Nine channels, one SAGARIS cadence, a domain that lasts

A real editor, and there are two of them sitting over one model: a linear list and a node canvas. The canvas compiles to the same executable steps the list writes, so a sequence built in either one runs identically and neither view is a picture of the other. Any of the shipped cadences opens as a graph you can rewire, and a cadence drawn on the canvas is the same rows in the same table as one typed as a list. Neither editor is the one that merely draws.

A reservation taken at dispatch, not a setting. Before a sequence email goes out, a database function locks the workspace, the account and the UTC day, and returns false if any contact at that account already received one inside it. The loser defers a day rather than sending, so four colleagues on one buying committee get four separate days. Because the lock is held inside the transaction, two scheduler ticks that overlap cannot both win, which is the case a planner-level check quietly loses.

In the contact's local time, not yours. Email sends between 08:00 and 10:00, SMS between 09:00 and 12:00, and calls surface between 09:00 and 17:00 Monday to Friday, all resolved against the contact's own timezone before any hour is compared. The other six channels carry no time of day constraint, because a LinkedIn touch is performed by a person on Today, or by a connected agent inside the hours set on its LinkedIn account, and a wait step is nothing but time. Windows are configurable per workspace; the defaults are the ones above.

Every channel stops, not just the one they replied on. A reply pauses the whole enrollment, so the follow-up SMS does not arrive after the prospect already answered the email. Separately, a cross-channel window of 24 hours keeps two different channels from landing on the same person inside a day.

It is not allowed to, and that is enforced in code rather than asked for in a prompt: a grounding gate requires every specific claim about the recipient to trace back to a non-model source before the draft can send. The gate extracts every specific assertion about the recipient from generated copy, funding rounds, money, counts, percentages, multiples, named entities, partnerships, awards, dates and recent events, and requires each one to trace back to material that came from a non-model source: operator-supplied prompt context, real retrieved artefacts, or known CRM facts. The model's own research summary is deliberately not evidence, because letting a generation vouch for itself is circular. An empty corpus does not mean nothing to check, it means nothing is grounded, so every specific blocks. This exists because a generated email once told a real prospect about a funding round that never happened.

A warmup ramp that both sending planes read from one committed file. A mailbox is capped at 5 a day in its first week, then 10, 20, 30 and 40, with a hard ceiling of 50; a domain runs 15, 30, 60, 90 and 120 with a ceiling of 150. No week may be more than twice the one before it. On top of that, suppression is fail-closed, so a read error blocks the send rather than letting it through; every commercial send carries a physical address and RFC 8058 one-click unsubscribe headers; and a domain whose spam rate crosses 0.30 percent is throttled, with 1 percent pausing it outright.

A short checklist, and the product walks your team through it. The outbound gate ships closed, so a new workspace sends nothing until somebody opens it. A first send is then held until the domain, mailbox, warmup, deliverability, placement testing, region capacity and compliance checks all pass. Your team also personalizes at least three drafts and reviews five contacts. That is why a first campaign lands in inboxes rather than in spam.