Technical Architecture
The instrument, in full.
14 signal families. 4 clades. A 7-stage pipeline. Isomorphism-based cross-domain cascade inference. A governance scheduler that maps fiduciary obligations and generates drafts without analyst intervention.
14 Signal families
Four clades. One taxonomy.
Augurus is built on a typed taxonomy of operational signal families, grouped into four functional clades. Each family shares the same 7-stage pipeline, the same EWMA-based deviation detection, and the same 3-state vocabulary. What differs is the domain, the metrics, and the clade-role it plays in a cascade.
Monitor clade — continuous state surveillance → alert/escalation
Monitor
Continuous observation of operational, human, and environmental state. Monitor clade families provide the baseline signal that downstream clades depend on. When a monitor family transitions, it propagates upstream to sentinel and quality-gate families through the cascade graph.
IS — Infrastructure State
Monitors core compute and network infrastructure health. Detects degradation in the underlying systems that operational workloads depend on.
Key metrics:cpu_utilization, memory_pressure, network_error_rate
State vocabulary:
OI — Operational Intelligence
Monitors process throughput, error rates, and cycle time variance across production and operational systems. The principal bridge between physical operations and financial outcomes.
Key metrics:throughput_rate, error_rate, cycle_time_variance
State vocabulary:
HS — Human State
Monitors workforce and leadership cognitive and functional capacity. Critical for companies where key-person dependency or workforce fatigue creates operational fragility.
Key metrics:hrv, cognitive_load_score, functional_status_score
State vocabulary:
ES — Environmental State
Monitors external environmental stressors that create operational risk — resource availability, facility conditions, and environmental anomaly rates that precede operational degradation.
Key metrics:environmental_stress_index, resource_deficit_fraction, anomaly_rate
State vocabulary:
Optimizer clade — flow/throughput optimization → dispatch/advisory
Optimizer
Monitors and optimizes flow and throughput across supply chains, energy grids, logistics networks, and organisational capacity. Optimizer clade families are the primary initiators of the PE cascade pattern — supply chain deviations here propagate to quality gate and then sentinel clades.
MO — Movement Optimisation
Monitors physical and logistical flow — blockages, transit delays, and flow rate deviations across transportation and distribution networks.
Key metrics:blockage_fraction, flow_rate, transit_delay_fraction
EN — Energy Navigator
Monitors power grid and energy system health — load factors, frequency deviations, and generation balance. Critical for manufacturing-intensive portcos with high energy dependency.
Key metrics:load_factor, frequency_deviation, generation_balance_delta
SC — Supply Chain
The primary initiator in the PE cascade pattern. Monitors supplier health, inventory depth, and lead time variance. A SC deviation at −2.8σ or greater triggers the full QS→OI→FS→RC cascade within 6–16 days in validated scenarios.
Key metrics:inventory_days, lead_time_variance, supplier_health_score
State vocabulary:
OS — Organisational State
Monitors organisational capacity, headcount attrition, and process cycle time. Bridges human state deviations to operational throughput impact.
Key metrics:capacity_utilization, headcount_attrition_rate, process_cycle_time_variance
Quality Gate clade — conformance/integrity threshold → disposition/restrict
Quality Gate
Monitors conformance, integrity, and regulatory compliance thresholds. Quality gate families are the middle stage of the primary PE cascade — they receive from optimizer families and propagate to sentinel families. QS deviation following an SC event is the first observable consequence on the production floor.
QS — Quality Sentinel
Monitors defect rates, yield, and rework fractions. The first production-floor signal to respond to a supply chain event. QS degradation typically follows SC deviation by 4–6 days in hardware and manufacturing portcos.
Key metrics:defect_rate, yield_rate, rework_fraction
State vocabulary:
RC — Regulatory Compliance
Monitors regulatory conformance scores, open audit findings, and compliance signal rates. In PE portfolios, RC is the terminal cascade receiver — a QS or OI deviation that persists for more than one reporting cycle typically surfaces as an RC finding.
Key metrics:conformance_score, open_findings_count, audit_signal_rate
State vocabulary:
SI — Structural Integrity
Monitors physical asset and infrastructure structural health — strain, fatigue cycles, and structural anomaly rates. Critical for heavy manufacturing, aerospace, and civil infrastructure portcos.
Key metrics:strain_index, fatigue_cycle_fraction, structural_anomaly_rate
Sentinel clade — threshold/breach detection → contain/capital action
Sentinel
Monitors financial solvency, threat vectors, and production constraint thresholds. Sentinel families are the terminal layer of the cascade — they receive propagated signals from optimizer and quality gate families and trigger governance actions, capital alerts, and fiduciary obligations. FS is the primary sentinel family in the PE deployment.
FS — Financial Sentinel
Monitors liquidity runway, operating cashflow, and accounts receivable aging. The financial expression of the operational cascade — when SC→QS→OI degrade, FS flags covenant proximity and bridge demand, typically 16–21 days after the initiating supply chain event.
Key metrics:runway_months, operating_cashflow_ratio, ar_days
State vocabulary:
TI — Threat Intelligence
Monitors behavioral anomalies, lateral movement patterns, and threat confidence scores across the operational environment. Sentinel-clade positioning means TI deviations trigger contain and capital-action recommendations directly.
Key metrics:behavioral_anomaly_score, lateral_movement_rate, threat_confidence
PC — Production Constraint
Monitors bottleneck utilization, constraint fraction, and queue depth. Identifies where production constraint limits throughput and where a constraint shift will amplify an existing OI degradation.
Key metrics:constraint_utilization, bottleneck_fraction, queue_depth_ratio
7-stage pipeline
OBSERVE → VALIDATE. Seven stages. All families.
Every signal family runs the same 7-stage pipeline on every evaluation cycle. The stages are fixed. The family-specific logic lives in each stage's NORMALIZE and DETECT_STATE implementations — everything else is shared infrastructure.
Stage 1
OBSERVE
Ingests raw metric readings from the configured source — webhook, REST, NetSuite, CSV. Observation adapters are family-specific but produce a typed Reading object with a consistent schema: metric name, raw value, timestamp, and source identifier. The adapter layer is the only family-specific code at this stage.
Stage 2
NORMALIZE
Converts raw readings to dimensionless signals suitable for EWMA processing. Handles unit normalization (percentages to fractions, absolute values to ratios), missing-value imputation using last-good-value carry-forward, and warmup flagging. Confidence scores are suppressed during the warmup period (first 10 observations per metric).
Stage 3
DETECT_STATE
Applies EWMA baseline tracking (α = 0.10) and z-score deviation detection. AT_RISK fires at ≥ 2.0σ in the operationally bad direction. ALERT/CRITICAL fires at ≥ 3.5σ. Raw threshold safety nets catch extreme deviations regardless of baseline state. Confidence score is derived from coefficient of variation (CV) post-warmup — low CV (stable baseline) yields high confidence scores.
Stage 4
DETECT_TRANSITION
Compares current state to the prior state held in the signal store. Emits a transition event when state changes — nominal→watch, watch→alert, alert→nominal. De-duplication logic prevents re-firing on the same state within the de-dup window, even when the triggering z-score remains elevated. Only genuine state transitions produce downstream events.
Stage 5
ESTIMATE_OUTCOME
Projects the financial and operational consequences of the detected state. For FS: covenant breach probability and runway compression timeline. For SC: cascade propagation estimate — which downstream families are likely to receive a TransferProposal and in what timeframe. For QS: yield loss rate and rework cost estimate. Outcome estimates are carried into the governance scheduler for obligation mapping.
Stage 6
RECOMMEND
Produces a typed recommendation with a reason code, a recommended action, the required approval tier (GP / analyst / automated), and a HALT path for governance conditions not met. HALT prevents recommendation delivery when pre-conditions are unmet — for example, when a co-investor notification requires confirmed receipt of a prior alert that has not yet been acknowledged.
Stage 7
VALIDATE
Emits a CalibrationSignal when a golden-deviation fires (≥ 2.0σ), carrying the coefficient of variation (CV) and z_severity. This is the signal that triggers cross-domain cascade propagation. The CalibrationSignal is consumed by the isomorphism engine, which computes structural similarity scores to all other active families and issues TransferProposals to families above the propagation threshold.
Signal algorithm
Adaptive baselines. Directional z-scores.
Each metric maintains an exponentially weighted moving average (EWMA) baseline. α = 0.10 weights recent observations more heavily than historical data, allowing the baseline to adapt to a company's evolving operational state without resetting after each reading. The result is a baseline that reflects where the company actually operates, not where it was when the platform was first configured.
Update equations
mean += α × (observation − mean)
var = (1 − α) × (var + α × diff²)
z = (observation − mean) / max(std, MIN_STD_FLOOR)
z ≥ 2.0σ (bad direction) → AT_RISK
z ≥ 3.5σ (bad direction) → ALERT
Design decisions
Minimum standard deviation floor
The MIN_STD floor prevents zero-variance collapse on companies with very stable baselines — a common failure mode when monitoring healthy portcos that suddenly experience an event. Without the floor, a sustained period of stability drives std toward zero, causing the first genuine deviation to produce an astronomically high z-score and triggering alert states that cannot be calibrated.
Directional detection
Directional z-scoring is critical: only deviations in the operationally bad direction trigger flags. A supplier health score improving from 0.82 to 0.96 does not fire an alert — only deterioration from 0.92 to 0.74 does. Each family defines the bad direction per metric. For runway_months, below baseline is bad. For defect_rate, above baseline is bad. The directionality is encoded in each family's DETECT_STATE implementation.
α = 0.10 tuning
An α of 0.10 balances responsiveness and stability. Higher α values (e.g. 0.3) react quickly to transient events but produce noisy baselines and frequent false positives. Lower α values (e.g. 0.05) are too slow to adapt to genuine structural shifts in a company's operational posture. The 0.10 value was selected based on observed cycle times in PE portco data — a 10-day half-life that captures genuine operational trends without chasing noise.
Cross-domain cascade
How a SC signal becomes a FS warning.
When a CalibrationSignal fires in one family, Augurus computes its structural similarity to all other active families using Jaccard-based isomorphism scoring across the 8-rank taxonomic position of each signal species. Families with similarity above the propagation threshold (0.30) receive a TransferProposal — a confidence-weighted invitation to update their own baseline in light of the upstream deviation.
The transfer is clade-gated: proposals cannot create amplification loops (A→B→A). Within a clade, transfers flow in the optimizer → quality → sentinel direction. Cross-clade transfers follow the cascade direction of the business event. The cascade graph is acyclic by construction.
| Source family | Target family | Similarity score | Cascade direction |
|---|---|---|---|
| SC Supply Chain | QS Quality Sentinel | 0.88 | Optimizer → Quality Gate |
| QS Quality Sentinel | OI Ops Intelligence | 0.82 | Quality Gate → Monitor |
| OI Ops Intelligence | FS Financial Sentinel | 0.76 | Monitor → Sentinel |
| OI Ops Intelligence | PC Production Constraint | 0.74 | Monitor → Sentinel |
| OI Ops Intelligence | RC Regulatory Compliance | 0.65 | Monitor → Quality Gate |
| FS Financial Sentinel | RC Regulatory Compliance | 0.61 | Sentinel → Quality Gate |
Fiduciary automation
Signal → obligation → draft. One automated cycle.
The governance scheduler runs as a post-evaluate hook on every pipeline cycle. It monitors company state transitions, maps triggered obligations, computes deadlines, generates draft notifications, and queues them for GP review. No analyst intervention is required between signal detection and draft generation — the GP's only action is approve, reject, or defer.
The scheduler loop
-
IngestRunner.run_once()
-
PortcoPipeline.run()
-
signal_aggregator.aggregate()
-
FiduciaryMonitor.evaluate()
-
TaskCreator.create_tasks()
-
ReserveMonitor.snapshot()
-
portfolio_report()
-
Review queue
-
delivery.deliver_approved()
What gets generated
Co-investor notifications
Side letter obligations to notify co-investors when a portco crosses a specified operational or financial threshold. Draft includes counterparty, threshold description, current metric values, and GP commentary field.
LP notice obligations
Fund document obligations to notify LPs of material events — covenant breach, key-person departure, material litigation. Deadline computed from event date per fund document. GP approves the draft; delivery is logged.
Material Adverse Change counsel
MAC counsel triggers are mapped to specific signal combinations — sustained FS distressed state combined with RC breach, or FS distressed combined with OI failure. Draft is flagged for GP review with legal urgency tag.
ASC 820 re-mark triggers
Fair value re-mark triggers are generated when operational cascade signals suggest a material change in enterprise value that predates the next scheduled valuation cycle. Audit committee notification draft includes supporting signal evidence.
Board meeting agenda items
Cascade events above the alert threshold are automatically queued as agenda items for the next board meeting, with supporting analysis and GP commentary field pre-populated.
See it for your portfolio
We walk through the full pipeline with your portco data — from raw signal to cascade inference to governance draft — in a single 90-minute session.