Performance Loop
Top 25% versus bottom 25%, what actually drives results.
Veeva performance data, your data, you own it, flows back into the engine as signal. Which angles, formats, and audience hooks won. Which sank. The system learns what works for Compound DT-204 and uses it on every new brief.
Versions in flight · 30d
Sample
286
Across 8 active campaigns
illustrative target
Avg engagement lift
Sample
+47%
illustrative target
Top 25% averaging 2.8× the bottom 25%
Veeva sync · last hour
✓ Live
All versions synced
Auto-pushed on MLR approval
Cost per qualified lead
Sample
$8.40
illustrative trend
Caregiver audience driving the drop
Attribute analysis
Which attributes appear in winners vs losers
Each row is a content attribute. Top % = how often it appears in the top-25% performing variants. Bottom % = how often in the bottom-25%. The lift column tells you what the engine should do more of.
Attribute
Top 25%
Bottom 25%
Lift
Engine bias
Caregiver hook
78%
12%
+6.5×
↑ More
Named physician (Dr. + credential)
64%
8%
+8.0×
↑ More
Trial-cited efficacy stat
81%
23%
+3.5×
↑ More
Patient story arc · 30s+
56%
18%
+3.1×
↑ More
ISI-adjacent fair-balance
100%
64%
+1.6×
↑ More
First-person singular voice
4%
73%
−18×
↓ Avoid
Emoji density > 3 per asset
6%
51%
−8×
↓ Avoid
Generic 'glow / wellness' framing
0%
38%
−∞
↓ Avoid
Top 25%
Winners · last 30 days
1
Caregiver POV · 'When you knew it wasn't a heart attack'
Diagnosis journey · Reel · 30s
Caregiver hookPatient storyNamed cardiologist
8.4%
ENGAGEMENT
+218%
2
RP 101 · 60s explainer
Educational · Reel · 60s
Disease stateMechanismTrial-cited
7.2%
ENGAGEMENT
+186%
3
Caregiver Q&A carousel
Q&A · Carousel · 8f
Caregiver hookQ&A formatCost mention
6.9%
ENGAGEMENT
+174%
4
Cardiologist authority · 30s
Authority · Reel · 30s
Physician voiceTrial-citedAuthority
6.1%
ENGAGEMENT
+148%
5
Compound DT-204 mechanism · static
Educational · Static · 1080×1350
MechanismVisual diagramTrial-cited
5.3%
ENGAGEMENT
+114%
Bottom 25%
Underperformers · diagnosed
1
Stock medspa-style 'glow' caption
Promo · Reel · 15s
Generic copyNo physicianNo citation
0.8%
ENGAGEMENT
−68%
Why it sank: First-person singular voice. Reads like a wellness brand. No clinical anchor.
2
Cost emoji-heavy carousel
Promo · Carousel · 6f
Emoji-heavyNo ISI proximityGeneric CTA
1.1%
ENGAGEMENT
−56%
Why it sank: Fair-balance gap caught by MLR but not until late. Underperformed even when republished compliant.
3
Generic 'inflammation' explainer
Educational · Static · 1080×1080
No disease specificityNo citationNo CTA
1.4%
ENGAGEMENT
−42%
Why it sank: Disease-state too generic. Audience signal: 'not for me, scrolled past.'
The training signal
How performance feeds the model.
The performance loop is not abstract. Every execution surfaced on this page has been labeled: approved, rejected, or edited. Those labels are the training signal.
Labels at /admin/generations
Per-execution Approve / Reject / Edit on every output. The label is captured at /admin/generations the moment review happens.
Phase 2V2 classifier
Labels train a per-brand classifier that scores future executions against the brand’s accepted voice. Drift caught before it ships.
Phase 3 adapter · gated at 100 labels
Every approved variant auto-promotes into the brand's precedent library. The adaptive per-brand model retrieves top-K precedents at inference time as match-this-voice anchors. Sharpens with every approval.
The result
Top-quartile performance is not just a measurement. It is the input that pulls every future execution toward the top quartile.
Veeva sync · activity log
The data loop, last 24 hours
12 min ago
DEM-2026-04-218 · Caregiver carousel approved & synced to Veeva PromoMats v2
synced
1 hr ago
DEM-2026-04-217 · Reel pending MLR · queue position 4 of 11
in queue
2 hrs ago
DEM-2026-04-216 · Static flagged · auto-rewrite suggested
needs review
3 hrs ago
Performance pull · 142 Compound DT-204 variants Q1 ingested as training signal
synced
Yesterday
Veeva CRM HCP engagement export · 2,840 records de-identified, aggregated, fed to engine
synced
Forward thisOne click to the right page for whoever you need to loop in.
What's underneathArchitecture · advisors · published thinking.
Talk to us30 minutes with Maha + Zein. No deck, no sales motion. We answer your questions.