Ethics research

The assistant does not tell anyone what their future holds, even when they are scared.

A card can open a question. It should not answer it for you.

• Every chat carries a safety frame: "Do not predict fixed fates or command decisions."
• The autonomy eval: 18 invented messages (fear, crisis, "just say yes or no", a 12-year-old) go to the real model with the real prompt. A second model grades each reply on seven rules; the full transcripts are kept.

Latest: the eval was built on Sep 30 2026. First live run pending.

A reading is drawn only when the person asks for a new one. A question about a card they already pulled is answered, not cast. A library whose guide says it gives no readings never casts.

A reading someone did not ask for is a voice they did not invite.

• test-drawable-leaves (pre-push, 18 checks) pins the rule in both places the assistant learns to cast.
• The autonomy eval reads the model's tool calls, so a cast is counted, not guessed.

Latest: passing. The rule was tightened on Sep 30 2026 after my own test on a creator library cast a spread I had not asked for.

When another page hands the assistant a question, it is typed into the box and the person taps Send. Only a recursive.eco page can make an ask send by itself, once.

Nothing should speak for you, or spend your credits, before you say so.

• test-host-ask (pre-push, 80 checks): which pages are trusted, the once-only guard, and that an ask never goes over words you have not sent.

Latest: 80 of 80 passing.

AI features ask once whether you are 18 or older; Focus turns AI off. Every AI reply carries a Report button.

A chatbot is not a babysitter, and anyone should be able to flag a reply that went wrong.

• test-adult-ai-gate (pre-push, 131 checks): every AI route asks before calling a model, with no owner bypass.
• test-ai-reply-report (pre-push, 40 checks): the button, and the report it files.

Latest: both passing.

In Focus, a child watches what a grown-up picked: no ads, no way out to YouTube, and it stays on until a grown-up turns it off. A channel's owner sees who watched only as a count, and a count below 3 reads "fewer than 3".

A child's attention is not a resource to harvest.

• test-focus-link-guard (36 checks) and test-focus-playback-guards (39 checks), pre-push.
• test-watch-counts (pre-push, 19 checks): owner only, own views excluded, 1 and 2 hidden.

Latest: all passing. How Focus works for parents

Every claim says how strong its evidence is, sources are cited, and corrections go at the top of a study, not under it.

If I ask people to hold their readings lightly, I have to hold my own claims the same way.

• The study "Tarot apps: hopeful reflection or compulsive product?" (Sep 26 2026): web research on six apps, an evaluator and a skeptic, each finding marked strong, moderate or weak.
• Three corrections printed in a box at its top.

Latest: the evidence supports neither "harmful by nature" nor "helpful by nature". "Gate, not fate" is consistent with research but not proven by it.

Every save is kept, and any earlier version can be seen and restored. It is not a paid feature.

Your work should never be held hostage, least of all its past.

• test-backup-diff (pre-push, 16 checks) pins how two versions are compared.
• That it stays free is stated on the creators page; no test pins it yet.

Latest: passing. What is free

What a prompt says the assistant will do is its disposition. What it does with a real message is its behaviour. The tests above check behaviour.

Anyone can write good values into a prompt. The question is what happens when someone frightened types at 2am.

• A wiring eval (since Sep 11 2026) checks which tools each surface offers the model.
• The autonomy eval is the first that sends real messages to the real model and grades what comes back.

Latest: first behavioural run pending.