⚠ CONCEPT DEMONSTRATION — ALL DATA IS SYNTHETIC AND ILLUSTRATIVE. NOT REPRESENTATIVE OF ANY REAL SYSTEM OR CUSTOMER. ⚠
Phase 2 Demo — Closed-Loop Intelligence
CONSEQUENCE CLOSED.
From anomaly detection to resolved consequence chain in under 60 seconds. Walk through the full Synestra lifecycle — detect, trace, propose, authorize, close.
1
Detected
2
Traced
3
Proposed
4
Authorized
✓
Closed
Phase 01 — Anomaly Detection
Anomaly detected.
T+0:00 · PHX-A Campus · Hall B Row 4
Thermal Anomaly · Warning
CDU-B4 supply air temperature exceeding safe operating threshold
Current temp
44.1°C
30d baseline
38.5°C
Delta
+14.5%
Rate of change
+0.6°C/min
Source signals
← Phaidra · CDU-B4 thermal advisory · event_id ph_9821 · 44.1°C supply
← BACnet · CRAC-B4 supply air sensor · 44.08°C · corroborating
← DCGM · H100-B4 cluster · GPU junction temp rising · 78→84°C
Phase 02 — Consequence Tracing
3-hop chain. 89% confidence.
T+0:00.8 · Chain traced in 800ms
HOP 1
THERMAL
CDU-B4 supply air +14.5% above baseline. Rate of change indicates CRAC coil restriction or refrigerant pressure drop. Not a transient spike.
Source: Phaidra thermal advisory + BACnet corroboration · confidence 94%
↓
HOP 2
COMPUTE
H100-B4 cluster GPU junction temp: 78→84°C and rising. Thermal throttle threshold is 87°C. At current rate, throttle fires in 9 seconds. Predicted throughput loss: 18%.
Source: NVIDIA DCGM · 12 GPUs at risk · H100-B4 rows 12–14
↓
HOP 3
POWER
GPU throttle will reduce PDU-B4 draw by ~42 kW. That headroom cannot be reallocated while the thermal constraint is active — it becomes stranded capacity. Not a failure, but 42 kW of paid-for power is effectively offline.
Source: Power model · PDU-B4 headroom blocked · $1,260/hr at $0.03/kWh compute rate
Consequence score
0.89
Cross-domain impact
Thermal → Compute → Power
Time to consequence
9 seconds
kW at risk
42 kW
Phase 03 — Intervention Proposed
Two actions. 42 kW recovered.
T+0:00.8 · Proposal generated · awaiting authorization
Recommended Intervention
01
Adjust CRAC-B4 setpoint: 72°F → 68°F
Reduces CDU-B4 supply air temp by estimated 5–6°C over 90 seconds. Addresses thermal root cause. No workload impact during adjustment.
02
Shift compute workload: H100-B4 rows 12–14 → H100-C1 rows 1–3
Moves at-risk jobs to cooler cluster zone while thermal correction propagates. Prevents GPU throttle from firing. Jobs resume at full speed on target cluster within 30 seconds.
kW recovered
42 kW
Thermal normalized
~90s
Workload impact
None
EA score recovery
+1.6%
This proposal was generated automatically. No human identified this chain. Synestra traced 3 domains, identified the intervention, and estimated the outcome — all within 800ms of the triggering event arriving at the ingest API.
Phase 04 — Authorization
Manual or auto. Your call.
Manual Mode
Operator reviews and authorizes each intervention. Synestra proposes; humans decide. Recommended during Phase 1 onboarding.
Auto Mode
Synestra executes interventions automatically within pre-approved action boundaries. Phase 2 capability. Site-configurable per action type and severity.
Executes both recommended actions · logs to ServiceNow + PagerDuty · outcome tracked against 30-day baseline
✓
CONSEQUENCE CLOSED
PHX-A · Hall B Row 4 · chain cc_7q9rx · opened 14:22:11 · closed 14:22:58
Time to close
0:47
from detection to resolved
kW recovered
42 kW
headroom restored to PDU-B4
EA score
+1.6%
94.2% → 95.8%
GPU throttle
Avoided
H100-B4 cluster at full speed
Execution log
14:22:11DETECTCDU-B4 thermal anomaly received · source: Phaidra + BACnet
14:22:11TRACE3-hop consequence chain constructed · 800ms · confidence 89%
14:22:12PROPOSEIntervention proposed · CRAC-B4 setpoint + workload shift
14:22:14NOTIFYPagerDuty alert created · P3 · chain cc_7q9rx · 42kW at risk
14:22:58EXECUTECRAC-B4 setpoint → 68°F · API call confirmed · BACnet write
14:22:59EXECUTEWorkload migration H100-B4 → H100-C1 · scheduler API confirmed
14:23:28VERIFYCDU-B4 supply air: 44.1 → 41.2°C · trend normalizing
14:23:58CLOSEDConsequence chain resolved · 42kW recovered · chain duration 0:47
14:23:58REPORTServiceNow INC-84291 updated: Resolved · Palantir dataset write complete
Step 1 of 5