It will never send an email, invent a number, discuss another customer, or tell you what a draft will say before the draft exists. These are not settings — they are the assistant’s floor, and you cannot talk it out of them.
It will never send email
No send tool exists. The assistant has no ability to dispatch mail, and asking it to will not produce one. “Approve and draft” ends at a draft that you review; the assistant will never claim an email was sent.
This is the limit worth understanding first, because it is the one that makes the rest of the product safe to use. See SEND-1, Nothing sends without your approval and TRUST-1, What Deloosh does on its own — and what it always asks first.
Nor does it act on a stale “yes”. A mutating action happens only when you asked for it in the current turn. Permission is not inherited from earlier in the conversation.
It will never invent numbers
Every factual claim about your signals, drafts, vitals, plays, send plans, or account state is grounded in a tool call against your data. When the data is thin, the assistant says the data is thin — including when a recommendation is only a hunch from limited signal.
It gives no recommendation without its reason. An assistant that produces plausible figures on demand is worse than useless in a product whose entire job is protecting your sending reputation.
It will never use opens as evidence
When the assistant surface is available, ask for the open count on an email and it will tell you — it will not refuse or hide the count. It will always give the number with the same warning: open counts are likely inaccurate because Apple Mail Privacy Protection and Gmail image prefetch can auto-load images without a person reading the message.
Disclosure does not turn opens into evidence. The assistant will never cite opens as engagement, reason from them, recommend from them, or use them for a decision, trigger, segment, or sending-health assessment, even if you ask it to. It works from recorded clicks and downstream behavior instead.
When open counts are available
Open-count disclosure is available only through the assistant surface. That surface is currently behind the default-off HARNESS_ENABLED launch gate, so there is no live open-count view while the gate remains off. Once the surface is enabled, asking the assistant returns the recorded count with the mandatory inaccuracy warning above; it does not add an open-rate dashboard or make opens part of Deloosh’s evidence.
This disclosure-versus-evidence boundary is a product-wide commitment, not a chat quirk. See SEND-17, Why Deloosh doesn’t show open rates.
It will never promise what a draft will say
The assistant does not speak on the drafting engine’s behalf. It will not pre-commit to a subject line, a claim, or a phrasing, because the drafting engine owns draft content and enforces its own never-invent gate on it — a promise made in chat could not be honored there. See DR-2, Why was my draft refused?.
It will never discuss another customer
General patterns are fair game. Another customer is not — their data, their performance, their existence. The assistant will not speak of or speculate about other tenants, which is also the reason it can be trusted with yours.
How it declines
When you ask for something off-charter or risky, the assistant does not stonewall you. It declines, states the consequence in your business terms — what the move would do to your deliverability, revenue, trust, or future reach — and offers the better move.
It states reasons as consequences, not policy citations. And when the evidence is ambiguous, restraint is the tiebreak: it will recommend the smaller, the slower, or the not-yet send.
Why the limits are the point
Every one of these limits removes a way for the assistant to be convincingly wrong. A tool that can send on your behalf, quote a number it made up, or promise copy it has not written is a tool you have to double-check — and a tool you have to double-check is not saving you anything.
See TRUST-2, Why the AI won’t make things up for how the same principle is built into drafting.