Emulo v0.6.7: claims match the evidence

Fixed

• Public copy claimed a benefit that no published test shows. Found by a claims audit of every public surface against the published evidence. The README tagline said agents "act like they just met you. Emulo fixes that". PyPI, the MCP registry entry and the Claude marketplace described a profile "your agent loads before every task". The Codex plugin said "Your agents stop starting from zero". The ClawHub skill and the load_emulo_profile tool description, which every MCP client shows its model, said the profile makes the agent "act like them instead of starting cold". RUN_ME.md told the mining agent the profile makes an agent behave "like someone who already knows them". The evidence is narrower: an appended mined profile changed output against a cold baseline; in a pre-registered 60 run study (August 2026) it did not separate from a length-matched invented profile on any of 9 comparisons; in a September rerun with the July profile and a new invented profile it separated on 5 of 9 on Fable 5.1 and 1 of 9 on Opus 5, measuring whether output changed, not whether it got better. Each surface now says what Emulo does (mines your typed messages into a you.md with dated receipts and installs it where your agent can read it), and the README has a "What is tested so far" section with those results, linked to /placebo and /fable. The home page and /agent-profile say "can read" and "can only have an effect" where they said "reads" and "only helps".
• Host claims now match a host-by-host load test; Claude Code and Codex skills are not auto-loaded reliably and the docs now say how to load them. The test used Emulo 0.6.6 from PyPI, a canary profile and fresh headless host sessions on 2026-09-25. A project AGENTS.md (--target agents --repo .) reached the model in Codex in one end-to-end test, and it and --target opencode were loaded into OpenCode's instructions. The Claude Code you skill installs but was left out of what Claude Code listed to the model; typing /you loads it. The Codex skill is opened only when a request matches its description. The Claude Code native plugin installs and registers its five skills. For the MCP server, Claude Code connected and listed the tool; Codex called it and received the no-profile message. Cursor, Gemini, OpenClaw and Hermes Agent loading is not verified. The README badge said "works with claude, codex, cursor, copilot, openclaw, hermes", although Copilot CLI is a log source only; it is now two badges, one for log sources and one for install targets. The support matrix lists installing and loading separately for every host. RUN_ME.md, the README, docs/PROFILE-FLOW.md, the home page, /agent-profile, llms.txt and the bootstrap skill lead with --target agents --repo . for Codex and OpenCode and /you for Claude Code, and the MCP section no longer names Claude Desktop and Cursor as clients that load the profile. Where a route was verified once, the docs say "verified end to end in one test" rather than "always on".
• RUN_ME.md now asks before installing anything, and says AGENTS.md is usually shared. Found by the quality review of this release before it shipped: an earlier draft of this release told the mining agent to run --target agents --repo . first, which appends the whole profile, verbatim dated quotes from private sessions included, to AGENTS.md in whatever folder the agent is in, and listed three install commands without saying to pick only the ones the person uses. RUN_ME.md now tells the agent to ask which coding agents the person uses and run only those installs, to name the exact folder and get a yes before --target agents, and to say first that the file is usually committed and shared. The README, docs/PROFILE-FLOW.md, the home page, /agent-profile, llms.txt, the bootstrap skill and the ClawHub skill say the same wherever they recommend --target agents. The emulo verify message no longer says to cut the rule twice.
• The README's Reddit example read as a result. It is now labelled as one run each way with no control, trimmed to each reply's opening plan and title, without "and it worked" or the unsourced "from an account with no followers", with the post numbers stated as shown in the screenshot, and linked to the controlled tests.
• /bench presented one run per task with no placebo as the product's effect. It now opens with a note that it is an early July 2026 comparison, that the model without the profile never saw the rules it was scored against, and that /placebo and /fable supersede it. Its title and descriptions no longer lead with 12 against 1, and "the plain model breaks the rules you told it, and the profiled one mostly does not" and "so only the profile could move the result" are gone.
• emulo verify said a missing quote "is invented, and it makes the agent confidently wrong about the person it is describing". A text match cannot know either. It now says a missing quote may be invented, paraphrased, or from history that was not mined, and that nothing in the mined corpus backs the rule. The README, the home page and docs/PROFILE-FLOW.md say the same, and RUN_ME.md no longer tells the mining agent that a rule without a receipt makes the agent "confidently wrong".
• /scan described a test it did not link. It said a with and without test of a rules file "made the agent measurably worse" and linked the note about CI enforcement. The test is in /notes/do-your-agent-rules-do-anything, which found the same four defects either way, a run 3.2 times longer with the rules, and no measurable difference in defects found. The page now says that and links it.
• Numbers without a source. "About 3M tokens" has no source in this repository and was removed. "8 months" and "9 months" now read "a profile the author mined in early July 2026 (1,656 sessions)" wherever that profile is meant, attributed rather than presented as a measured result, and the September rerun names its own input, "a later July 2026 profile (dated 2026-07-30)"; the README card image comes from a different run of the miner and is now captioned as such; /bench, which ran on 16 July with a profile whose session count is not recorded, no longer names one; 1,843 sessions and 1.74M tokens stay on /placebo, where the August study corpus is meant. The video skill's "321 upvotes" had no link or date and was removed.
• The README said "There is nothing to buy" while the site offers the paid Profile Build. It now says the tool is free and that the site offers one optional paid editing service.
• The video skill shipped one person's private film pipeline. It pointed at files that are not in this repository (videos/_kits/, videos/_handoff/, docs/MASTER-PROMPT-STANDARD.md), quoted private feedback and film measurements, and told every user that one person adds their own captions. It still loads the video profile first and keeps the method, written so it works in any project.
• Stale docs. docs/OPENCLAW_HERMES.md still said Ditto, used ~/.claude/skills/ditto where --target claude now writes ~/.claude/skills/you/, linked the old repository, and said a copied profile means "day one can start warm". ROADMAP.md linked a docs/FEEDBACK.md that does not exist, listed four Codex skills instead of five, described the 160K cap as the starter default, and planned a leaderboard for results already published. docs/SOURCES.md and docs/PROFILE-FLOW.md said OpenClaw and Hermes Agent load the profile; they now say discovery was checked in July 2026 on earlier versions and not re-verified. The example profiles no longer say the agent "acts like you".

Verified

• 3 new Python tests and one extended, each of which failed before the change for the reason it names: the MCP tool description must not promise the agent will act like the user (new); the existing invented-quote test now also requires that emulo verify not call a missing quote invented or say it makes the agent wrong; and (new) the RUN_ME.md a real run writes must print --target agents --repo . first, name /you, and drop "already knows them" and "confidently wrong". The /you check was confirmed by deleting that note from the RUN_ME.md template: the test failed, and passed again once the note was restored. (new) The generated RUN_ME.md must ask which agents the person uses before the first install command, say to run only those, require a confirmed folder for --target agents, and carry the committed-and-shared warning; with the warning deleted from the template the test failed, and it passed again once restored. The same invented-quote test now also requires the output to say cut only once.
• The CLI run end to end on a fixture history in a clean home: the generated RUN_ME.md reads as above; emulo verify on a profile with one real and one invented quote exited 1 and printed the new message; emulo mcp answered tools/list with the new description and tools/call, with no activated profile, with no active Emulo profile; run emulo.
• /bench and the home page rendered in headless Chrome and were read at full size.
• Full suite on Python 3.11 with the [pro] extra: 571 tests, 5 skipped, 0 failures. On Python 3.13 without cryptography: 549 tests, 13 skipped, 0 failures. On Windows the symlink tests skip, and the two release pin tests skip until the v0.6.7 tag exists.
• Site tests after a clean npm ci: 46 tests, 0 failures.
• The v0.6.7 bootstrap runtime pins emulo.py to SHA-256 c18cbcf0f1b84dce (prefix); MINING_PROMPT.md is unchanged at ee22077c2cda3c1c (prefix). Full digests in .agents/skills/emulo/runtime.json.
• Not verified here: the host results above come from the separate host-by-host test and were not re-run for this release. Python 3.8, because no 3.8 interpreter is installed on the release machine, and pip install emulo==0.6.7, because the version is not published at the time of writing.