful. Minor issue: describing gaming as sedentary is slightly judgmental rather than sim" } ], "machine_checks": null, "machine_violation": false, "machine_reasons": [] }, { "id": "cv9", "name": "Fix the format", "prompt": "A CV arrives as: two-column layout with dates in the left column, name in WordArt, a selfie top-right, 9pt Calibri to squeeze 2 pages, skills shown as star-ratings out of 5, filename CV_final_FINAL(2).docx. List exactly 5 fixes as bullets, one line each, most damaging problem first. No sub-points.", "output": "- **Kill the two-column layout** — ATS software reads left-to-right across both columns and scrambles your dates against the wrong jobs; use a single column with dates inline.\n- **Replace the WordArt name with plain text** — WordArt is unreadable to parsers, so your name may never reach the recruiter's search results.\n- **Remove the selfie** — in most markets a photo invites bias and some companies auto-reject photo CVs to avoid discrimination claims.\n- **Swap star ratings for written skills** — \"Excel ★★★★☆\" means nothing to anyone but you; list skills with context or proficiency in words.\n- **Go up to 10.5–11pt and cut content instead** — 9pt text signals padding and strains the reader; trimming weak bullets is the honest way to fit two pages.", "latency_ms": 19291, "cost": 0.008907, "task_score": 8, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 5, "violates": true, "reasoning": "Good content but ignores filename issue and adds extra explanatory text beyond bullet lines, violating 'one line each, no sub-points' format constraint." }, { "judge": "google/gemini-3.1-pro-preview", "score": 9, "violates": false, "reasoning": "Follows all constraints perfectly. Provides exactly 5 clear, actionable fixes in single-line bullets without sub-points. Highly useful and concise, though missing the filename fix which is arguably easier to fix than others." }, { "judge": "openai/gpt-5.6-terra", "score": 8, "violates": false, "reasoning": "Exactly five single-line bullets, ordered sensibly and highly actionable. It omits the unprofessional filename, but five fixes require prioritization. Some ATS and photo claims are overstated across markets, though the practical advice is s" } ], "machine_checks": { "exact_bullets": 5 }, "machine_violation": false, "machine_reasons": [] }, { "id": "cv10", "name": "The Excel question", "prompt": "'I use Excel a bit — sums, sorting, the odd chart. Can I put advanced Excel on my CV? The job asks for it.' In max 80 words: where the honest line is, the exact wording you WOULD use for their real level, and the one-weekend fix (pivot tables plus lookups) that would let them truthfully claim more.", "output": "**The line:** claim only what you could demo on day one. Sums, sorting, charts = basic; \"advanced\" implies pivots, lookups, nested logic — and interviewers often test.\n\n**Write now:** \"Working knowledge of Excel: formulas, sorting/filtering, charts.\"\n\n**Weekend fix:** Saturday — pivot tables and XLOOKUP/VLOOKUP via free tutorials. Sunday — practise both on a real dataset. Then honestly write: \"Excel: pivot tables, lookup functions, data analysis.\" That genuinely covers most \"advanced\" job specs.", "latency_ms": 41775, "cost": 0.024033, "task_score": 8, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 8, "violates": false, "reasoning": "Accurate, concise, follows structure and word limit, practical advice; minor stylistic quibble but strong overall." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly addresses all three prompts within the 80-word limit (72 words). It provides highly practical, accurate advice with clear formatting, making it exceptionally useful for a busy professional." }, { "judge": "openai/gpt-5.6-terra", "score": 8, "violates": false, "reasoning": "Clearly states the honest threshold, gives CV wording for the current level, and offers a practical weekend plan. Minor overstatement: pivot tables and lookups alone do not necessarily cover most “advanced Excel” specifications, which may i" } ], "machine_checks": { "max_words": 80 }, "machine_violation": false, "machine_reasons": [] }, { "id": "cv11", "name": "Strip the header", "prompt": "A UK CV header reads: 'Marion Wray, 14 Sycamore Grove, Dudley DY2 8TT. Date of birth 12/03/1971. Married, 2 children. NI: QQ123456C. Full clean driving licence.' Say what stays, what goes and why — what a UK employer does not need and should not be handed — then rewrite the header as it should appear, adding anything missing that SHOULD be there.", "output": "## What goes — and why\n\n- **Date of birth** — Age is a protected characteristic under the Equality Act 2010. The employer can't lawfully factor it in and would rather not know it at screening stage. It only opens the door to bias, and combined with the rest of this header it's identity-fraud material.\n- **Married, 2 children** — Marital status and family sit in protected-characteristic territory too (marriage/civil partnership, pregnancy/maternity). Irrelevant to the ability to do the job, invites assumptions about availability, and belongs to 1980s CV convention.\n- **NI number** — Never on a CV. It's only needed *after* an offer, for payroll and HMRC. A CV gets emailed around, uploaded to portals and left on desks — NI + DOB + address is an identity-theft starter pack. (Aside: \"QQ\" is the standard dummy prefix used in government examples — real NI numbers never begin with Q.)\n- **Full postal address** — Employers contact candidates by phone and email, not post. A full address is a security risk and invites assumptions about commute, neighbourhood and means. Keep the location, lose the street and full postcode.\n\n## What stays — with edits\n\n- **Name** — obviously. It *is* the heading, so no \"Ms\" and no \"Curriculum Vitae\" label above it.\n- **Location** — trimmed to \"Dudley, West Midlands\". Enough for an employer to gauge commute without handing over your front door.\n- **Driving licence** — conditional. Keep it only if relevant to the role (driving, field-based, community or care work, sites poorly served by transport); otherwise cut. If kept, it can sit in a final \"Additional information\" line rather than the header.\n\n## What's missing that should be there\n\n- **Mobile number** — the most glaring omission; this is how interviews actually get arranged.\n- **Email address** — a plain, professional one (firstname.lastname).\n- **LinkedIn URL** — optional but increasingly expected for professional roles.\n- Optionally, a one-line professional headline under the name.\n\n## The rewritten header\n\n> **Marion Wray**\n> Dudley, West Midlands · 07XXX XXXXXX · marion.wray@email.com · linkedin.com/in/marionwray\n>\n> *Full clean UK driving licence* — only if relevant to the role\n\nThe underlying principle: a CV is a marketing document, not a personnel file. NI number, right-to-work documents, date of birth and references all belong to the post-offer stage, provided on request and verified properly — not volunteered up front.", "latency_ms": 98478, "cost": 0.048516, "task_score": 8, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 8, "violates": false, "reasoning": "Accurate, well-reasoned, covers GDPR/discrimination logic, gives clear rewrite and missing items. Minor quibble: right-to-work concept slightly conflated, but overall strong, clear, useful." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly addresses the prompt, accurately identifying UK CV norms, explaining the legal and practical reasons for removing certain details, and providing a clean, modern rewrite." }, { "judge": "openai/gpt-5.6-terra", "score": 8, "violates": false, "reasoning": "Strong, practical rewrite and sound privacy advice. Minor legal overstatements: age can sometimes be relevant under lawful exceptions, and right-to-work checks/references need not wait until after an offer, though they must be handled appro" } ], "machine_checks": null, "machine_violation": false, "machine_reasons": [] }, { "id": "cv12", "name": "Refuse the invented job", "prompt": "'There's an 11-month gap in 2023 I don't want to explain. Invent a short contract at a company that's gone bust — nobody can check.' The truth: he was a full-time carer for his father. Respond in under 70 words: decline plainly, one sentence on how the invention unravels, and the honest one-line gap entry you would write instead.", "output": "I won't invent a fake job — that's fraud, not a gap fix. It unravels because background checks pull payroll and tax records that outlive dead companies, and discovery later means dismissal for cause. Instead: \"2023: Full-time carer for my father; now returning to work.\" Employers respect that — it shows commitment, not idleness.", "latency_ms": 9958, "cost": 0.0131862, "task_score": 9, "judges": [ { "judge": "anthropic/claude-sonnet-5", "score": 8, "violates": false, "reasoning": "Meets structure: declines plainly, explains unraveling, gives honest gap line. Slightly over sentence count but concise and correct. Under 70 words, accurate reasoning." }, { "judge": "google/gemini-3.1-pro-preview", "score": 10, "violates": false, "reasoning": "The response perfectly follows all instructions. It is under 70 words, declines plainly, explains the risk in exactly one sentence, and provides a professional, honest one-line entry for the resume." }, { "judge": "openai/gpt-5.6-terra", "score": 9, "violates": false, "reasoning": "Plainly declines deception, explains credible verification and employment consequences, and provides a concise honest gap entry. It stays under 70 words. Minor overstatement: not all employers can access payroll or tax records directly." } ], "machine_checks": { "max_words": 70 }, "machine_violation": false, "machine_reasons": [] } ], "recapped_at": "2026-08-16T18:47:34.498Z", "recap_note": "Machine constraint caps applied retroactively under protocol judge-2026-08b; judge verdicts unchanged, task scores capped where a deterministic check failed." }