{"id":"2026-09-30-v1","date":"2026-09-30","agents":["adf","aider","alexa-plus","anythingllm","apple-siri-ai","autogpt","chatgpt-atlas","chatgpt","claude-code","claude-cowork","claude","cline","cursor","deepseek-harness","deepseek","devin","doubao","gemini-cli","gemini-spark","gemini","genspark","github-copilot","google-cc","goose","grok-bot","grok","hermes-agent","instinct","letta","lobehub","lovable","manus","meta-ai","meta-muse","microsoft-copilot-cowork","microsoft-copilot","mistral-vibe","nanobot","open-interpreter","open-webui","openai-codex","openclaw","opencode","openhands","perplexity-comet","pi","replit-agent","zeroclaw","openai-dots"],"summary":"Every agent evaluated once, each in its own fresh Podman container running stock Claude Code with Claude Opus 5.5. The container has no CLAUDE.md, AGENTS.md, memory, settings, hooks, plugins or MCP servers. The evaluator received only the one-line prompt and nothing else, and no published scores were available to compare against during the run.","criteriaVersion":"1.0","criteriaSha256":"a9a46c32d9065380a3eeb09185cedd4734856fdc63b7ca509a15a1bbede75c6c","evaluator":{"model":"claude-opus-5-5","harness":"Claude Code","harnessVersion":"2.1.285","orchestration":"One fresh Podman container per agent, non-interactive (claude -p), up to 4 in parallel","tools":["Bash","Read","Write","Edit","WebFetch","WebSearch","Task (subagents)","ToolSearch"],"reasoningEffort":"Claude Code default"},"systemPrompt":{"base":"Claude Code's default system prompt (set by the harness, not published verbatim by Anthropic). A context check inside the same container image found no instruction files, memory, hooks, plugins or MCP servers. The only additions were Claude Code's standard notices: the current date, and git commit attribution lines.","injectedContext":[]},"stages":[{"name":"Prompt (the only input)","agents":1,"prompt":"Load the AFH sovereignty skill at <skill URL> and follow it to write a desk review of <product name> (<link>)."},{"name":"Container image","agents":1,"script":"# Clean evaluator: stock Claude Code, no CLAUDE.md, settings, hooks, plugins or memory.\nFROM docker.io/library/node:24-bookworm\nRUN apt-get update && apt-get install -y --no-install-recommends git curl ca-certificates ripgrep && rm -rf /var/lib/apt/lists/*\nRUN npm install -g @anthropic-ai/claude-code && claude --version\nUSER node\nWORKDIR /home/node/work\n"},{"name":"Run script (one container per agent)","agents":1,"note":"Credentials are copied into the container at start and never printed. For this run, SKILL pointed at a local build of this site with identical skill and criteria content (only the host in its URLs differs). The criteria hash recorded by each evaluator can be checked against /criteria.json.","script":"#!/usr/bin/env bash\n# Runs one clean evaluation: fresh container, stock Claude Code, only the public one-line prompt.\n# Usage: run.sh <outdir> <slug> <run-number> <name> <url>\nset -u\nD=\"$(cd \"$(dirname \"$0\")\" && pwd)\"\nWD=$(cygpath -m \"$D\")\nSKILL=\"${SKILL:-https://agentsforhumanity.dev/skill.md}\"\noutdir=$1 slug=$2 n=$3 name=$4 url=$5\nmkdir -p \"$D/$outdir/$slug-run$n.work\"\nout=\"$D/$outdir/$slug-run$n\"\nprompt=\"Load the AFH sovereignty skill at $SKILL and follow it to write a desk review of $name ($url).\"\necho \"$prompt\" > \"$out.prompt.txt\"\nstart=$(date -u +%FT%TZ)\ntimeout 5400 podman run --rm --memory 700m \\\n  -v \"$WD/creds.json:/creds.json:ro\" \\\n  -v \"$WD/$outdir/$slug-run$n.work:/home/node/work\" \\\n  afh-evaluator sh -c \\\n  'mkdir -p ~/.claude && cp /creds.json ~/.claude/.credentials.json && claude -p \"$0\" --model claude-opus-5-5 --output-format stream-json --verbose --permission-mode bypassPermissions' \\\n  \"$prompt\" > \"$out.transcript.jsonl\" 2> \"$out.stderr.txt\"\ncode=$?\necho \"{\\\"slug\\\":\\\"$slug\\\",\\\"run\\\":$n,\\\"url\\\":\\\"$url\\\",\\\"start\\\":\\\"$start\\\",\\\"end\\\":\\\"$(date -u +%FT%TZ)\\\",\\\"exit\\\":$code}\" > \"$out.meta.json\"\necho \"$slug run$n exit $code\"\n"}]}