FM04: Call path, spend mix, and early-warning safety case

FM04 execution receipt · 2026-09-17 · Contract revision 2

Audit status: DONE for the authorized read-only audit and parent handoff. The corrected target outcome is not marked green: current location-specific spend is 9.898015765671%, widget-only click volume is not verifiable from the current event schema, and nearby urgent call quality has no separate outcome evidence.

Corrected goal carried forward: strictly more than 50% of spend on location-specific keywords, at least five verified Google-Ads-originated phone-widget clicks, and separately evidenced nearby urgent call quality. A widget click is not a connected call, and a connected call is not automatically nearby or urgent.

Scope and identity: Google Ads customer 4891907898, Csupi Autómentés, HUF, Europe/Budapest, auto-tagging enabled. Active target scope is 21 campaigns with the current EXACT | structure. Measurement property is GA4 266986146 and the site container is GTM-TH67S8M. The parent is task 01a0a269-78d2-7191-9709-7009b8d01715 on local.

Safety boundary: No Ads, budget, campaign, keyword, ad, GTM, GA4, consent, website, or deployment mutation was issued. No paid ad was clicked, no fake gclid was created, no synthetic conversion was sent, no form was submitted, no real call was placed, no caller PII was read, and no private evidence was published.

1. Deliverable A, all active final URLs and response integrity

Fresh read-only URL probe at 2026-09-17T11:38:20Z (13:38:20 CEST) used curl -sS -L with the exact active URL inventory from the Ads readback. Result: 21/21 HTTP 200, 21/21 final paths unchanged, and no redirect observed. Every response had a title, an H1, one GTM-TH67S8M marker, one tracking-consent.js, one phone-reveal.js, one widget-phone, and at least one gfn-track-phone phone link. The response body sizes were 47,700 to 54,808 bytes and the observed HTML response times were 0.206 to 0.423 seconds. This is a response check, not a Core Web Vitals lab measurement.

2. Representative desktop and mobile checks

The paid-path reveal control is backed by the live phone-reveal.js source, which calls event.preventDefault() before reveal and copy, while retaining the tel: href and gfn-track-phone class. Because that retained class is intentionally observed by GTM, the control itself was not clicked. The live tap-to-dial path therefore remains an explicitly unperformed behavioral check, not a false pass.

3. Failure-path checks for landing and phone UX

Scenario 9 verdict: PREVENTED at the checked technical layer. All 21 active URLs responded with intact marker content, and representative direct, paid-state, desktop, and mobile paths rendered the phone route without overflow or console errors. Residual per-page interaction and performance risk is UNKNOWN because a real paid session and lab performance trace were not permitted.

Scenario 14 verdict: PREVENTED for observed UI visibility, with the five-click target still UNKNOWN. The widget is visible and labelled on desktop and mobile, the paid masked state exposes a clear reveal control, and the mobile layout has no overflow. No live tap was performed, and the current event model does not isolate the floating widget, so this is not conversion or call-quality proof.

4. Deliverable B, keyword-level location-spend classification

Fresh Google Ads API readback covered customer 4891907898, active target campaign IDs only, and the closed customer-timezone period 2026-08-18 through 2026-09-16. The scope query required campaign.status = ENABLED and the 21 current target IDs. It did not filter keyword status, so cost from a criterion later paused or removed during the period was not silently dropped. Cost is metrics.cost_micros / 1,000,000.

The keyword classifier normalizes case and Hungarian accents, tokenizes the keyword text, and marks a row location-specific only when the keyword itself contains one of these explicit tokens: biatorbagy, budakeszi, budaors, buda, csepel, diosd, erd, pomaz, szentendre, solymar, pilisvorosvar, piliscsaba, zsambek, torokbalint, m0, m1, m5, m6, m7, or m31. The buda token was explicitly included in the corrected rerun. Campaign names and ad-group names were not used as classification evidence.

Reproducible scenario 15 calculation: the query sums integer micros before rounding: location-specific cost L = 13,160.631008 HUF (13,160,631,008 micros), generic cost G = 119,801.685072 HUF (119,801,685,072 micros), and total T = L + G = 132,962.316080 HUF (132,962,316,080 micros). The displayed campaign-level total rounds to 132,962.32 HUF. Share is L / T × 100 = 9.898015765671220248%. The strict test is 2 × L > T, or 26,321.262016 > 132,962.316080, which is false. Therefore scenario 15 is not passed.

Two anti-error checks are explicit: buda autómentés is counted as location-specific even though its criterion was later paused, while autópálya autómentés in the M5 campaign is generic because the keyword text has no route token. Conversely, location terms inside the Generic campaign are counted as location-specific. This prevents campaign-name-only classification.

Scenario 15 verdict: EXPOSED. Generic terms consumed 90.101984234329% of the measured active-target spend, leaving location-specific terms at 9.898015765671%. No spend recommendation or Ads mutation was made before separating delivery and measurement integrity.

5. Read-only measurement and call-quality evidence

The current GTM evidence uses a broad Click URL contains tel: trigger for phone-link events. The live page assigns gfn-track-phone widget-phone to the floating widget, but the current provider event rows do not expose a widget-specific dimension. Therefore the 18 paid GA4 events and 17.59 Ads all-conversion value are retained as proxies, not promoted to the corrected five verified widget-click gate.

6. Daily ledger definition and early-warning controls

Daily join key: one complete Europe/Budapest date, with the normal provider reporting-delay note. No caller name, phone number, recording, transcript, vehicle registration, or exact address is stored. Missing outcome fields are UNKNOWN, never zero.

Current ledger baseline: the supplied FM04 tracker did not contain a completed daily join. Current aggregate evidence therefore records verified_widget_clicks=UNKNOWN and connected_nearby_urgent_calls=UNKNOWN. GA4's 18 paid phone-link events and Ads proxy values remain separate evidence rows.

Scenario 16 verdict: EXPOSED. The tracker now defines a daily joined evidence model and trigger schedule, but the current supplied evidence does not contain a completed daily widget-to-call-quality ledger. The shortfall can be detected from the next complete read, but it is not yet proven historically.

7. Owned scenario verdict receipt

8. Acceptance receipt

• [x] The owned scenarios 9, 14, 15, and 16 each have a PREVENTED or EXPOSED verdict, evidence, a negative/failure-path check, an owner, and a pre-day-14 trigger.
• [x] All 21 active final URLs are listed with campaign ID, RSA ad ID, exact URL, HTTP result, unchanged path, and marker summary.
• [x] Representative direct desktop, direct mobile, location-specific mobile, and paid-state mobile checks include repeatable viewport, DOM, widget, overflow, and console evidence, with screenshots captured in the run.
• [x] Scenario 15 uses a reproducible keyword-text classifier and strict arithmetic. The result is 9.898015765671%, so the greater-than-50% acceptance criterion is correctly exposed as not met.
• [x] The daily ledger separates Ads spend, impressions, Ads clicks, widget-only verified clicks, connected calls, nearby calls, urgent calls, and the nearby-urgent intersection. Missing call outcomes remain UNKNOWN.
• [x] The corrected goal remains intact: more than 50% location-specific spend, at least five verified phone-widget clicks, and separately evidenced nearby urgent call quality.
• [x] No real call, paid ad click, synthetic conversion, credentials exposure, data publication, account mutation, website change, or deployment occurred.
• [x] The tracker remains valid HTML and every visible text node added by this receipt is inside #agent-progress, whose CSS forces blue text.

Known limits: A genuine paid-session click, a real floating-widget click, one-to-one event identity, connected call, nearby classification, and urgent classification were not performed or available. They are deliberately not claimed. This audit exposes the current gates and gives the next complete daily run a deterministic acceptance test.

Local privacy note: The tracker remains a private local artifact. No public review host was used because the evidence is customer-specific and the contract forbids publication.