Your sends are rationed. Prove which ones deserve to exist.
RevenueOS ranks every client program's next best send, routes high-stakes messages for human approval, and publishes a client-ready ledger proving which angles moved replies, meetings, and pipeline.
Built for outbound agencies and GTM teams running 5,000+ monthly sends across multiple client programs.
The sizing math — can your volume prove anything? — runs before the invoice does · or write directly: info@revenueos.app
ON THE RECORD — RevenueOS is a proof layer for cold outbound: built and internally verified to run randomized holdout experiments across a program's email angles, publishing confidence bands — with the arithmetic public — showing which messages cause replies, meetings, and pipeline. Customer enrollment opens with the design-partner pilot. For agencies on Smartlead, Instantly, and Clay.
(the holdout is the slice of each list deliberately never mailed, so the winner has something honest to beat)
Opened two ops roles in three days — the operations-scaling angle is the one under test on this program. Ranked above 41 other sends competing for today’s capacity.
Maya — saw Halden opened two ops roles this week. Most teams hit that hire because routing is manual, not because headcount is short. Worth fifteen minutes on how the other three we work with sequenced it?
The page you show the client.
One row per angle under test. The band is the range of lifts compatible with the data so far — a band sitting wholly above 1.0× rules out no-effect at the stated 90% level; a band straddling 1.0× is honestly unresolved. The gate is the pass/fail bar we wrote down before seeing any data, and the two columns are allowed to disagree: an angle can be retired because it cannot clear the gate while the evidence about it stays undecided. The bands hold even when checked weekly. No adjectives.
Measured here · angle lift — does one angle beat another
SPECIMEN — SEEDED DEMO TENANT
SPECIMEN — SEEDED DEMO TENANT
SPECIMEN — SEEDED DEMO TENANT
reading a row — band: the lifts compatible with the data so far · gate: the bar fixed before any data · rows: collected of required — per the published table, or sized per tenant where the table publishes no cell. A celled row’s target sets its rows required — sizing only; the decision is judged against the gate alone, and the gate never moves.
Three jobs. One deliverable.
Not a dashboard you interpret. A queue you sign, a ledger that grades the claims, and a book that keeps what survived.
The best next send, first.
Every morning, each client program’s proposals ranked by priority, soonest-expiring first — each row carrying the model’s action confidence when a model supplies it. Who to reach, what to send, why now, reasoning attached — rationed volume goes to the sends that deserve it.
a approve · r reject · e edit · u undo
Specimen — seeded demo tenant
Today — Tue 07 Jul12 pending
Dana Whitfield · Northwind Analytics
soft-cadence follow-up
0.91action confidence
Maya Okonjo · Halden Logistics
operations-scaling angle
0.67action confidence
Priya Raghavan · Coastline Freight
social-proof opener
0.58action confidence
the morning sort — no expiry pressing, so Dana’s row leads, written firm
seeded example — action confidence is how strongly the model behind a suggestion backs its own proposed action: a reading to weigh, never a measured outcome probability. Live rows show it only when a model supplies one — without it the row says so and asks for your judgment; the register never invents a number
Every send carries a name.
Anything high-stakes routes to a human before it leaves — approve, edit, or reject in seconds. The signature is the audit trail’s first entry, and your deliverability’s last defense.
wine-red marks human judgment — the machine never wears it
09:14 — one decision
Dana Whitfield · soft-cadence follow-up
→ email · draft reviewed
→ cleared to send · the decision is recorded — who decided it, and when
the same row — now it carries a name
A ledger clients can audit.
Message angles run as honest experiments. Verdicts publish with confidence attached — wins, nulls, and still-collecting alike. Validated angles join the book and compound across clients.
rungs: reply → positive reply → meeting
Dana’s angle, graded by the math
soft-cadence follow-up
EVIDENCE · FAVORABLE EFFECT ESTABLISHEDDECISION · PROMOTE
positive-reply · target 2× · cell B3 · rows 12,944 of 12,944 required · band [1.21×, 1.58×]
gate 1.15×
0.8×
1.0×
1.8×
the band clears the gate — entry closed, in the book
The weekly deliverable
What was tested. What won. What failed. How confident we are. What changes next week. Client-ready, recomputable, yours to forward.
WK 27 · acme-co-proof.pdf · 3 claims — 1 promoted
the 1 promoted: soft-cadence follow-up — the entry above, posted to the book
We qualify you before we invoice you.
Design-partner pilot
design-partner pilot · san diego, ca · pressed 07 jul 2026
$5–15K /mo
scoped by program count
flat — never a % of lift · the band’s top is the pilot’s whole-book ceiling up to fifteen programs · exact figure on the call
Forward-deployed — the engineer who built it joins every call.
We author your first angle book with you.
Your own claim ledger, from your own sends, in weeks.→ schedule a
Weekly readout ritual — objections tallied, roadmap steered.
The published anytime-valid floor — anytime-valid meaning the band stays honest however often you check it — previewed on your numbers here, confirmed in writing before the invoice.
Clause — the pilot operating threshold · checked live, identity after the math
Meets the pilot operating threshold.
The sizing math runs on your real volumes on the call — can they prove anything? — before any invoice.
sends 6,500 against 5,000 (meets) · programs 2 against 2 (meets)
Illustrative published sample requirement — cell A1, conservative 0.64% reply
≈151.9 months to a verdict (~12.7 yrs)
21,928 rows · 0.1% → 0.64% (holdout → proof reply rate)
assumes your 2 programs share the monthly volume equally
months at 6,500 sends/mo across 2 programs — the volumes entered above; the rows required never change with volume, only the months-to-verdict do
the conservative floor — a healthier reply rate and the accelerated design shorten this sharply; we run your real numbers on the call. published anytime-valid table · generated 2026-07-03 · cell A1 · the full table
Two clocks run here: first learnings land in weeks, but a statistically decisive verdict under the registered design takes the sample the table says — and if that is quarters or worse at your volume, we say so before any invoice.
the pilot operating threshold is public: 5,000+ monthly sends · 2+ programs — nothing stored, nobody asked who you are. the pilot, on one page
G. MAHN — countersigned 07 jul 2026
founder · the engineer on every pilot call
your countersign
no mail app? write directly: info@revenueos.app
Rider — dated 07 jul 2026
Self-serve and platform tiers open after the design-partner cohort. Join the waitlist.
Schedule A
The first six weeks
activities, not verdicts — results take the time they take
Connect Smartlead / Instantly / SendGrid + your CRM (HubSpot, Salesforce, or Copper); every parser must prove itself on captured live events before anything sends.
Angle book v1 authored together; experiments enrolled.
First sends signed from the queue — a approve · r reject · e edit · u undo.
Outcomes land as typed rungs; the first bands begin to draw.
Claims on the record — each with its evidence state and its decision, settled or not.
First weekly readout pressed — client-ready, recomputable, yours to forward.
Countersigned above — Grant Mahn, founder, and the engineer on every pilot call.info@revenueos.app
Straight answers.
Do we have enough volume for this to work?
The pilot operating threshold is roughly 5,000 sends a month across programs — a commercial bar, and a separate question from the published sample requirement below. Before any invoice, we run your real volumes through the minimum-n table — the rows needed to detect 1.25×, 1.5×, or 2× lifts on a 2% positive-reply base. If honest detection takes quarters, we say quarters and suggest the audit cadence instead of the subscription.
(Exhibit A entered — the minimum-n table. It runs.)
Exhibit A — the published anytime-valid table, on your numbers
The numbers below are the shipped engine’s own published table — generated by the same engine the product runs, never a textbook formula. This exhibit only selects a cell; it never invents a number.
① First — prove the outreach works · program incrementality — does outreach beat no-outreach
≈123.4 months to a verdict (~10.3 yrs)
21,928 rows · 0.1% → 0.64% (holdout → proof reply rate)
conservatively snapped to the published grid — your 1% reply rate → the published 0.64% reply rate.
months at 4,000 sends/mo per program — the volume entered above; the rows required never change with volume, only the months-to-verdict do
② Then — prove one angle beats another · angle lift, at a 2% reply base
Two clocks run here: first learnings land in weeks, but a statistically decisive verdict under the registered design takes the sample the table says — and if that is quarters or worse at your volume, we say so before any invoice.
These ARE the anytime-valid floors — conservatively snapped to the published grid, dated, recomputable (α .1 — a 90% band; safe to peek weekly). published anytime-valid table · generated 2026-07-03 · cell A1. the full table · your exact per-tenant table ships with the pilot (the pilot, the floor clause).
Why isn’t pricing tied to the lift you find?
Because a proof layer that takes a percentage of the number it reports has an integrity problem measuring its own paycheck. Flat fee, stated up front — the standing hypothesis is $5–15K flat for the pilot, sized to volume and revised with the cohort — and for books up to fifteen programs the band’s top is a ceiling for your whole book: an agency running twelve programs pays the top of the band, not twelve times a per-program rate; larger portfolios are scoped on the call. We sell the audit, never the outcome of the audit.
What if the readout says our best angle does nothing?
Then you stop spending capped sends on it — that’s the product working. Every row publishes whether or not it clears the gate, and the ledger keeps two things apart that most tools merge: whether the evidence settled, and what the pre-registered rule decided. An angle can be retired because its band rules out the gate while the evidence about it stays honestly undecided; those are different findings and the ledger will not merge them. A proof layer that only returns good news is an ad. Retiring an angle that cannot reach the gate at n=1,102 is worth real money at 2026 volumes.
Why does a human approve every send?
Two reasons. Deliverability: 2026 complaint ceilings do not forgive autopilot. Accountability: when a prospect gets an email, someone’s name is on that decision. The queue makes the signature take seconds — a approve, r reject — minutes a day, not meetings.
Where does our data go?
Into your tenant, and nowhere you haven’t signed for. Per-tenant row-level isolation; no decision commits without its own record — who decided it, and when — written in the same tenant-scoped write as the decision itself; processing runs on the subprocessors named in the DPA; GDPR DSR paths built and timed. Cross-client benchmarks are opt-in, aggregate-only, and never reported below a five-tenant floor — the consent clause is in the DPA on day one, not retrofitted across a signed book.
(Exhibit B — the subprocessor list, as printed in the colophon: Clerk · Supabase · Stripe · SendGrid · Fly.io · Vercel · Anthropic · OpenAI · Upstash · Sentry · Grafana — under DPA.)
Are you SOC 2?
SOC 2 is planned ahead of mid-market procurement. Pilots today receive the security-posture one-pager — architecture, per-tenant isolation, encryption, and audit logging of administrative access, data-subject-request execution, and tenant erasure, each written before the operation proceeds — plus the DPA with the full subprocessor list and the pooling-consent clause included from day one.
current as of sep 2026 — this answer will be revised on the record when soc 2 lands.
Where do replies live — do you have an inbox?
In your sending tool, where they already are. Smartlead’s Master Inbox, Instantly’s Unibox, and the equivalent in whichever platform you run already centralize every reply, and that is where your team keeps reading and answering it. RevenueOS is send-side: it ranks the next send, routes it to a person to approve, and reads each reply back as a typed outcome — reply, positive reply, meeting — so the ledger can grade what worked. We read the outcome, not the inbox, and we never ask you to move a conversation to us.
Certification
The foregoing answers are given straight, on the record, and revised only in writing — struck, corrected, left legible.
G. Mahn — the engineer answering · san diego, ca · jul 2026
Everyone else helps you send. We prove what was worth sending.
The design-partner pilot. The math before the money.