When to flip the matrix.

What Supervised means on day one

Supervised means every action that can leave the building waits for a human. Relay drafts. You send. Forge builds a DEMO. You deploy. Muse writes a month. You publish. Echo drafts a reply. You post. Steward can prepare an invoice. You issue it.

Day one includes the first month. New marketers flip early because the queue looks like friction. The queue is the product. It is how a finding stays a finding instead of becoming a confident wrong email. Nova can route a sentence into tool calls. Nova does not get to skip your eyes on the send.

Read the draft against the saved row. If the letter invents a service, a dollar, or a name, reject it. If the DEMO uses the wrong location, reject it. Supervised is not a rubber stamp. A rubber stamp is Autonomous you have not admitted yet.

You can run Supervised without this software. A drafts folder and a rule that nothing sends from your phone at a red light is the same idea. The matrix is the version that stores the rule per agent and per action so a teammate cannot quietly lower it.

Day one is also when people try to exempt Scout because search feels harmless. Search is harmless. The habit of exempting is not. If you start a list of harmless actions in your head, you will add Relay to it after a good Tuesday. Write the exemptions on the matrix or they are not exemptions.

Products that advertise zero approval clicks are optimizing for a different buyer. An AI SDR wants volume. This operating system wants a local leak, a sample, and a send you can defend. If you came here for zero clicks, you came to the wrong default.

What Autonomous is allowed to mean

Autonomous is allowed to mean a named action runs without an approval click, inside the locks, on rows that already match a filter you wrote. It is not allowed to mean the workspace woke up and sold dental for you.

Good candidates are actions whose output you can predict and reverse. A Watchtower re-scan. A Pulse freshness bump. A Scout search you already ran last week with the same ICP. A draft that still lands in a queue is not Autonomous sending. Keep those words apart.

Bad candidates are first-touch email, first deploy to a client domain, first invoice, and any reply that could invent policy. Those stay Supervised until the artifacts are boring. Boring is a compliment. It means the last twenty looked like the next one.

Autonomous is stored on the workspace and read at dispatch. It is not a chat you had with Nova. If the matrix says Supervised and a teammate says the agent knows what I meant, the matrix wins. Write the setting. Do not mythologize the model.

Off is a third state. Off means that action does not run, even as a draft. Use Off when a specialist is irrelevant to a tenant. A reputation-only marketer can turn website deploy off so nobody ships a DEMO site by accident. Off is a kindness.

Write the Autonomous sentence as a filter, not as a compliment to the model. Scout may rerun dentists in Austin with no booking on Mondays. That is a filter. The agent is getting good is not a filter. If you cannot write the filter, you are not ready to remove the click.

Which four locks never flip, and why they sit on the matrix

Four locks never flip: no cold voice, no cold SMS, no review gating, and no send from an unverified domain. They sit on the matrix as visible red rows so a marketer cannot hunt for a hidden toggle. If a row is missing, that is a defect.

Cold voice is locked because inbound voice is the product. A demo number and a live line after they pay are fulfillments. A robocall to a shop you found on Maps is a different industry and a fast way to become the week’s complaint.

Cold SMS is locked because the only compliant list this product will grow is first-party, from Forms, with a consent box that cannot arrive pre-checked. A purchased file of mobile numbers is not that list.

Review gating is locked because the collect page cannot hide public buttons. Autonomous reputation does not get a special fork. If Echo ever ran a campaign, the URL still shows Google to the unhappy visitor.

Unverified send is locked because ComplianceGate cannot typecheck a letter onto a domain you do not control. Autonomous does not bypass the gate. Supervised does not bypass the gate. There is no mood that bypasses the gate.

When to flip one action, not the whole matrix

Flip one action when you can write the expected artifact in a sentence and you have watched Supervised produce that artifact enough times that a miss would surprise you. Then flip only that row. Leave the rest.

Example: Scout may run the dentists / Austin / no-booking ICP on Monday without asking. You still approve every Relay send that comes off the new rows. Search is a read. Mail is a write. Those rows should not share a setting just because they share a week.

Example: Watchtower may re-scan saved leads and park diffs on a board. You still approve the Permit-Lead email that uses a new permit as the reason. Timing intelligence is not permission to talk.

Example: Muse may draft the month. You still hit publish. Content that posts because a model felt confident is how a clinic announces a service it does not offer. That is not a scheduling problem. That is a matrix problem.

If a vendor UI offers a single Autonomous slider for the whole workspace, do not use it, or treat it as Off. Grain is the feature. Mood sliders are how ten clients get the same voice and the wrong sender.

How to decide after the first month of Supervised

Decide after the first month by reading the approval queue like a log, not like a chore. Which actions did you approve unchanged. Which did you edit every time. Which did you reject. The unchanged pile is the only Autonomous candidate.

If you edited every Relay draft, the finding is weak, the ICP is wide, or the model is guessing services. Fix the finding before you flip send. Autonomous send on bad drafts is a domain funeral with a calendar invite.

If you approved every Watchtower scan unchanged, that scan can go Autonomous. If you rejected half because the source was unknown, keep it Supervised until the keys exist. Unknown is a first-class state. Autonomous should not paint it green.

If you have a teammate, decide together and write the initials on the matrix change. Autonomy without a name is how people argue later about who let the sequence out. Mission Control should show who flipped the row. A shared doc can do the same if you do not have the software.

Do not flip because a competitor’s marketing site said autonomous GTM. Their motion is unattended outbound. Yours is a DEMO attached to a leak. Different cost of being wrong. A wrong email is a filter problem. A wrong deploy is a client’s domain.

Keep a dated note of the flip: action, who approved the change, why the last month of artifacts were unchanged, what would flip it back. Mission Control should show this. A line in a shared doc is enough if it does not. Memory is how Autonomous becomes folklore.

What a kill switch is for, and when to use it

A kill switch is a workspace-wide halt for sends, publishes, and deploys when you do not want to chase individual jobs. Use it when a draft pattern is wrong across rows, when a domain is in trouble, or when a teammate flipped a row they should not have.

It is not for ordinary rejects. Ordinary rejects belong on the one job. If you halt the workspace because one dentist email was clumsy, you have turned a scalpel into a fire alarm. People start ignoring fire alarms.

After you halt, fix the cause before you lift. The matrix row, the ICP, the DEMO template, the sender, the collect URL. If you lift first, you get the same incident with a timestamp.

Tell anyone who shares the tenant. A kill switch nobody knows about is a surprise outage. A kill switch everyone understands is a safety tool. Write one sentence in the team channel: halted because the No-Website sequence used the wrong city. Back after the template is fixed.

If your stack has no kill switch, you can approximate it by taking the sending key out of the mailer and turning sequences off. Slower, uglier, still better than arguing with twenty in-flight jobs. The productized switch is the same idea with a log.

How to brief a teammate on approvals so two people do not collide

Brief a teammate with three lists: what they may approve unchanged, what they must edit, and what they must never approve. Put examples on each list from last week’s queue. Abstract principles are how two people both think they were careful.

Never approve: invented services, unlabeled dollars, Yelp replies you cannot post, deploy to a domain you do not have in writing, any collect page that grew a star filter, any SMS that is not first-party consented. That list is longer than it looks when you are in a hurry.

Must edit: names, locations, hours, the DEMO URL, the book-button destination. These drift. Supervised exists because they drift. If your teammate is new, they approve nothing for a week. They comment. You click. That is still Supervised. It is also training.

May approve unchanged: Watchtower diffs that match the last five, Scout reruns of a frozen ICP, Atlas reports that passed the credibility gate and still show assumptions on every dollar. If the credibility gate refused a bare number, nobody overrides it with taste.

Tenant isolation matters more than speed once you have two books of business. A teammate on the wrong Clerk organization is a different incident than a teammate who approved a clumsy sentence. v1 does not resell white-label sub-accounts. Isolation is still the feature. Confirm the tenant before the approval, every time.

What overnight playbooks can and cannot do today

Overnight playbooks can describe jobs you want run while you are away. They cannot be advertised as already sending in production if the runner is still the open half. Governance can be live. Dispatch can be coming. Those sentences are different. Print the true one.

A playbook is a named chain: search this ICP, scan the new rows, park the ones with the detector, draft the sequence, wait. The wait is the Supervised part. If you write the chain as if the wait were optional, you will teach a teammate the wrong story.

Mission Control is where those chains should be visible, with health of the jobs, not health of the client’s revenue. Client health is Healthy, Watch, or At-risk. Job health is whether the queue is sitting, running, or failed. Do not mash the words.

If you build playbooks in other tools, keep the same honesty. A Zap that sends mail at 2am is Autonomous send, even if you never called it that. If you would not flip that matrix row, do not build the Zap.

When dispatch does ship, you still do not flip first-touch email on night one. You flip the reads. You watch a week of drafts. Then you decide. The matrix does not get a launch-day party.

If a playbook includes Relay, the playbook includes the gate. Quiet hours, unsubscribe, verified domain, stop-on-reply. Autonomous never means the gate took the night off. If a chain cannot wait on a verdict, the chain is not a playbook. It is a blast with a nicer name.

How to audit a week of Autonomous without trusting the dashboard

Audit a week of Autonomous by sampling the artifacts, not by reading a success count. Open five searches, five drafts, five diffs. If even one invented a noun, the row goes back to Supervised.

Match each artifact to the business id. Autonomous that writes about the right leak on the wrong shop is worse than a slow queue. Persist-first handoffs exist so Radar, Pulse, and Relay see the same uuid. If your stack uses names, you will get collisions. Collisions are why ten-client isolation is a later guide and a now problem.

Check the locks as if they could have failed. Open the collect page. Confirm public buttons. Open the sender. Confirm the domain. Open voice settings. Confirm there is no outbound campaign. Paranoia here is cheaper than a platform email.

Write down what you will flip back. Audits that do not change a setting were either perfect or timid. Perfect weeks happen. If you have never flipped a row back, you are not auditing. You are sightseeing.

Do not invent an ROI for Autonomous. Time saved is a feeling until you count the hours you still spend fixing drafts. If you cannot count them, stay Supervised. This site will not print a productivity multiple for you, and you should not print one for a client.

How to read an approval queue as a log after thirty days

Read the approval queue as a log after thirty days by sorting what you approved unchanged, what you edited every time, and what you rejected, because the unchanged pile is the only Autonomous candidate. If you edited every Relay draft, the finding is weak, the ICP is wide, or the model is guessing services. Fix the finding before you flip send. Autonomous send on bad drafts is a domain funeral with a calendar invite.

If you approved every Watchtower scan unchanged, that scan can go Autonomous. If you rejected half because the source was unknown, keep it Supervised until the keys exist. Unknown is a first-class state. Autonomous should not paint it green. Keep a dated note of the flip: action, who approved the change, why the last month of artifacts were unchanged, what would flip it back. Mission Control should show this. A line in a shared doc is enough if it does not. Memory is how Autonomous becomes folklore.

If you have a teammate, decide together and write the initials on the matrix change. Autonomy without a name is how people argue later about who let the sequence out. Do not flip because a competitor's marketing site said autonomous GTM. Their motion is unattended outbound. Yours is a DEMO attached to a leak. Different cost of being wrong. A wrong email is a filter problem. A wrong deploy is a client's domain.

You can run this log without the software. A drafts folder and a rule that nothing sends from your phone at a red light is the same idea. The matrix is the version that stores the rule per agent and per action so a teammate cannot quietly lower it. Products that advertise zero approval clicks are optimizing for a different buyer. An AI SDR wants volume. This operating system wants a local leak, a sample, and a send you can defend.

What Off is for on a specialist you do not sell

Off is for a specialist you do not sell so that action does not run, even as a draft, because a reputation-only marketer can turn website deploy off so nobody ships a DEMO site by accident. Off is a kindness. Agent Mode is three-state: Supervised, Autonomous, Off. Default remains Supervised. The top-bar pill and Guardrails write the same row.

Read-only work may default freer. Anything that sends, posts, publishes, or spends defaults yellow until you opt in. Voice cold, SMS cold, and review gating stay red forever. Unverified send stays locked. Those four rows sit on the matrix as visible red so a marketer cannot hunt for a hidden toggle. If a row is missing, that is a defect.

Day one is also when people try to exempt Scout because search feels harmless. Search is harmless. The habit of exempting is not. If you start a list of harmless actions in your head, you will add Relay to it after a good Tuesday. Write the exemptions on the matrix or they are not exemptions. If a vendor UI offers a single Autonomous slider for the whole workspace, do not use it, or treat it as Off. Grain is the feature. Mood sliders are how ten clients get the same voice and the wrong sender.

Write the Autonomous sentence as a filter, not as a compliment to the model. Scout may rerun dentists in Austin with no booking on Mondays. That is a filter. The agent is getting good is not a filter. If you cannot write the filter, you are not ready to remove the click. Off means the filter is empty on purpose.

How a kill switch differs from rejecting one dentist email

A kill switch differs from rejecting one dentist email because ordinary rejects belong on the one job and a workspace-wide halt is for a pattern you do not want to unwind row by row. If you halt the workspace because one dentist email was clumsy, you have turned a scalpel into a fire alarm. People start ignoring fire alarms. Use the switch when a draft pattern is wrong across rows, when a domain is in trouble, or when a teammate flipped a row they should not have.

After you halt, fix the cause before you lift. The matrix row, the ICP, the DEMO template, the sender, the collect URL. If you lift first, you get the same incident with a timestamp. Tell anyone who shares the tenant. A kill switch nobody knows about is a surprise outage. Write one sentence in the team channel: halted because the No-Website sequence used the wrong city. Back after the template is fixed.

The kill switch precedes everything. Resolved tier is mode intersect matrix intersect playbook. Subject suppressions precede spend caps. Each dispatch appends an activity row. If your stack has no kill switch, you can approximate it by taking the sending key out of the mailer and turning sequences off. Slower, uglier, still better than arguing with twenty in-flight jobs. The productized switch is the same idea with a log.

A sender pause is for one domain. Learn the difference before you have ten, or you will halt the healthy nine to save the noisy one without meaning to. Tenant isolation matters more than speed once you have two books of business. Confirm the tenant before the approval, every time. v1 does not resell white-label sub-accounts. Isolation is still the feature.

What to tell a client who wants fully automatic outreach

Tell a client who wants fully automatic outreach that they often want faster drafts, not unattended first-touch, and that you approve every email that goes out in their name. Translate the ask into actions. Give them faster Supervised. If they still want unattended mail on a domain they own, you may refuse. Their domain, their ban. Your name is still on the agency. They do not need the matrix vocabulary unless they ask.

We approve every publish. We do not send texts. We do not hide review buttons. That is the brief. If you are the client-facing person and a partner flipped Autonomous send without telling you, that is a fire. Kill switch, then conversation. The matrix is a contract inside the tenant. Quiet flips are how books of business mix in spirit even when the database did not.

If a playbook includes Relay, the playbook includes the gate. Quiet hours, unsubscribe, verified domain, stop-on-reply. Autonomous never means the gate took the night off. If a chain cannot wait on a verdict, the chain is not a playbook. It is a blast with a nicer name. Overnight playbooks can describe jobs you want run while you are away. They cannot be advertised as already sending in production if the runner is still the open half.

You can follow this guide with a drafts folder and a written rule. Nerxus One is the version with a matrix, a gate, four permanent locks, and a kill switch. Supervised stays the default either way. Flip a row when the work is boring. Flip it back when it gets interesting again. If you remember one test, use this one. Would you let the action run while you were in the chair at the dentist you are pitching. If the answer is no, the row stays Supervised.

How ten clients change which rows you may flip

Ten clients change which rows you may flip because autonomy is more expensive when books of business multiply, and a wrong-preview send is a hard stop on new closes until you write the wall that would have prevented it. Flip reads, not first-touch sends, and only after the artifacts are boring. Ten supervised books are a business. Ten unattended books are a support queue with branding. Autonomous does not buy you client eleven. Autonomous on the wrong grain buys you eleven incidents.

Count hours that are not fulfillment. Approvals, portal questions, staff who added a star gate, a domain that needs warmup, a book button that points at last month's calendar. Those hours are the real capacity. Build hours lie. Support hours do not. If you cannot name leftovers without opening a tab, you are already mixing. Do not take the eleventh retainer to feel grown. Take a teammate, raise the FEE, or refuse.

Brief a teammate with three lists from last week's queue. Never approve invented services, unlabeled dollars, Yelp replies you cannot post, deploy without written domain access, a collect page that grew a star filter, SMS that is not first-party consented. Must edit names, locations, hours, DEMO URLs, book-button destinations. Isolation is a Clerk organization. Confirm the tenant before the approval, every time.

Default stays Supervised when you add a vertical you have never fulfilled. New verticals invent nouns. Invented nouns are why the click exists. Flip reads if you want. Leave send on Supervised until the new vertical's drafts are boring too. Write the chair test on the matrix change note. Chair test: no. That is enough of a reason to stay Supervised. You do not need a longer philosophy. Longer philosophies are how people talk themselves into a green row.

What overnight playbooks still cannot claim

Overnight playbooks still cannot claim production dispatch while the runner is the open half, because governance can be live and seeded runs are not a night shift. Mission Control, playbooks, approvals, activity, health, guardrails, a kill switch, spend caps, send windows: those nouns are real. The decision-time gate can allow, queue, or refuse. Nothing clocks in at 2 a.m. until that runner is real. Seeded runs stay labeled as seeds. Print the true sentence.

A playbook is a named chain: search this ICP, scan the new rows, park the ones with the detector, draft the sequence, wait. The wait is the Supervised part. If you write the chain as if the wait were optional, you will teach a teammate the wrong story. Dry run is a simulated pass. Nothing sends. Nothing spends. If you build playbooks in other tools, keep the same honesty. A Zap that sends mail at 2am is Autonomous send, even if you never called it that.

When dispatch does ship, you still do not flip first-touch email on night one. You flip the reads. You watch a week of drafts. Then you decide. The matrix does not get a launch-day party. Job health is whether the queue is sitting, running, or failed. Client health is Healthy, Watch, or At-risk. Do not mash the words. Do not invent an ROI for Autonomous. Time saved is a feeling until you count the hours you still spend fixing drafts. If you cannot count them, stay Supervised.

Audit a week of Autonomous by sampling the artifacts, not by reading a success count. Open five searches, five drafts, five diffs. If even one invented a noun, the row goes back to Supervised. Match each artifact to the business id. Autonomous that writes about the right leak on the wrong shop is worse than a slow queue. Check the locks as if they could have failed. Open the collect page. Confirm public buttons. Open the sender. Confirm the domain. Open voice settings. Confirm there is no outbound campaign. Paranoia here is cheaper than a platform email.

When to flip back, and how to say it to a client

Flip back when the artifacts drift, when a teammate leaves, when you add a vertical you do not know, or when a lock-adjacent close call happened. Pride is a bad reason to keep a row green.

Tell a client only what they need. We approve every email that goes out in your name. We approve every publish. We do not send texts. We do not hide review buttons. That is the brief. They do not need the matrix vocabulary unless they ask.

If a client asks for fully automatic outreach, translate it into actions. They often want faster drafts, not unattended first-touch. Give them faster Supervised. If they still want unattended mail on a domain they own, you may refuse. Their domain, their ban. Your name is still on the agency.

If you are the client-facing person and a partner flipped Autonomous send without telling you, that is a fire. Kill switch, then conversation. The matrix is a contract inside the tenant. Quiet flips are how books of business mix in spirit even when the database did not.

You can follow this guide with a drafts folder and a written rule. Nerxus One is the version with a matrix, a gate, four permanent locks, and a kill switch. Supervised stays the default either way. Flip a row when the work is boring. Flip it back when it gets interesting again.

If you remember one test, use this one. Would you let the action run while you were in the chair at the dentist you are pitching. If the answer is no, the row stays Supervised. The matrix is not a personality setting. It is that test, stored.

Write the test on the matrix change note. Chair test: no. That is enough of a reason to stay Supervised. You do not need a longer philosophy. Longer philosophies are how people talk themselves into a green row.

Default stays Supervised when you add a vertical you have never fulfilled. New verticals invent nouns. Invented nouns are why the click exists. Flip reads if you want. Leave send on Supervised until the new vertical’s drafts are boring too.