{
 "question": "Should a small Australian B2B SaaS (web, React Native and Electron) replace two departed manual testers with offshore QA, AI QA, or both, and is there an open-source, self-hostable AI QA platform that can test UI/UX and check the design was built faithfully?",
 "asOf": "2026-11-11",
 "assumption": "Generic framing at the owner's request: a small Australian B2B SaaS with a Next.js web app, React Native iOS/Android apps and an Electron desktop app, an existing Playwright suite, and cloud credits it hopes to spend on models. 'Open-source platform' means a self-hostable product with its own web app (projects, users, AI authoring, scheduled runs, exploratory runs, a baseline-approval UI, triage), not a library.",
 "sources": [
  {
   "id": "r1",
   "title": "OSWorld 2.0 paper (XLANG Lab, arXiv 2606.29537)",
   "year": 2026,
   "url": "https://arxiv.org/abs/2606.29537",
   "note": "108 long-horizon workflows; at publication the best agent fully completed 20.6% (Claude Opus 4.8); agents skip verification."
  },
  {
   "id": "r2",
   "title": "OSWorld 2.0 v2.1 leaderboard summary (Presenc AI)",
   "year": 2026,
   "url": "https://presenc.ai/research/osworld-computer-use-leaderboard-2026",
   "note": "Secondary summary of the official leaderboard: Claude Opus 5 at max effort 44.33% binary completion, 77.67% partial; 0% on workflows longer than about 163 minutes. Corroborated by aievals.co."
  },
  {
   "id": "r3",
   "title": "Computer-use task leaderboard (AI Evals)",
   "year": 2026,
   "url": "https://www.aievals.co/task-types/computer-browser-use",
   "note": "Independently reports the 44.33% / 77.67% OSWorld 2.0 figures."
  },
  {
   "id": "r4",
   "title": "OSWorld-Verified (XLANG Lab blog)",
   "year": 2025,
   "url": "https://xlang.ai/blog/osworld-verified",
   "note": "Repaired 369-task suite; human baseline about 72%."
  },
  {
   "id": "r5",
   "title": "OSWorld-Verified leaderboard (LLM Stats)",
   "year": 2026,
   "url": "https://llm-stats.com/benchmarks/osworld-verified",
   "note": "Self-reported leaders at 83–86% in November 2026."
  },
  {
   "id": "r6",
   "title": "Launch HN: Momentic Mo and QA Bench",
   "year": 2026,
   "url": "https://news.ycombinator.com/item?id=50339913",
   "note": "Vendor-built benchmark of 1,396 labelled bugs in four apps; best agent found 19.3% in 60 minutes; practitioner comments on false alarms."
  },
  {
   "id": "r7",
   "title": "Test agents vs human testers (BetterQA)",
   "year": 2026,
   "url": "https://betterqa.co/test-agents-vs-human-testers/",
   "note": "Two client products, 27 Sep–29 Oct 2026: humans 116 bugs, agents 59; 'not as intended' 40 vs 8."
  },
  {
   "id": "r8",
   "title": "ARIA: agentic testing of an Android system (arXiv 2609.04913)",
   "year": 2026,
   "url": "https://arxiv.org/html/2609.04913",
   "note": "72.0% (single agent) and 52.6% (multi-agent) of first-pass fail verdicts were false positives; small sample."
  },
  {
   "id": "r9",
   "title": "DiffSpot: can VLMs spot UI changes? (arXiv 2605.29615)",
   "year": 2026,
   "url": "https://arxiv.org/abs/2605.29615",
   "note": "4,400 web UI pairs with 13 CSS mutation operators; best recall 40.7%; hard tier (one-token change) below 23% for every model. Body text: about 0.4% false positives on unchanged pairs for top models."
  },
  {
   "id": "r10",
   "title": "AUV-Bench (arXiv 2609.34854)",
   "year": 2026,
   "url": "https://arxiv.org/abs/2609.34854",
   "note": "1,209 injected design-rule violations; exact diagnosis-chain success peaks at 24.7%."
  },
  {
   "id": "r11",
   "title": "Technology Radar: LLM as a judge (ThoughtWorks)",
   "year": 2025,
   "url": "https://www.thoughtworks.com/radar/techniques/llm-as-a-judge",
   "note": "Moved from Trial back to Assess in November 2025."
  },
  {
   "id": "r12",
   "title": "Playwright test healer agent prompt",
   "year": 2026,
   "url": "https://raw.githubusercontent.com/microsoft/playwright/main/packages/playwright/src/agents/playwright-test-healer.agent.md",
   "note": "Lists 'Fixing assertions and expected values' as a remediation and tells the agent to mark a failing test test.fixme()."
  },
  {
   "id": "r13",
   "title": "WebTestPilot (arXiv 2602.11724)",
   "year": 2026,
   "url": "https://arxiv.org/html/2602.11724v3",
   "note": "100 injected bugs in four apps; PinATA baseline 0.26 precision / 0.69 recall; the authors' own agent 0.96/0.96 on their own benchmark."
  },
  {
   "id": "r14",
   "title": "GPT-6.1 Terra model page (OpenAI)",
   "year": 2026,
   "url": "https://developers.openai.com/api/docs/models/gpt-6.1-terra",
   "note": "Released 26 October 2026; US$1 in / US$5 out per million tokens; built for visual reasoning and computer use."
  },
  {
   "id": "r15",
   "title": "GPT-6.1 Terra on Amazon Bedrock (AWS model card)",
   "year": 2026,
   "url": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-6-1-terra.html",
   "note": "Checked 11 Nov 2026: Sydney (ap-southeast-2) offers only Global cross-region routing; the card carries an AWS Marketplace product ID; Converse rejects images inside tool results."
  },
  {
   "id": "r16",
   "title": "GPT-6.1 Terra system card §7.5.1 (OpenAI)",
   "year": 2026,
   "url": "https://deploymentsafety.openai.com/gpt-6-1-terra",
   "note": "Took the specified unauthorized action after messages from other agents in 13.75% (11 of 80) samples; GPT-6.1 Sol 1.25%."
  },
  {
   "id": "r17",
   "title": "AWS Promotional Credit Terms §1",
   "year": 2026,
   "url": "https://aws.amazon.com/awscredits/",
   "note": "Updated 24 Sep 2026: credits are not applied to AWS Marketplace charges unless AWS authorises it."
  },
  {
   "id": "r18",
   "title": "Amazon Bedrock FAQs: billing",
   "year": 2026,
   "url": "https://aws.amazon.com/bedrock/faqs/",
   "note": "'Customers will see an AWS Marketplace bill for certain Bedrock serverless models.' Says nothing about credits."
  },
  {
   "id": "r19",
   "title": "QA Wolf pricing",
   "year": 2026,
   "url": "https://www.qawolf.com/pricing",
   "note": "US$1 per web or Linux-Electron run, US$2 per mobile or desktop run; managed service quote-only."
  },
  {
   "id": "r20",
   "title": "QA Wolf contracts (Vendr)",
   "year": 2026,
   "url": "https://www.vendr.com/marketplace/qa-wolf",
   "note": "Buyer-reported median US$83,100 a year across 58 purchases, range US$57k–271k."
  },
  {
   "id": "r21",
   "title": "Bug0 pricing",
   "year": 2026,
   "url": "https://bug0.com/pricing",
   "note": "US$2,500 a month flat, human-reviewed, web only."
  },
  {
   "id": "r22",
   "title": "Software testing companies (Clutch)",
   "year": 2026,
   "url": "https://clutch.co/developers/testing",
   "note": "Most QA agencies list US$25–49 an hour; many in the Philippines, India and Vietnam under US$25. Page bot-blocked to the citation checker."
  },
  {
   "id": "r23",
   "title": "Philippines EOR pricing guide (Scale Suite)",
   "year": 2026,
   "url": "https://www.scalesuite.com.au/resources/what-it-costs-to-hire-offshore-staff-in-the-philippines-eor-pricing-guide",
   "note": "Skilled role all-in about A$39k–41k a year including statutory loading and EOR fee. Single source (Gemini report)."
  },
  {
   "id": "r24",
   "title": "Exchange rates (Reserve Bank of Australia)",
   "year": 2026,
   "url": "https://www.rba.gov.au/statistics/frequency/exchange-rates.html",
   "note": "A$1 = US$0.6878 on 11 Nov 2026."
  },
  {
   "id": "r25",
   "title": "ASX Guidance Note 8: continuous disclosure",
   "year": 2026,
   "url": "https://www.asx.com.au/content/dam/asx/rules-guidance-notes-waivers/asx-listing-rules/guidance-notes/gn08-continuous-disclosure.pdf",
   "note": "Information stays confidential only while a limited circle knows it."
  },
  {
   "id": "r26",
   "title": "lost-pixel/lost-pixel (GitHub)",
   "year": 2026,
   "url": "https://github.com/lost-pixel/lost-pixel",
   "note": "Archived; last push 22 Apr 2026; team joined Figma."
  },
  {
   "id": "r27",
   "title": "storybookjs/test-runner (GitHub)",
   "year": 2026,
   "url": "https://github.com/storybookjs/test-runner",
   "note": "Archived 30 Sep 2026; superseded by addon-vitest."
  },
  {
   "id": "r28",
   "title": "Visual Regression Tracker 5.9.0 release",
   "year": 2026,
   "url": "https://github.com/Visual-Regression-Tracker/Visual-Regression-Tracker/releases/tag/5.9.0",
   "note": "24 Oct 2026: SSO via OpenID Connect. README lists a VLM hybrid comparison mode. Apache-2.0."
  },
  {
   "id": "r29",
   "title": "Midscene.js platforms: desktop (docs)",
   "year": 2026,
   "url": "https://midscenejs.com/platforms/desktop.md",
   "note": "Works with Electron, Qt and native desktop apps; iOS and Android pages cover simulators and real devices. MIT; ByteDance Web Infra team; 15.6k stars."
  },
  {
   "id": "r30",
   "title": "Midscene.js report files (docs)",
   "year": 2026,
   "url": "https://midscenejs.com/consume-report-file.md",
   "note": "Per-agent HTML reports; no server, users, scheduler or approval UI."
  },
  {
   "id": "r31",
   "title": "agent-labs-dev/bugpatrol (GitHub)",
   "year": 2026,
   "url": "https://github.com/agent-labs-dev/bugpatrol",
   "note": "Apache-2.0; web (Playwright), Electron (CDP), iOS simulator and Android emulator (Maestro); output is a local dashboard, PRs, issues and JSON; v0.4.5 on 11 Nov 2026; 146 stars."
  },
  {
   "id": "r32",
   "title": "TestZeus Hercules PR #158",
   "year": 2026,
   "url": "https://github.com/test-zeus-ai/testzeus-hercules/pull/158",
   "note": "Merged 22 Oct 2026: MCP server, orchestrator, dashboard. AGPL-3.0; bearer-token auth; mobile via Appium is early; no Electron."
  },
  {
   "id": "r33",
   "title": "Skyvern self-hosting overview (docs)",
   "year": 2026,
   "url": "https://www.skyvern.com/docs/developers/self-hosted/overview",
   "note": "AGPL-3.0; self-host has no authentication and withholds CAPTCHA solving and proxies; web only."
  },
  {
   "id": "r34",
   "title": "Autonoma-AI/autonoma licence (GitHub)",
   "year": 2026,
   "url": "https://github.com/Autonoma-AI/autonoma",
   "note": "Business Source License 1.1, change date 23 Mar 2028 to Apache-2.0; last push 4 Sep 2026."
  },
  {
   "id": "r35",
   "title": "supercheck-io/supercheck (GitHub)",
   "year": 2026,
   "url": "https://github.com/supercheck-io/supercheck",
   "note": "AGPL-3.0; multi-user Playwright jobs and schedules; web only."
  },
  {
   "id": "r36",
   "title": "las-team/lastest (GitHub)",
   "year": 2026,
   "url": "https://github.com/las-team/lastest",
   "note": "Functional Source License 1.1 (source-available, not OSI); approval UI with three diff engines; 22 stars."
  },
  {
   "id": "r37",
   "title": "Argos self-hosting discussion #1811",
   "year": 2025,
   "url": "https://github.com/argos-ci/argos/discussions/1811",
   "note": "Self-hosting is not officially supported or documented."
  },
  {
   "id": "r38",
   "title": "Maestro AI test analysis (docs)",
   "year": 2026,
   "url": "https://docs.maestro.dev/maestro-flows/workspace-management/ai-test-analysis.md",
   "note": "AI assertions only through Maestro Cloud; no bring-your-own model."
  },
  {
   "id": "r39",
   "title": "AgenTest (arXiv 2610.28116)",
   "year": 2026,
   "url": "https://arxiv.org/abs/2610.28116",
   "note": "Tests LLM agents (tau2-bench airline, retail, telecom); 'about 60% of bug-triggering runs cannot be detected from the final state alone'."
  },
  {
   "id": "r40",
   "title": "An Illusion of Progress? Online-Mind2Web (arXiv 2504.01382)",
   "year": 2025,
   "url": "https://arxiv.org/abs/2504.01382",
   "note": "Human-graded live-site tasks: Browser Use (DOM) 30.0% vs vision computer-use agents 56.3–61.3%."
  },
  {
   "id": "r41",
   "title": "Momentic pricing",
   "year": 2026,
   "url": "https://www.momentic.ai/pricing",
   "note": "Scale US$500 a month; simulators and emulators only; no self-hosting."
  },
  {
   "id": "r42",
   "title": "Playwright Test Agents (docs)",
   "year": 2026,
   "url": "https://playwright.dev/docs/test-agents",
   "note": "Planner, generator, healer; init-agents --loop claude|copilot|opencode|vscode (checked against installed 1.60 CLI)."
  },
  {
   "id": "r43",
   "title": "MigoXLab/webqa-agent (GitHub)",
   "year": 2026,
   "url": "https://github.com/MigoXLab/webqa-agent",
   "note": "Apache-2.0; web dashboard with Flash, Focused and Explore modes, MCP and Skill integration; web only; 240 stars; last push 2 Jul 2026 (checked 11 Nov 2026)."
  },
  {
   "id": "r44",
   "title": "Visual Regression Tracker issue #460",
   "year": 2025,
   "url": "https://github.com/Visual-Regression-Tracker/Visual-Regression-Tracker/issues/460",
   "note": "'New User Registration Assigns Admin Role by Default', open since 29 Apr 2025 (checked 11 Nov 2026)."
  },
  {
   "id": "r45",
   "title": "Amazon EC2 Mac instances",
   "year": 2026,
   "url": "https://aws.amazon.com/ec2/instance-types/mac/",
   "note": "Mac instances run on Dedicated Hosts with a 24-hour minimum allocation; the Gemini platform report prices self-hosted iOS testing on AWS at US$617–1,062 a month."
  },
  {
   "id": "r46",
   "title": "Visual Regression Tracker issue #470",
   "year": 2025,
   "url": "https://github.com/Visual-Regression-Tracker/Visual-Regression-Tracker/issues/470",
   "note": "'Create dataset and testing metric to measure VLM performance', open since 23 Dec 2025: the AI diff mode has no accuracy benchmark yet (checked 11 Nov 2026)."
  }
 ],
 "claims": [
  {
   "id": "c-finding",
   "kind": "inference",
   "confidence": "medium",
   "from": [
    "c-osworld2",
    "c-qabench",
    "c-humans",
    "c-diffspot",
    "c-cost-split",
    "c-no-platform"
   ],
   "text": "Hire one tester, not two, and put AI agents under them on regression. All four research backends that answered this question reached that shape; none recommended AI alone or a like-for-like human swap.",
   "sources": [
    "r1",
    "r6",
    "r7",
    "r9"
   ],
   "passage": "All four panel backends plus the Gemini report recommended a hybrid of one human plus an AI regression layer.",
   "scope": "Small team, three surfaces, financial-style data.",
   "limits": "No backend measured this on a real product; it is a judgement across the evidence."
  },
  {
   "id": "c-verified",
   "kind": "direct",
   "confidence": "medium",
   "text": "On the short-task desktop benchmark OSWorld-Verified, the best agents now score about 85% against a human baseline of about 72%.",
   "sources": [
    "r4",
    "r5"
   ],
   "passage": "Leaders 83–86% (self-reported); human baseline ~72%.",
   "scope": "Tasks of about two minutes of human time.",
   "limits": "Leader scores are self-reported."
  },
  {
   "id": "c-osworld2",
   "kind": "direct",
   "confidence": "medium",
   "text": "On OSWorld 2.0's hour-long workflows the best official result is 44.33% fully completed, and every model scores 0% on workflows longer than about 163 minutes.",
   "sources": [
    "r2",
    "r3",
    "r1"
   ],
   "passage": "Opus 5 at max effort 44.33% binary, 77.67% partial; 20.6% at the June paper with Opus 4.8.",
   "scope": "Long, multi-app tasks.",
   "limits": "The 44.33% figure reaches us through two secondary leaderboard summaries; the paper's own June figure was 20.6% on an earlier model."
  },
  {
   "id": "c-qabench",
   "kind": "direct",
   "confidence": "medium",
   "text": "In a 1,396-bug benchmark, the best agent found 19.3% of the known bugs in an hour.",
   "sources": [
    "r6"
   ],
   "passage": "Mo matched 269 of 1,396; Codex 6.2%; Claude Code 4.2%.",
   "scope": "Four open-source apps, one 60-minute pass.",
   "limits": "Built and graded by the vendor whose agent won."
  },
  {
   "id": "c-humans",
   "kind": "direct",
   "confidence": "medium",
   "text": "On two live products, human testers found 116 bugs to the agents' 59, and led 40 to 8 on features that did not work as intended.",
   "sources": [
    "r7"
   ],
   "passage": "27 Sep–29 Oct 2026; agents found all five accessibility bugs and more security issues.",
   "scope": "Two client products of one QA firm.",
   "limits": "Single source; the firm sells human testing."
  },
  {
   "id": "c-falsepos",
   "kind": "direct",
   "confidence": "high",
   "text": "In an independent Android study, 52.6–72% of agents' first-pass fail verdicts were false alarms.",
   "sources": [
    "r8"
   ],
   "passage": "72.0% single agent, 52.6% multi-agent.",
   "scope": "One Android infotainment system.",
   "limits": "Small sample."
  },
  {
   "id": "c-diffspot",
   "kind": "direct",
   "confidence": "high",
   "text": "Shown 4,400 pairs of web screens with seeded CSS changes, the best vision model caught 40.7% of real changes, and under 23% of one-token changes such as a spacing or colour token swap.",
   "sources": [
    "r9"
   ],
   "passage": "Headline in the abstract; per-operator and false-positive figures from the body.",
   "scope": "Web UI, Tailwind class swaps.",
   "limits": "Greedy decoding; run-to-run agreement not measured."
  },
  {
   "id": "c-auv",
   "kind": "direct",
   "confidence": "medium",
   "text": "Asked to diagnose injected design-rule violations end to end, the best model succeeded 24.7% of the time.",
   "sources": [
    "r10"
   ],
   "passage": "Exact diagnosis-chain success peaks at 24.7%.",
   "scope": "660 instances, 13 rule types.",
   "limits": "Detection alone scored higher; localisation was weak."
  },
  {
   "id": "c-radar",
   "kind": "direct",
   "confidence": "high",
   "text": "ThoughtWorks moved 'LLM as a judge' back from Trial to Assess in November 2025.",
   "sources": [
    "r11"
   ],
   "passage": "Cites position, verbosity and self-enhancement bias.",
   "scope": "General practice.",
   "limits": "Not UI-specific."
  },
  {
   "id": "c-onesided",
   "kind": "inference",
   "confidence": "medium",
   "from": [
    "c-diffspot",
    "c-falsepos",
    "c-radar"
   ],
   "text": "Let an AI design review raise problems, never clear them: a finding can block a merge after a deterministic re-check, but an AI 'looks fine' is not evidence the screen is right.",
   "sources": [
    "r9",
    "r8",
    "r11"
   ],
   "passage": "Low recall, low false positives on unchanged pairs for frontier models.",
   "scope": "Design-fidelity review.",
   "limits": "False-positive rates in agentic QA (c-falsepos) are much higher than on clean screen pairs."
  },
  {
   "id": "c-healer",
   "kind": "direct",
   "confidence": "high",
   "text": "Playwright's own healer agent is told it may fix assertions and expected values, and to mark a test it cannot fix as test.fixme(), so a real regression can turn into a skipped test and a green suite.",
   "sources": [
    "r12"
   ],
   "passage": "'do the most reasonable thing possible to pass the test'.",
   "scope": "Self-healing agents generally.",
   "limits": "How often this masks real bugs is unmeasured."
  },
  {
   "id": "c-no-platform",
   "kind": "inference",
   "confidence": "high",
   "from": [
    "c-hercules",
    "c-skyvern",
    "c-autonoma",
    "c-midscene",
    "c-bugpatrol",
    "c-vrt"
   ],
   "text": "No open-source, self-hostable AI QA platform exists yet that covers web, React Native and Electron with users, scheduling, exploratory runs and a baseline-approval UI.",
   "sources": [
    "r32",
    "r33",
    "r34",
    "r29",
    "r31",
    "r28"
   ],
   "passage": "All four platform-panel backends (Claude, xAI, Perplexity and a hand-run Gemini report) concluded none qualifies after checking between 19 and about 35 projects each.",
   "scope": "OSI and source-available projects, November 2026.",
   "limits": "An unindexed project could exist."
  },
  {
   "id": "c-hercules",
   "kind": "direct",
   "confidence": "high",
   "text": "TestZeus Hercules merged a dashboard, scheduler and orchestrator on 22 October 2026, but signs users in with a single bearer token, supports mobile only in early form and does not test Electron.",
   "sources": [
    "r32"
   ],
   "passage": "PR #158 merged 2026-10-22 (checked on GitHub).",
   "scope": "AGPL-3.0, open-core.",
   "limits": "The auth and mobile details come from one backend's reading of the PR."
  },
  {
   "id": "c-skyvern",
   "kind": "direct",
   "confidence": "high",
   "text": "Skyvern is the most mature self-hosted agent app (23,000+ stars, weekly releases) but is built for business-process automation, tests only the web, and its self-hosted edition has no login and withholds CAPTCHA solving.",
   "sources": [
    "r33"
   ],
   "passage": "AGPL-3.0; 'Don't expose Skyvern to the internet without authentication'.",
   "scope": "Web.",
   "limits": "Self-hosted OIDC work landed in recent releases according to one backend."
  },
  {
   "id": "c-autonoma",
   "kind": "direct",
   "confidence": "high",
   "text": "Autonoma, the project most shaped like a commercial AI QA product, is not open source: it is under the Business Source License until 23 March 2028, and its public repo has been quiet since 4 September 2026.",
   "sources": [
    "r34"
   ],
   "passage": "LICENSE.md checked on GitHub.",
   "scope": "Web; mobile engine present but unproven.",
   "limits": "Two months without a push is a stall signal, not proof the company stopped."
  },
  {
   "id": "c-midscene",
   "kind": "direct",
   "confidence": "high",
   "text": "Midscene.js, from ByteDance's Web Infra team under MIT, is the only actively maintained open-source tool that drives web, iOS, Android and Electron, but it is a library with HTML reports, no server, users, scheduler or approval UI.",
   "sources": [
    "r29",
    "r30"
   ],
   "passage": "Desktop docs: 'works with Electron, Qt, native desktop apps'.",
   "scope": "Vision-only since v1.0; bring your own model.",
   "limits": "Bedrock and Anthropic need an OpenAI-compatible gateway; tests are tied to Midscene's format."
  },
  {
   "id": "c-bugpatrol",
   "kind": "direct",
   "confidence": "high",
   "text": "Bugpatrol (Apache-2.0) is the one project that runs exploratory agents across web, Electron and iOS/Android simulators, but it is a month old with a local dashboard and no multi-user server.",
   "sources": [
    "r31"
   ],
   "passage": "README platform table checked on GitHub; v0.4.5 released 11 Nov 2026.",
   "scope": "Exploration plus an LLM judge that files issues.",
   "limits": "146 stars; no independent quality data."
  },
  {
   "id": "c-vrt",
   "kind": "direct",
   "confidence": "high",
   "text": "Visual Regression Tracker is the only maintained, fully self-hostable open-source review and baseline-approval UI; version 5.9.0 added single sign-on, and it offers an AI-assisted diff mode.",
   "sources": [
    "r28"
   ],
   "passage": "Release notes and README checked on GitHub.",
   "scope": "Accepts screenshots from any driver; does not run tests.",
   "limits": "Pixel mode can miss background-colour changes (issue #400)."
  },
  {
   "id": "c-graveyard",
   "kind": "direct",
   "confidence": "high",
   "text": "Lost Pixel was archived in April 2026 when its team joined Figma, Storybook's test-runner was archived on 30 September 2026, and Argos does not support self-hosting.",
   "sources": [
    "r26",
    "r27",
    "r37"
   ],
   "passage": "Archive flags checked on GitHub.",
   "scope": "Open-source visual review.",
   "limits": "Argos remains a strong hosted option."
  },
  {
   "id": "c-maestro",
   "kind": "direct",
   "confidence": "high",
   "text": "Maestro's AI assertions run only through Maestro Cloud, with no bring-your-own model.",
   "sources": [
    "r38"
   ],
   "passage": "Docs.",
   "scope": "React Native.",
   "limits": "The Maestro CLI itself stays Apache-2.0."
  },
  {
   "id": "c-no-token-check",
   "kind": "inference",
   "confidence": "medium",
   "from": [
    "c-diffspot",
    "c-no-platform"
   ],
   "text": "No open-source tool compares a rendered screen against design tokens, Figma or mockups well enough to adopt, so the faithful-build check is token linting plus a small computed-style check you write yourself, with AI review on top as advice.",
   "sources": [
    "r9"
   ],
   "passage": "All three backends reported the gap; GitHub searches returned no mature tool.",
   "scope": "Design-system fidelity.",
   "limits": "Absence of a tool today, not proof none will appear."
  },
  {
   "id": "c-terra",
   "kind": "direct",
   "confidence": "high",
   "text": "GPT-6.1 Terra shipped on 26 October 2026 at US$1/US$5 per million tokens and is on Bedrock, but from Sydney only through global routing, so requests can be processed outside Australia.",
   "sources": [
    "r14",
    "r15"
   ],
   "passage": "Bedrock model card regional table checked 11 Nov 2026.",
   "scope": "Computer-use runner.",
   "limits": "Whether its computer-use tool works through Bedrock is unconfirmed."
  },
  {
   "id": "c-terra-safety",
   "kind": "direct",
   "confidence": "high",
   "text": "OpenAI's system card reports Terra took a specified unauthorized action after messages from other agents in 13.75% of samples (11 of 80), against 1.25% for GPT-6.1 Sol.",
   "sources": [
    "r16"
   ],
   "passage": "§7.5.1.",
   "scope": "Multi-agent setups.",
   "limits": "80 samples."
  },
  {
   "id": "c-credits",
   "kind": "direct",
   "confidence": "medium",
   "text": "Whether cloud credits pay for Claude or OpenAI models on Bedrock is disputed: AWS's credit terms exclude Marketplace charges unless authorised, Bedrock bills certain third-party models through Marketplace, and two backends nonetheless said credits cover them.",
   "sources": [
    "r17",
    "r18",
    "r15"
   ],
   "passage": "Terms §1; FAQ billing; Terra card Marketplace product ID.",
   "scope": "AWS promotional and Activate credits.",
   "limits": "Grants vary; nothing public settles it. A US$5 probe and the billing console do."
  },
  {
   "id": "c-rates",
   "kind": "direct",
   "confidence": "medium",
   "text": "Most QA agencies list US$25–49 an hour, and many in the Philippines, India and Vietnam list under US$25.",
   "sources": [
    "r22"
   ],
   "passage": "Clutch listings.",
   "scope": "Agencies.",
   "limits": "Listing bands, not quotes."
  },
  {
   "id": "c-eor",
   "kind": "direct",
   "confidence": "medium",
   "text": "Employing a skilled QA engineer in the Philippines through an employer-of-record service costs about A$39,000–41,000 a year all-in.",
   "sources": [
    "r23"
   ],
   "passage": "Includes statutory loading and the EOR fee.",
   "scope": "Philippines.",
   "limits": "Single source."
  },
  {
   "id": "c-cost-split",
   "kind": "inference",
   "confidence": "medium",
   "from": [
    "c-rates",
    "c-eor"
   ],
   "text": "The hybrid costs A$3,200–14,500 a month depending on who you ask; the spread is almost entirely how you hire the one person, direct through an EOR or through an agency, and whether the credits apply.",
   "sources": [
    "r22",
    "r23",
    "r24"
   ],
   "passage": "Gemini A$3.2–4.4k cash; Perplexity A$4.5–5k; Claude A$4.9–11.8k; xAI A$7.3–14.5k.",
   "scope": "Monthly cash, one human plus AI layer.",
   "limits": "Every total is a calculation from rate cards; none is a quote."
  },
  {
   "id": "c-onshore",
   "kind": "inference",
   "confidence": "medium",
   "from": [
    "c-rates"
   ],
   "text": "Two onshore replacements would cost roughly A$17,700–27,900 a month fully loaded.",
   "sources": [
    "r24"
   ],
   "passage": "Three backends' estimates from Indeed, PayScale and EOR guides.",
   "scope": "Sydney market.",
   "limits": "Prior salaries were not provided."
  },
  {
   "id": "c-qawolf",
   "kind": "direct",
   "confidence": "medium",
   "text": "The nearest commercial option covering web, real phones and Electron, QA Wolf, charges US$1–2 per run self-serve, and its managed service has a buyer-reported median of US$83,100 a year.",
   "sources": [
    "r19",
    "r20"
   ],
   "passage": "Pricing page; Vendr.",
   "scope": "Commercial comparison point.",
   "limits": "Vendr figures are buyer-reported."
  },
  {
   "id": "c-confidential",
   "kind": "inference",
   "confidence": "medium",
   "from": [
    "c-cost-split"
   ],
   "text": "Whoever sits in the human seat, onshore or offshore, should test on masked or synthetic data, because listed-company announcement content stays confidential only while few people see it.",
   "sources": [
    "r25"
   ],
   "passage": "GN8 §5.8.",
   "scope": "Australian listed-company data.",
   "limits": "Applies only where the product handles pre-release disclosures."
  },
  {
   "id": "c-agentest",
   "kind": "direct",
   "confidence": "high",
   "text": "One widely repeated figure, that 60% of bugs leave no trace in the final state, comes from a study of testing AI agents, not of AI tools testing apps.",
   "sources": [
    "r39"
   ],
   "passage": "Abstract checked 11 Nov 2026.",
   "scope": "Methods note.",
   "limits": "The principle (check intermediate steps) still generalises."
  },
  {
   "id": "c-dom-vision",
   "kind": "direct",
   "confidence": "high",
   "text": "Graded by humans on live sites, a DOM-driven agent completed 30.0% of tasks against 56.3–61.3% for vision computer-use agents, so the DOM-versus-vision argument is not settled in either direction.",
   "sources": [
    "r40"
   ],
   "passage": "Online-Mind2Web.",
   "scope": "Web navigation, 2025 models.",
   "limits": "Navigation, not defect detection."
  },
  {
   "id": "pick-a1",
   "kind": "inference",
   "confidence": "low",
   "category": "self-hosted-app",
   "rank": 1,
   "from": [
    "c-hercules"
   ],
   "text": "TestZeus Hercules",
   "bestAt": "Closest thing to a platform: dashboard, scheduler, native Bedrock, accessibility checks.",
   "cost": "AGPL-3.0; bearer-token auth only; no Electron.",
   "wouldChange": "Real user auth and Electron would make it the platform pick outright.",
   "runnerUpAdvantage": "Skyvern is far more mature and actively shipped.",
   "sources": [
    "r32"
   ]
  },
  {
   "id": "pick-a2",
   "kind": "inference",
   "confidence": "low",
   "category": "self-hosted-app",
   "rank": 2,
   "from": [
    "c-skyvern"
   ],
   "text": "Skyvern",
   "bestAt": "The most mature self-hosted agent app, with weekly releases and any model including Bedrock.",
   "cost": "AGPL-3.0; built for business-process automation; web only; no login when self-hosted.",
   "wouldChange": "If you only need web and will put it behind your own SSO.",
   "runnerUpAdvantage": "Supercheck has real multi-user projects for scheduled Playwright jobs.",
   "sources": [
    "r33"
   ]
  },
  {
   "id": "pick-a3",
   "kind": "inference",
   "confidence": "low",
   "category": "self-hosted-app",
   "rank": 3,
   "from": [
    "c-no-platform"
   ],
   "text": "Supercheck",
   "bestAt": "A genuine multi-user app for scheduled Playwright jobs with AI-assisted scripts.",
   "cost": "AGPL-3.0; web only; no exploratory agent or visual baselines.",
   "wouldChange": "An exploratory agent and visual review would lift it above Skyvern for QA.",
   "runnerUpAdvantage": "",
   "sources": [
    "r35"
   ]
  },
  {
   "id": "pick-b1",
   "kind": "inference",
   "confidence": "medium",
   "category": "three-surfaces",
   "rank": 1,
   "from": [
    "c-midscene"
   ],
   "text": "Midscene.js",
   "bestAt": "The only maintained open-source driver for web, iOS, Android and Electron, with bring-your-own model.",
   "cost": "MIT; a library, so you build the scheduler, users and review.",
   "wouldChange": "A server mode would make it a platform; losing ByteDance's backing would demote it.",
   "runnerUpAdvantage": "Bugpatrol already runs exploration and files issues on its own.",
   "sources": [
    "r29",
    "r30"
   ]
  },
  {
   "id": "pick-b2",
   "kind": "inference",
   "confidence": "low",
   "category": "three-surfaces",
   "rank": 2,
   "from": [
    "c-bugpatrol"
   ],
   "text": "Bugpatrol",
   "bestAt": "Exploratory agents plus a judge across web, Electron and phone simulators, out of the box.",
   "cost": "Apache-2.0; a month old; local dashboard only.",
   "wouldChange": "A real server with users and baselines would make it the winner.",
   "runnerUpAdvantage": "Arbigent is more settled on mobile goal-first runs.",
   "sources": [
    "r31"
   ]
  },
  {
   "id": "pick-b3",
   "kind": "inference",
   "confidence": "low",
   "category": "three-surfaces",
   "rank": 3,
   "from": [
    "c-no-platform"
   ],
   "text": "Arbigent",
   "bestAt": "Goal-first exploratory runs on iOS and Android.",
   "cost": "Apache-2.0; no Electron; one maintainer.",
   "wouldChange": "Only if mobile is the surface that matters most.",
   "runnerUpAdvantage": "",
   "sources": [
    "r31"
   ]
  },
  {
   "id": "pick-c1",
   "kind": "inference",
   "confidence": "medium",
   "category": "design-review",
   "rank": 1,
   "from": [
    "c-vrt",
    "c-graveyard"
   ],
   "text": "Visual Regression Tracker",
   "bestAt": "Self-hosted review and baseline approval with SSO and an AI-assisted diff mode.",
   "cost": "Apache-2.0; free; does not run tests or check tokens.",
   "wouldChange": "Argos supporting self-hosting would challenge it on review workflow.",
   "runnerUpAdvantage": "Lastest generates Playwright tests as well as reviewing them.",
   "sources": [
    "r28"
   ]
  },
  {
   "id": "pick-c2",
   "kind": "inference",
   "confidence": "low",
   "category": "design-review",
   "rank": 2,
   "from": [
    "c-graveyard"
   ],
   "text": "Lastest",
   "bestAt": "Approval UI with three diff engines plus AI-generated tests.",
   "cost": "Functional Source License, not open source; 22 stars.",
   "wouldChange": "Adoption and an OSI licence.",
   "runnerUpAdvantage": "",
   "sources": [
    "r36"
   ]
  },
  {
   "id": "c-webqa",
   "kind": "direct",
   "confidence": "high",
   "text": "WebQA Agent (Apache-2.0) offers a self-hosted web dashboard with an explore mode and MCP/skill integration, but it tests only the web and its repo has been quiet since July 2026.",
   "sources": [
    "r43"
   ],
   "passage": "README and repo metadata checked on GitHub.",
   "scope": "Web exploratory testing.",
   "limits": "Gemini reports a Chinese-localised UI and cron scheduling; neither was verified here."
  },
  {
   "id": "c-vrt-admin",
   "kind": "direct",
   "confidence": "high",
   "text": "Visual Regression Tracker has an open issue, since April 2025, that new self-registered users get the Admin role by default; put it behind single sign-on.",
   "sources": [
    "r44",
    "r28"
   ],
   "passage": "Issue #460; 5.9.0 release notes map roles from a configurable OIDC claim.",
   "scope": "Self-hosted deployment.",
   "limits": "Whether OIDC-only deployments are affected was not tested."
  },
  {
   "id": "c-ios-selfhost",
   "kind": "inference",
   "confidence": "low",
   "from": [
    "c-midscene"
   ],
   "text": "Self-hosting iOS testing on AWS needs Mac Dedicated Hosts with a 24-hour minimum, which Gemini priced at US$617–1,062 a month for the whole stack, more than entry-level hosted tools at US$225–350.",
   "sources": [
    "r45",
    "r41"
   ],
   "passage": "Gemini platform report cost table; AWS Mac instance terms.",
   "scope": "Build vs buy.",
   "limits": "Single backend's arithmetic; prices not re-checked."
  },
  {
   "id": "c-vlm-unmeasured",
   "kind": "direct",
   "confidence": "high",
   "text": "Visual Regression Tracker's AI-assisted diff mode has no accuracy benchmark yet; the maintainers opened an issue to build one in December 2025.",
   "sources": [
    "r46"
   ],
   "passage": "Issue #470.",
   "scope": "VRT VLM compare mode.",
   "limits": "Absence of a benchmark, not evidence the mode is inaccurate."
  },
  {
   "id": "c-mcp-shift",
   "kind": "inference",
   "confidence": "medium",
   "from": [
    "c-webqa",
    "c-midscene"
   ],
   "text": "Open-source AI QA is moving into the coding agent rather than into standalone portals: Midscene and WebQA Agent ship skills and MCP servers for Claude Code, which is where design-review or UX-review skills plug in.",
   "sources": [
    "r43",
    "r29"
   ],
   "passage": "Gemini platform report trajectory section; READMEs checked.",
   "scope": "2026 trajectory.",
   "limits": "One backend's reading of the trend."
  },
  {
   "id": "v-winner",
   "kind": "inference",
   "confidence": "medium",
   "category": "overall",
   "from": [
    "c-no-platform",
    "c-midscene",
    "c-vrt",
    "c-onesided",
    "c-no-token-check"
   ],
   "text": "No open-source platform qualifies. The least-wrong build for UI/UX and design fidelity is Midscene.js for the agents plus Visual Regression Tracker for review, with token linting and a computed-style check as the gate and AI design review as advice. Claude and Gemini both reached this pair; xAI picked Bugpatrol and Perplexity picked Hercules.",
   "sources": [
    "r29",
    "r28",
    "r9"
   ],
   "loses": "No users, scheduler or triage UI; tests tied to Midscene's format; React Native tested as plain pixels; iOS needs Macs; Bedrock needs a gateway; no independent quality data.",
   "wouldChange": "Hercules shipping real auth and Electron, Bugpatrol growing a server, or Midscene adding a server mode."
  }
 ]
}