ASIL predicts each viewer’s session intent and adapts the streaming homepage in real time.
Problem
Sessions surface content that doesn’t match viewer intent
An estimated 28% of sessions end without meaningful playback.
Solution
Adaptive Streaming Intelligence Layer (ASIL)
Real-time intent classification that adapts hero selection, row hierarchy, and autoplay to match each session’s context.
Impact
+8.8%
Weekly Engaged Hours per Subscriber (WEH/S)
avg. across confirmed experiments
Modeled baseline4.5 hrs / wk
With ASIL4.9 hrs / wk
Weekly Engaged Hours per Subscriber (WEH/S)+8.8% lift
Modeled baseline4.5 → ASIL4.9 hrs / wk+8.8% WEH/S
How It Works
🔌
Session Signals
Device · Time · Recency · Scroll
Captures context within the first 3 interactions of a session before any content decision is made.
→
↓
🤖
Intent Model
Classifies session in <12ms
Weights signals into one of five intent archetypes: binge, comfort, quick-break, re-entry, or discovery.
→
↓
🏠
Orchestration
Hero · Row order · Autoplay
Reconfigures the homepage surface in real time to match the classified intent before the subscriber sees it.
→
↓
📈
Experiment & Adapt
WEH/S tested via A/B framework
Experiment outcomes continuously retrain and calibrate the model, compounding accuracy over time.
Session Intent Clusters
A subscriber’s viewing goal can change from one session to the next. ASIL classifies each session into one of five probabilistic intent archetypes within seconds of platform entry.
Continuous guardrail monitoring at every ramp gate
System Architecture & Intent Modeling
Real-time session signals in, intent-aware homepage out. ASIL sits between the platform’s recommendation engine and presentation layer to classify intent, orchestrate the response, and refine through continuous experimentation.
Model confidence scores across all 5 intent clusters. Session intent type is classified as the highest-fit cluster, driving all downstream orchestration. Example session: TV · 8pm · 1-day recency · 3 eps this week
A/B Test OutcomesCluster Delta AnalysisGuardrail MonitoringModel RetrainingGround Truth RefinementConfidence Score Drift
Inference Latency
12ms
P95 · Async pipeline
Input Features
9
Signal ingestion layer
Intent Clusters
5
Probabilistic · Exhaustive
Classification Accuracy
84%
Held-out test set
Design Decisions
Behavioral Clustering
The five intent archetypes were derived from session behavioral patterns rather than defined top-down. Cluster count and separation were validated in Phase 1 silent mode, confirming significant behavioral variance across all five groups before any live UI changes. Clusters are probabilistic and exhaustive: every session receives a score across all five, and the model is retrained as ground truth accumulates from experiment outcomes.
Device as Probabilistic Signal
Device type informs, but does not determine, intent classification. A TV session correlates with lean-back patterns in aggregate but not at the individual session level. Treating device as a rule would misclassify a material fraction of sessions.
Asynchronous Inference & Graceful Fallback
The model runs concurrently with homepage render. The UI surfaces a static default configuration while inference completes, updating within a 50ms threshold. If inference fails or times out, the homepage degrades to the standard non-personalized layout with no subscriber-visible disruption.
Presentation Layer Only
ASIL orchestrates how recommendations are presented, not what is recommended. Keeping the long-horizon taste model unchanged is a deliberate boundary: it constrains scope, reduces execution risk, and enables clean A/B attribution by isolating the surface variable.
Explicit Non-Goals (v1)
✕Mid-session or in-playback personalization (ASIL orchestrates the homepage at session entry only)
✕Subscriber segmentation or CRM targeting (ASIL is session-level, not subscriber-profile-level)
✕Full visual redesign or navigation overhaul (ASIL changes content orchestration, not the surrounding interface)
✕Replacement of the long-horizon recommendation engine
✕Cross-service identity, entitlement, or catalog integration
✕Ad-tier monetization optimization
Metrics Tree Hierarchy
How subscriber behavior breaks down into measurable layers, from session-level discovery and depth signals up through weekly supporting metrics to Weekly Engaged Hours per Subscriber, the single north star ASIL is designed to move.
North Star
Supporting Metric
Session Signal
North Star
4.5h
Weekly Engaged Hours / Subscriber
Total hrs watched ÷ Active subscribers (weekly) · Modeled baseline
7-Day Return Rate
58%
Sessions / Week
5.5
Minutes / Session
49m
Avg Session Depth
54%
Time to First Play
68s
Session Abandonment
28%
Hero CTR by Intent
34%
Completion Rate
54%
Supporting Metrics
7-Day Return Rate
Habit
58%
Share of subscribers who open the app within 7 days of a given session. A subscriber who returns within the week is substantially less likely to churn.
Sessions / Week
Habit
5.5
Number of distinct viewing sessions a subscriber starts in a rolling 7-day window. Higher frequency means the app is part of a weekly routine, not an occasional visit.
Minutes / Session
Depth
49m
Total time watched per session, averaged across all sessions. Longer sessions mean the subscriber found content worth staying for.
Avg Session Depth
Depth
54%
Share of a content item watched per session started. Rises when orchestration matches content length to available session time.
Session Signals
Time to First Play
Discovery
68s▼ lower is better
Seconds from app open to first content playback. Drops when the right content is surfaced immediately without requiring scrolling or search.
Session Abandonment
Discovery
28%▼ lower is better
Share of sessions that end with no playback initiated. The primary signal of intent-surface mismatch, named as an explicit failure mode rather than measured as a gap in another metric.
Hero CTR by Intent
Discovery
34%
Click-through rate on the hero banner, broken down by classified intent cluster. The only metric that directly validates whether the right content is surfaced for each intent type.
Completion Rate
Depth
54%
Share of a content item watched before the subscriber exits or switches titles. Indicates whether content length matched the subscriber's available session time.
Experimentation Roadmap
Three independent Phase 2 experiments targeting hero configuration, homepage density, and autoplay behavior. Mutual exclusion enforced across all treatment groups; Return After Gap and Discovery Mode interventions are planned for future validation.
Experiment
Hypothesis
Primary Measure
Secondary
Cohort Alloc.
Decision
Exp A: Context-Aware Hero
Lift: +6.7% · 96.4% conf.
Aligning hero runtime and format to inferred session intent increases WEH/S and reduces time to first play.
Δ WEH/S over 4 weeks
Playback start rate, abandonment rate
20%
✓ Go
Exp B: Homepage Compression
Lift: +3.4% · 88.1% conf.
Reducing homepage density for short-intent sessions decreases browse friction and increases completion rate.
Completion rate
Session frequency, scroll depth, minutes/session
15%
⌛ Watch
Exp C: Binge Acceleration
Lift: +10.8% · 98.2% conf.
Lean-back acceleration increases episodes per session and series completion, strengthening 7-day return rate.
Episodes per session
7-day return rate, series completion, churn delta
20%
✓ Go
Experiment A
Context-Aware Hero Module
WEH/S Lift
+6.7%
Confidence
96.4%
p-value
0.036
Sample
142K
● Significant
Session Metrics — Baseline vs. Treatment
WEH/S Delta▲ +6.7%
Baseline: 4.5 hrs/wkTreat: 4.8 hrs/wk
Abandonment Rate▼ 20.0%
Baseline: 30%Treat: 24%
Time to First Play▼ 25.7%
Baseline: 19.1sTreat: 14.2s
Analysis & Decision
✓ Decision
Go: include in Phase 3 multi-intent rollout
WEH/S divergence crosses the 95% threshold at week 8. Abandonment and time-to-play improvements confirm the mechanism: hero alignment reduces friction at session entry.
Experiment B
Homepage Compression (Short Intent)
WEH/S Lift
+3.4%
Confidence
88.1%
p-value
0.119
Sample
98K
Trending
Session Metrics — Baseline vs. Treatment
Completion Rate▲ 18.8%
Baseline: 48%Treat: 57%
Session Frequency▲ +6.0%
Baseline: 5.5/wkTreat: 5.8/wk
Abandonment Rate▼ 13.3%
Baseline: 30%Treat: 26%
Analysis & Decision
⌛ Decision
Watch: extend run window, re-evaluate at 95% threshold
Confidence plateaus at 88.1%, below threshold. All session metrics are directionally positive; completion rate is the strongest signal. Re-evaluate at 90% confidence given consistent direction across the full run window.
Experiment C
Binge Acceleration (Lean-Back)
WEH/S Lift
+10.8%
Confidence
98.2%
p-value
0.018
Sample
187K
● Significant
Session Metrics — Baseline vs. Treatment
Episodes per Session▲ +18.0%
Baseline: 2.1Treat: 2.5
7-Day Return Rate▲ +11.0%
Baseline: 62%Treat: 69%
Completion Rate▲ 17.5%
Baseline: 63%Treat: 74%
Analysis & Decision
✓ Decision
Go: include in Phase 3 multi-intent rollout
98.2% confidence and +10.8% WEH/S lift at week 7. Episodes per session and 7-day return both show clear separation: autoplay friction reduction compounds within the session and carries into the following week.
Experiment Decision Simulator
Replay the confidence trajectory for each experiment. Drag to any week to see where each stood, and when Exp A and C crossed the go/no-go threshold.
Exp A: Context-Aware Hero
⌛ Watch
Week 5of 12
Conf. threshold: 95%83.4%
Sample Size89K
WEH/S Lift+3.8%
Exp B: Homepage Compression
Insufficient data
Week 6of 12
Conf. threshold: 95%76.6%
Sample Size59K
WEH/S Lift+1.9%
Exp C: Binge Acceleration
⌛ Watch
Week 5of 12
Conf. threshold: 95%88.1%
Sample Size134K
WEH/S Lift+7.6%
Phase 3: Multi-Intent Scaling
Gradual traffic ramp pending Exp B resolution. Each gate requires explicit sign-off against WEH/S lift and guardrail thresholds before the next stage unlocks.
1
Shadow
5%
Combined multi-intent model scoring at scale, no UI changes
Gates: No latency regression vs. baseline, intent distribution within expected range
Gates: +Δ WEH/S at 95% conf., all guardrails within Phase 2 thresholds
3
Multi-Intent Rollout
25%
All Phase 2 validated clusters active simultaneously
Gates: Composite WEH/S lift holds at 25%, all guardrails within threshold
4
Geo Expansion
50%+
Sequential market rollout, validating cluster signal transfer across markets
Gates: Consistent WEH/S signal across tier-1 markets, no cluster collapse in non-US cohorts
Guardrail Monitoring
ASIL monitoring covers seven non-negotiable metrics spanning performance, stability, content health, and subscriber experience. Any single breach automatically reverts all experiment traffic to the control condition pending root cause investigation.
Current Status · All Metrics Live · Q3 2026
App Latency P95/P99
✓ OK
312ms / 487ms
+4ms / +6ms vs. baseline
Threshold: < 400ms / 600ms
Async inference adds negligible overhead. Both percentiles tracked; stable vs. Phase 1 baseline.
Crash Rate
✓ OK
0.08%
▼ −0.01% vs. baseline
Threshold: < 0.20%
Marginal improvement vs. Phase 1 baseline of 0.09%. Platform stability unaffected by orchestration layer.
Playback Error Rate
✓ OK
0.31%
+0.02% vs. baseline
Threshold: < 1.0%
Negligible uptick from baseline 0.29%. Within measurement noise; orchestration not introducing delivery issues.
Autoplay Skip Rate
✓ OK
18.4%
▼ −1.8% vs. baseline
Threshold: < 25%
Improving from baseline 20.2%. Subscribers accepting autoplay at higher rates, confirming lean-back classification accuracy.
Content Diversity Index
✓ OK
94.2
−0.4 vs. baseline
Threshold: > 85
Negligible dip from baseline 94.6. Well above the filter bubble floor; catalog breadth unaffected.
Support Tickets (Nav)
⚠ Monitor
+2.1%
▲ +2.1% vs. baseline
Threshold: < +5.0%
Largest mover across all guardrails. Likely reflects user surprise at homepage reordering; under active investigation.
7-Day Churn Delta
✓ OK
−0.3%
▼ −0.3% vs. baseline
Threshold: < +0.5%
Phase 1 baseline was flat. Phase 2 showing improvement; ASIL is reducing churn, not just shifting engagement.
Risk Analysis & Mitigation
Identified risks ranked by severity. High risks have dedicated guardrail coverage; all mitigations were built into the ASIL design.
Model Misclassification at Scale
High
Wrong intent classification degrades the session for the misclassified population; ASIL becomes a friction source.
Mitigation
Silent Phase 1 validation and per-cluster guardrail monitoring throughout rollout.
Latency Regression
High
Inference overhead increases time-to-first-play, the direct inverse of ASIL's primary goal.
Mitigation
Async inference with static fallback. P95/P99 guardrails halt on first regression.
Orchestration Scope Creep
Medium
Hero and row changes can be perceived as a UI redesign, creating stakeholder misalignment.
Mitigation
Presentation-layer-only scope enforced. ASIL changes content, not surface design.
Content Diversity Narrowing
Medium
Aggressive short-intent orchestration could reduce catalog exposure over time.
Mitigation
Content Diversity Index enforced as an explicit guardrail with per-session-type floors.
A/B Experiment Interference
Medium
Three simultaneous experiments on related surfaces risk contamination and noisy attribution.
Mitigation
Mutual exclusion enforced. Sequential rollout available as fallback if contamination detected.
Cold Start: New Subscribers
Low
New subscribers lack behavioral signal and may receive poor intent classifications in early sessions.
Mitigation
Static orchestration default for sessions 1-5. Explicit cold-start handling in model spec.
Guardrail Breach Simulator
Simulate what happens when a guardrail metric crosses the breach threshold. The response sequence below reflects the live monitoring protocol.
Support Tickets — Navigation Category
Currently +2.1% vs. Phase 1 baseline, the only guardrail metric trending above baseline
Detection
5-min polling
PagerDuty alert <60s
Response
Auto-halt to control
Within 5 min, no manual step
Re-ramp
Dual sign-off required
PM + data science lead
Support Ticket Delta+2.1%
0%Threshold: +5.0%+10%
🛑ROLLOUT HALT TRIGGERED
Support Tickets (Navigation) has breached the +5.0% threshold at +6.3%. Per ASIL rollout protocol, all Phase 2 traffic has been reverted to the control condition pending investigation. No further ramp will occur until root cause is identified and signed off.
Traffic halted
✕ Exp A: Context-Aware Hero
✕ Exp B: Homepage Compression
✕ Exp C: Binge Acceleration
Protocol triggered
⚠ Breach at 1.3× above threshold
⚠ On-call PM & eng lead notified
⚠ Re-ramp blocked pending root cause sign-off
Session Intent Simulator
Five probabilistic session archetypes, each triggering a distinct homepage configuration. Select an intent type to see how ASIL classifies the session, what it changes on the homepage, and the observed metric impact.
Classification Signals Input
Input signals that determined this classification
→
↓
Homepage Orchestration Process
Lean-Back Binge
→
↓
Observed Impact Output
vs. Phase 1 baseline
Homepage Output Simulated
How a streaming homepage would be orchestrated for this session intent