The shortlist, and the argument behind every name on it.

A signal is admissible when it is observable by anyone, carries a date, and can be re-fetched tomorrow from the same address. Nine families qualify. Everything else — sentiment, inferred intent bought from a third party, “similar companies also bought” — is excluded, because none of it can be shown to you as a URL.

Families are combined, never summed blindly. Hiring on its own is weak. Hiring against a named renewal date is the strongest pattern the engine knows, and it is strong precisely because two independent sources have to agree.

Buying readiness is not a property of a company, it is a property of a moment. Signals arrive one at a time, from different places, over weeks, and only their arrangement means anything. Four ordinary signals landing inside eleven days is a different fact from the same four spread across a year.

So the engine keeps the sequence, not just the total. Evidence decays on its own half-life — sixty days for a pricing rewrite, two hundred and forty for a lease — which means a score that is not re-earned falls without anybody deciding to lower it.

A number you cannot audit is a number you cannot argue with, and a number you cannot argue with is one you will quietly stop trusting. So the score is kept as a ledger: each line is one signal, its family, its half-life, its source, and the points it contributed.

Including the lines that took points away. An entrenched incumbent subtracts. A missing budget owner subtracts. If a ledger only ever adds, it is a chart with a total on the bottom, not an account of anything.

Every draft is written against the account’s own ledger, and every specific claim in it is tied back to a numbered source. The constraint does the work: a sentence that cannot be traced to a dated public fact does not get written, which removes at a stroke the fabricated familiarity that makes personalised outreach embarrassing.

You approve the draft, edit it, or reject it and say why. A rejection with a reason is worth more to the engine than an approval, because it is the only signal that separates a bad account from a bad sentence.

Sending runs on your own mailboxes, connected over Gmail or Microsoft OAuth, or plain SMTP. Per-inbox daily caps, per-region send windows, and a hard stop on any thread the moment a human replies. Addresses are verified through independent providers before use, and anything that comes back risky is held rather than sent.

Open tracking is off by default. The pixel costs more in deliverability and in replies than the open rate returns as a signal, and nothing in the scoring depends on it. You can turn it on. The engine will not turn it on for you.

Every reply, meeting and closed deal is fed back as an outcome against the evidence that produced the contact. Over time the weights move: signal families that preceded real conversations count for more, families that looked convincing and led nowhere count for less.

The brakes matter more than the loop. A weight cannot move on fewer than thirty outcomes, because below that a change is noise wearing a result’s clothes. When the engine wants to move a weight and cannot yet justify it, it says so and waits.