STEP 03 · ANALYZE4-PASS AI ENGINE

Detect. Diagnose. Recommend. Narrate.

Sentrum runs four model passes per insight — a fast anomaly scan, a structured diagnosis against a locked seven-section cause-tree, action recommendations tied to the diagnosed cause, then a plain-English narration. Every insight cites the data, names the driver, and gives you a specific next move.

Get Started← Back to overview
THE PIPELINE

Four passes, each with a specific job.

Single-prompt LLM dashboards produce shallow takes. We split the job so the diagnostic logic is auditable and the narration stays grounded in cited data.

01DETECT

Anomaly scan

A fast pass over 40+ signals against per-account rolling baselines — not industry benchmarks. Flags only the metrics that broke your normal range, with severity score and time-to-detect.

Models: GPT-5-mini / Sonnet 4.7 · cached
02DIAGNOSE

Cause-tree walk

Walks the locked diagnostic order: Spend → Primary KPI → Efficiency → Supporting rates → Mix shifts → Optimization edits → Signal quality. Attributes the change to 1–2 drivers from the approved list.

Models: Sonnet 4.7 (deep) · structured output
03RECOMMEND

Cause-bound actions

Every recommendation maps to the diagnosed cause. No 'test new creative' without fatigue signals to point at. Includes expected directional impact and the watch-metric to verify the move worked.

Models: Sonnet 4.7 · cause-action lookup
04NARRATE

Plain-English story

Turns the structured diagnosis into the headline + What changed / Why / So what / Now what format. Always cites the numbers. Always names the driver. Never says 'the algorithm got better'.

Models: Sonnet 4.7 (narrator) · streaming
DIAGNOSTIC ORDER

Locked. Same on every insight.

The order matters. Diagnosing CTR before checking spend leads to wrong calls. Sentrum runs the same sequence every time so every insight is comparable to every other insight in the same account.

1
Spend (input)
Did budget change? Pacing or scaling?
2
Primary KPI volume (output)
Results = the optimized event (Purchase / LPV / ThruPlay).
3
Efficiency metric
Cost-per-result / rate. Is it directional or noise?
4
Supporting metrics
CPM, CTR, VTR, CVR — pinpoint the layer that moved.
5
Mix shifts (allocation)
Did spend share swing ≥20% across ad sets, ads, or placements?
6
Optimization edits
Match inflection points on the timeline to logged edits ±1–3 days.
7
Signal quality (2026)
EMQ, CAPI coverage, AEM event priority — check before recommending creative.
Sample insight · ASC Spring Catalog
Purchases up +18% on flat spend as CPM fell −6%.
WHAT CHANGED

Spend held at $2.4k/day. Purchases climbed from 38 → 45 PoP. CPM dropped from $19.20 → $18.05. CTR stable; CVR steady.

WHY

Auction pressure eased (CPM −6%) as the broader-targeting ad set caught delivery share from a tighter retargeting ad set you paused on day 3. Mix-shift driver, not creative.

SO WHAT

Sustainable for ~5–7 days under current frequency. Watch reach saturation; frequency 2.1 → 2.4 means you're approaching the saturation slope.

NOW WHAT

Raise daily budget +15% on the winning ad set while CPM stays low. If frequency crosses 3.0, refresh creative before scaling further.

EVERY INSIGHT INCLUDES

Four cuts, every time.

This format is enforced by the engine, not a suggestion. If an insight can't fill all four, it doesn't ship — that's how we keep generic advice out of the product.

WHAT CHANGED

Quantified delta + the period, the segment, the metric. No vibes.

WHY

1–2 drivers from the approved list (spend, CPM, CTR/VTR, CVR, mix shift, edits, signal quality). Cited.

SO WHAT

Risk or opportunity. Sustainable or short-lived. Confidence band based on volume + signal agreement.

NOW WHAT

Specific action tied to the diagnosed cause + the watch-metric to verify it worked.

Approved drivers · max 2 per insight
  • Spend / pacingBudget moved; volume followed.
  • Auction cost (CPM)Targeting pressure, fatigue, placement mix.
  • Engagement (CTR / VTR)Creative resonance or fatigue.
  • Conversion quality (CVR / LPV rate)Landing page, offer, traffic quality.
  • Mix shift (allocation)Spend share moved ≥20% across segments.
  • Optimization editsBudget / audience / creative changes ±1–3 days.
  • Signal quality (2026)EMQ, CAPI coverage, AEM event priority.
Significance thresholds · so we don't overreact
MetricFlag if change ≥
Spend / Primary KPI / Efficiency±10%
Mix shift in spend share±20%
CTR / VTR (relative or absolute)±10% or ±0.20pp
Segment callout (min spend share)≥10%
Segment callout (min KPI volume)≥15%

Confidence band (High / Medium / Low) accompanies every insight. Low confidence always explains why — usually low volume or short window, not model uncertainty.

WHAT YOU SEE IN THE PRODUCT

The cuts, by surface.

Daily Pulse

Morning brief — what moved, what needs attention, what to scale. Anomaly flags, AI summary at the top of the screen, Signal Quality card, ASC vs Manual delta when both exist.

Campaign Deep Dive

Per-campaign drill with the full diagnostic order, optimization-log overlay on every metric chart, breakdown tabs (age, gender, placement, ASC user_segment_key), creative-by-creative attribution.

Ask Soren

Natural-language Q&A grounded in your historical baselines + the cause-tree. Not a pasted CSV. Pro + Agency get follow-up questions and saved threads.

Cadence Reports

Weekly, biweekly, monthly, campaign-wrap. Same four-cut format per insight. Email delivery on your schedule. Excel + branded PDF export.

Creative Intelligence

Per-creative diagnostics: archetype classification, Gemini video analysis on hook strength, fatigue surfacing tightened for Andromeda's faster delivery shifts.

Signal Quality

EMQ, CAPI matched / Pixel-only / CAPI-only coverage, AEM event status, recommendations. 2026's most under-diagnosed CPA inflation cause.

NEVER · BY DESIGN

What Sentrum refuses to say.

  • “Test new creative” without a fatigue signal to cite
  • “Performance improved because the algorithm got better”
  • Recommendations not tied to a diagnosed cause
  • Vanity metrics without context (impressions ↑ celebrations)
  • Generic best-practice copy ungrounded in your data
  • CPA flag on a value-bid campaign without checking ROAS first
  • “Winners / losers” without spend share ≥10% or KPI ≥15%
  • AI hallucinated numbers — every figure cites raw_insights
ALWAYS · BY DESIGN

What Sentrum actually says.

The mirror of the list above. Every banned shortcut maps to a specific replacement — what Sentrum produces instead, grounded in the data it just read.

Instead ofSentrum says
“Test new creative”“CTR dropped 18% over 9 days while CPM held flat — fatigue signal. Refresh with a UGC archetype; you're currently 100% in static-product.”
“The algorithm got better”“CPM fell 6% as the broader ad set caught delivery share from the paused retargeting set. Mix-shift driver, not auction-pressure.”
Untied recommendation“CVR is down 22% with CTR stable — check the landing page. Likely cause: speed regression or offer mismatch from the May 8 deploy.”
Vanity metric celebration“Impressions +24% — but frequency rose from 1.8 to 3.1 and CTR fell. Reach saturation; not a win. Expand audience before scaling.”
Generic best-practice copy“In your account, 25–34 women drive 41% of purchases on 22% of spend. Shift +10% budget toward the ad set indexed there.”
CPA flag on value-bid“CPA rose $28 → $34 but ROAS held at 4.2× because value rules upweighted high-AOV customers. Net efficiency unchanged — ignore.”
Premature winner/loser“ASC-B has the best CPA but holds 6% of spend and 9% of KPI volume — below the threshold to call. Continue monitoring at scale.”
Hallucinated number“Spend $2,418 · purchases 45 · revenue $5,840 — all from raw_insights between May 12–18. ROAS 2.41×.”

Every figure cites raw_insights · every driver is from the approved list · every action ties to a diagnosed cause

READY

See the four-cut format on your own account.

Flat monthly pricing that never scales with your ad spend. Card required at signup. Diagnose on day one.

See pricing
Detect, diagnose, recommend, narrate — Sentrum’s analysis engine