Three Gates — R4 Activation · LP Measurement · Level King Proof

SOW · 2026-07-22 · status: awaiting Akash decisions D1–D5 · supersedes WS1 of sow-master-r4-to-daily-engine-2026-07-20.md

Objective

Get ads live, make the funnel readable, and prove client delivery once — in that dependency order.

Why it matters

Zero ads running, zero calls booked, zero verified client leads. Net cash is −$1,004.

You need to

Answer 5 decisions (D1–D5). ~10 min. Everything else is mine.

Do this first, before reading anything else. Level King campaign 120247483840870714 is ACTIVE and spending right now, and three real leads have been sitting unanswered since Jul-21. Cindy Shaw Romaguera gave her phone, ZIP and said "Texting please" — nobody ever replied to her in Messenger. That is a live client-harm and reputation issue, and it is the only thing in this document that is urgent today. See D1.

The picture in one screen

Ads live0of 2,769 Net cash−$1,004business $2,659 Business runway86.5 dat current burn Calls booked0pending Verified client leads0Level King Lifetime paid spend$1,403→ 7 leads

The honest read: all three problems are real, but only one is urgent (Level King), only one is expensive (R4), and the cheap one in the middle (measurement) is what makes the expensive one worth doing. Activating R4 before the landing page can be measured would repeat exactly the last month: money spent, nothing learned.

D1 — Level King: pause the campaign and answer the three leads needs you

Campaign 120247483840870714 is ACTIVE, $19.42 spent, budget_remaining $17.62, 3 Messenger starts at $6.47 each.
The bot has never handled a real inbound lead. All 3 real conversations were Meta's native Page auto-flow, imported after the fact by backfill_meta_messenger.py on Jul-22.
Every "bot qualified" row in the database is a synthetic or owned-number test.
Dan has never received an automated lead. The one transfer that fired went to your phone ending 5477, 18 hours late.

★ Recommended: pause the campaign now, and you personally text Cindy today (she asked to be texted and gave a valid number). Reply "pause LK" and I'll pause it; the two texts are yours to send, not mine.

Alternative if you'd rather not pause: leave it running but accept that every new lead lands in the same silence until WS-C ships. I don't recommend it — $6.47 conversations are cheap and worth having, but only if someone answers them.

D2 — R4 spend ladder: $100/day is not affordable needs you

R4 is built and verified — this is purely a "how much" question.

Daily budget	Business-cash runway	14-day cost	Expected leads @ $35.57 CPL
$100/day (as built)	20 days	$1,400	~39
$50/day	33 days	$700	~20
$40/day	38 days	$560	~16
$0 (stay dark)	86.5 days	$0	0

★ Recommended: $50/day for 14 days, $700 hard cap, starting only after WS-B lands (48h). Keeps ~33 days of business runway, buys ~20 leads — enough to judge the three concepts, not enough to judge 27 individual ads, and that's the right resolution for this budget.

Math: business cash $2,658.70; business-only burn ≈ $30.70/day (86.5-day runway). Adding $X/day gives $2,658.70 ÷ ($30.70 + X). Meta CPL $35.57 is your real last-30-day figure, not an estimate.

D3 — Sequencing: instrument before you spend needs you

Right now nothing on the landing page reports a form start or a form submit. There is no postMessage listener anywhere (grep returns zero hits).
PostHog $pageview is broken on the LP — 387 people fired LandingOfferAssigned and 380 fired $pageleave, but only 103 fired $pageview. Every dashboard built on it undercounts by ~3.7×.
So if you activate R4 today, the result is unreadable: you will know spend and completed applications, and nothing about where people actually drop.

★ Recommended: WS-B first (≤48h, no spend), then R4. Reply "instrument first" to accept, or "launch now" to activate R4 immediately and accept a blind read.

D4 — What counts as a win, and when I turn it off needs you

A kill criterion set now is worth more than a judgement call made at day 10 with money on the line.

Checkpoint	Kill / continue rule
Day 3 (~$150)	Kill any ad set with >$40 spend and zero landing-page views. Delivery failure, not message failure.
Day 7 (~$350)	Continue only if ≥5 applications AND blended CPL ≤ $60. Otherwise pause all, keep the learning, spend nothing further.
Day 14 (~$700)	Hard stop regardless of result. Write the verdict, decide round 5 separately.
Any time	Pause immediately if Meta's next charge would not clear from business cash.

★ Recommended: adopt as written. Reply "kill rules ok" or give me your numbers.

D5 — Two facts I need from you that I cannot get myself needs you

Page ad limit. Meta's API does not expose the tier; the "5,000" figure is a client-side default. 36 R2 ads carry error 3253001 "Ad Limit Reached," which points at ~250. Open Ads Manager → the page-limit panel and tell me the number. 2 minutes. Wrong assumption here silently caps the whole estate.
The url_tags revert. Three ads were repaired on Jul-19 — the new tagged creatives still exist and are correct — but all three ads now point back at the old untagged creatives. Read-only data cannot tell me when or why. Did you or anyone roll this back?

★ Recommended: answer both in one line. If you don't know about the revert, say so and I'll treat it as an unexplained system behaviour and add a guard.

WS-A — R4 activation

A. Context and verified state reference only

What R4 is: 27 ads = 3 concepts × 9 niches, tweet-static format. Built Jul-20 after a 45-ad draft was rejected.

c1 AI-First split → stpierre.ai/resources/manifesto/ (9 ads)
c3 805-report proof → stpierre.ai/resources/market-report/ (9 ads)
c4 direct offer control → start.stpierre.ai (9 ads)

Fact	Value
Campaign	`120248219884220390` "SP R4 \| AI-First Champion" — PAUSED, CBO $100/day, OUTCOME_LEADS
Ad sets	9, all PAUSED, no ad-set budget, US 25–65, FB feed + IG stream, fully broad (no interests, no exclusions)
Ads	27, all PAUSED, page `1197436790119715` (Alex St. Pierre)
Spend ever	Zero — maximum-window insights return 0 rows for all 9 ad sets
Approval on record	Contact sheet approved by Akash 2026-07-20 13:35 ET; scope = "build PAUSED authorized; ACTIVATION requires separate go"
Readback	`readback-verification.json` — campaign/adsets/ads pass, coverage exact, DR/DC guard pass, DOF opt-out pass
Canon check	0 fails / 0 warns across 189 copy strings
Destination QA	27/27 pass, cache-busted curl on all 27

Niche message-match: verified clean. I checked all 9 R4 slugs against the LP alias map — basement-waterproofing, concrete-leveling, crawlspace-repair, deck-replacement, drainage-correction, foundation-repair, septic-replacement, siding-replacement, window-replacement all resolve. The roofing_replacement / concrete_leveling_large_projects mismatch found in live traffic came from older campaigns, not R4.

Estate context: 54 campaigns, 240 ad sets, 2,769 ads, 3,813 creatives. Lifetime spend $1,403.20 → 18,252 impressions → 490 clicks → 147 landing-page views → 7 leads. 53 ads DISAPPROVED (33 Unacceptable Business Practices, 15 Financial/Insurance, 5 Fraud/Scams — Meta returned no per-asset detail). 43 WITH_ISSUES, all in R2.

A. The work no action — mine once D2/D3 land

#	Task	Owner	Gate
A1	Set campaign daily budget to the D2 figure (default $50/day)	Claude	D2
A2	Apply url_tags to all 27 R4 creatives before activation, carrying `message_variant_id` + `experiment_id` + ad/adset/campaign macros. Must happen pre-activation — url_tags is immutable post-creation, so doing it later forces a creative clone that re-triggers ad review.	Claude	D3
A3	Verify the c1/c3 destinations (manifesto, market-report) carry the same instrumentation as the LP — otherwise 18 of 27 ads are unreadable	Claude	WS-B
A4	Activate: 37 discrete unpause operations (1 campaign + 9 ad sets + 27 ads — each is individually PAUSED, not inheriting)	Claude	D2+D3+D4
A5	Readback verification within 15 min of activation; screenshot contact sheet to you	Claude	—
A6	Daily 09:00 delivery check against the D4 kill rules; report only on a rule trip or on schedule	Claude	D4
A7	Day-14 verdict written to brain + a decision row for round 5	Claude	—

Not in scope, deliberately: repairing the 53 disapproved ads (Meta gives no per-asset reason — needs a manual Ads Manager review first), unpausing anything outside R4, and the daily additive ad engine (WS2 of the Jul-20 master SOW). One clean test before more machinery.

A. Acceptance criteria reference only

All 27 R4 ads at effective_status ACTIVE, confirmed by independent readback, not by write-response.
All 27 creatives carry non-empty url_tags including both attribution params.
ads_running_or_in_review_count > 0 on the account.
Spend never exceeds the D2 cap; a hard stop fires at $700 cumulative.
Every application arriving during the window resolves to a specific ad id.

WS-B — Landing page: make it readable, then make it convert

B. The 56→10 number was wrong — here is what's actually true needs you

I reported "56 landing-page views → ~10 Typeform starts" earlier. That figure does not survive contact with the data, on either side.

The "56" undercounts by ~3.7×. PostHog $pageview fires for 103 people while $pageleave fires for 380 and LandingOfferAssigned for 387. The library is loading and working — the initial pageview simply isn't captured. It's a config bug at index.html:53, not bot filtering.
The "10" came from a contaminated source. Of 88 recent Typeform starts, 73 share one desktop Mac user-agent with no hidden fields — that's your own verify-*.mjs QA automation. Real human starts ≈ 15.
The ratio was accidentally about right (~103 human loads → ~15 human starts ≈ 15%), but built from two broken numbers.
Bot cohort 424638 is now mislabelling humans — it's defined as "fires LP events but never $pageview," which is exactly what a real visitor now looks like.

Consequence: every CPL, start-rate and funnel figure produced in the last month is unreliable. Not wrong in direction — unreliable in magnitude. Fixing this is 1–2 days and it is the highest-leverage work in this document, because it is what makes the R4 spend worth making.

★ Recommended: approve WS-B as a block. It is all instrumentation and layout — no offer or copy rewrite without separate approval per your preview-before-deploy rule. Reply "WS-B go".

B. Verified causes, ranked reference only

#	Cause	Confidence	Evidence
1	No form-start / form-submit instrumentation exists anywhere	certain	`grep` for postMessage listeners returns 0 hits across the LP
2	PostHog `$pageview` broken on the LP	certain	103 `$pageview` vs 380 `$pageleave` / 387 offer-assigned
3	Zero Typeform question visible above the fold on desktop; ~1 option on mobile	measured	Desktop iframe top 645px in an 800px viewport; mobile 534px in 812px before iOS chrome
4	`disable-auto-focus=true` forces a scroll and a click into the iframe before anything can be answered	certain	`index.html:126`
5	Attribution silently truncated at the Typeform boundary	certain	11 hidden fields the LP sends aren't declared on the form (incl. `sp_angle_id` vs form's `angle_id`). `message_variant_id` is declared and always sent, yet populated 1 time in 161 responses
6	The LP is your weakest destination	verified	Since Jul-15: 2 applications tagged `offer-lp` vs 5 `90day-projection` and 5 `manifesto`
7	Median dwell 11.3s = message rejection, not mechanics	strong	p50 11s / p75 21s; iframe interactive at 1.4s; proof bar sits below the form
8	Cold traffic asked to fill an application as its first act, with no lower-commitment door	strong	The resource pages that do have one out-produced the LP 5:1

Ruled out: load speed (no JS errors, no failed requests, Typeform interactive at 1,369ms), go.stpierre.ai 302s (both /privacy and /terms return 200 first-hop — that note is stale), and niche skip logic (verified working: ?sp_niche=foundation-repair correctly rewrites the H1 and opens on Q2).

B. The work no action

#	Task	Effort	Approval needed
B1	Wire Typeform's postMessage bus: listen for `form-ready`, `form-screen-changed`, `form-submit`. Emit to PostHog + a new `sp_funnel_events` row.	3h	no
B2	Fix `posthog.init` so `$pageview` fires. Backfill nothing; mark the break window in the metric contract.	1h	no
B3	Redefine bot cohort 424638 off the now-invalid "no `$pageview`" rule; exclude the QA user-agent from all funnel reporting.	1h	no
B4	Give QA automation a distinguishable `application_token` prefix so test traffic is separable at source.	1h	no
B5	Declare the 11 missing hidden fields on form `j4wHbfWD`, and rename `sp_angle_id`→`angle_id` to match. Read-only inspection is done; this is a form-definition edit.	1h	yes — touches the live form
B6	Raise the form above the fold: compress the 4-line H1 + subhead + sub-subhead + chips, move the proof bar above the form, remove `disable-auto-focus`.	3h	yes — visual preview first
B7	Confirm manifesto + market-report pages carry the same instrumentation (blocks reading 18 of 27 R4 ads).	2h	no
B8	Rebuild the funnel view on trustworthy events; publish LP-view → form-start → form-complete → booked as one chain.	3h	no

★ B5 and B6 touch live surfaces. Per your preview-before-deploy rule I will send a visual preview and wait. B1–B4, B7, B8 are mechanical and I'll just do them.

B. Acceptance criteria reference only

A single query returns LP view → form start → form complete → call booked, for any date range, with test traffic excluded — and the numbers reconcile with Typeform's own response count within 5%.
message_variant_id and experiment_id populate on >90% of new responses (baseline: 1 in 161).
At least one Typeform question is visible above the fold at 375×812 and 1280×800.
$pageview person-count is within 10% of $pageleave person-count.

WS-C — Level King: prove the delivery promise once

C. What is actually proven, step by step needs you

#	Step	Verdict	Evidence
1	FB ad → Messenger conversation	PROVEN	3 starts at $6.47, $19.42 spent
2	Conversation → the new bot	NEVER HAPPENED	All 3 real convos were Meta's native Page auto-flow, backfilled Jul-22 by script
3	Bot qualification	TEST ONLY	5 synthetic/owned-number leads. The one "qualified" real lead was marked by the backfill script
4	CRM contact created	PARTIAL	Postgres yes (8 leads). GHL does not exist for this client — `ghl_location_id` NULL, no PIT, no custom values
5	Qualified stage/tag	MECHANICALLY PROVEN	Stage events with idempotency keys
6	Notification to client	WRONG RECIPIENT	1 transfer, to Akash's phone (5477), 18h late. Dan has never received a lead
7	CAPI Lead event	FAILING	Status `error`, HTTP 400, `events_received` NULL. n8n relay `gp1NILFWhEfkUGx3` is inactive

Client record: id 5cb23431-b9d7-4b9a-87bf-c339058cad3f, status upcoming (not active). Onboarding checklist 10/15. The new-launch-qa hard gate is failing: 4 pass / 5 fail / 3 warn. Step 9 blocker on record reads "Published active instead of paused" — which is why the campaign is live right now.

The parity audit is already partly stale in your favour. Files written later the same morning — messenger-bot-config.js and messenger-followup.js — do implement Jean's personality, exactly 21 FAQ ids, runtime client config resolved from the Page's own row (fails closed), and the verbatim 5-touch cadence with quiet hours, STOP/DND and a 24h policy guard. So the gap list is smaller than the audit says. But the rollout gate last read 20/28 → HOLD, and I did not re-run it.

★ Recommended: pause (D1), then run WS-C as a strict sequence — no client-facing sends until the gate reads 28/28 and one real lead has been handled live end to end.

C. The work no action — mine, except C1 and C7

#	Task	Owner
C1	Text Cindy today; decide whether Richard and Ann are still worth a reply	Akash
C2	Re-run the 28-check rollout gate; report the exact failing checks	Claude
C3	Close whatever the gate actually fails (the 8 known blockers may already be fixed in the later code)	Claude
C4	Fix the CAPI 400 and activate n8n relay `gp1NILFWhEfkUGx3`; stay inside the approved 10-event pilot	Claude
C5	Repoint lead transfer from your 5477 to Dan, with you on a copy for the first 10	Claude
C6	Close the new-launch-qa gate: fix M2 naming, set M4 `inline_comment` to OPT_OUT, get the campaign paused so the gate can pass honestly	Claude
C7	Owned-number canary: you message the Page from a phone you control, and we watch all 7 steps fire live	Akash + Claude
C8	Only after C7 passes: reactivate the campaign and take the next real lead live, with 72h close monitoring	Claude
C9	Stamp `launched_at`, flip status to active, finish checklist 15/15	Claude
C10	Decide whether Level King needs GHL at all, or whether Postgres + Messenger is the real product now	Akash — not urgent

C. Acceptance criteria reference only

One real inbound Messenger lead handled start-to-finish by the deployed bot, with stage events showing actor stpierre-messenger-bot and no backfill flag.
That lead reaches Dan, within 5 minutes, on a channel he confirms he received.
One CAPI Lead event with events_received: 1 and no error.
Rollout gate 28/28. new-launch-qa PASS.
Zero unanswered inbound older than the follow-up cadence allows.

Sequencing

What happens in what order reference only

When	What	Blocked by
Today	D1 — pause Level King, you text Cindy	your word
Today	D5 — you read the page ad-limit tier from Ads Manager (2 min)	you
Day 1–2	WS-B instrumentation (B1–B4, B7, B8). No spend.	D3
Day 1–2	WS-C gate re-run + fixes (C2–C6), in parallel — different systems, no contention	D1
Day 2	B5/B6 previews to you for approval	WS-B
Day 3	A2 url_tags on all 27 R4 creatives, then A4 activation at the D2 budget	D2, D3, D4, WS-B
Day 3	C7 owned-number canary	C2–C6
Day 4–17	R4 runs to the D4 kill rules; C8 Level King reactivation once the canary passes	—

Why this order: WS-B is cheap, fast, and makes WS-A's money readable — so it goes first. WS-C shares no systems with either, so it runs in parallel. WS-A goes last because it is the only one that spends, and it is the only one that can't be undone by editing a file.

Open questions and things I could not verify

Honest gaps in this plan reference only

True page ad-limit tier — Meta exposes no field. Circumstantial evidence points at ~250. Needs your eyes (D5).
Why url_tags reverted on 3 ads — not determinable read-only.
Per-ad disapproval reasons — Meta returned only 3 generic policy strings and empty issues_info on all 53 ads. Repair without a manual review would be guesswork.
Whether the post-audit Messenger code is actually deployed and green at 28/28 — I did not re-run the gate (C2).
Whether 34 zero-spend campaigns ever served — Meta suppresses zero-impression rows, so "never built a deliverable ad" and "never delivered" are indistinguishable.
Your retainer price — repo grep returns a mix of $1,500/mo and $1,700/mo. I did not use a revenue figure anywhere in this SOW's math for that reason. Confirm it and I'll add breakeven and payback to the day-14 verdict.

Methodology & sources reference only

Meta: scripted Graph v24.0 via vault token, read-only, 2026-07-22 ~10:04 UTC. Zero writes. Account act_1096498922089551.
PostHog: project 493312, HogQL over events where $host='start.stpierre.ai', Jul-15→22.
Typeform: Responses API read-only GET on form j4wHbfWD. The live form was never driven or submitted.
Postgres brain: finance_metric_snapshots, sp_funnel_events, sp_contractor_applications, sp_client_leads, sp_client_messages, sp_client_lead_transfers, sp_capi_fired_events, clients, sp_onboarding_checklists, 25 session_handoffs Jul-18→22.
Cash figures: finance_metric_snapshots as_of 2026-07-21 — net cash −$1,004.40, business cash $2,658.70, weekly burn $1,124.13, business-only runway 86.5 days.
Browser: start.stpierre.ai loaded at 1280×800 and 375×812; geometry via getBoundingClientRect, timings via Resource Timing. Form never filled.
Source audits: messenger-bot-parity-audit-2026-07-22.html, round4-pack-2026-07-20/, sow-master-r4-to-daily-engine-2026-07-20.md, audit-lp-reality.md.

No external writes, sends, deploys, activations or form submissions were performed in producing this document. Everything in it is a read.