AUGURUS

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.

OI Domain-active

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:

healthy degraded critical

Separate-repo integration. Core IS families not in primary PE deployment.

OI Domain-active

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:

stable degraded failure

Separate-repo integration. Active in PE cascade as downstream receiver from SC and QS.

HS Domain-active

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:

effective impaired critical
ES Domain-active

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:

nominal stressed critical

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 Separate repo

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 Domain-active

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 Primary PE family

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:

stocked at_risk disrupted
OS Domain-active

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 Primary PE family

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:

conforming degraded critical
RC Primary PE family

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:

compliant at_risk breach
SI Domain-active

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 Primary PE family

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:

liquid stressed distressed
TI Domain-active

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 Separate repo

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

Separate-repo integration. Not in primary PE deployment.

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.

IP boundary: AIP contributes 7 stages (77.8% of the pipeline). PCI adds COMBINE and EXECUTE stages (22.2%) for composed products that orchestrate multiple families. The IP boundary is clean. The cascade inference lives in the VALIDATE→COMBINE link — AIP's VALIDATE stage emits the CalibrationSignal; PCI's COMBINE stage consumes it and drives the cross-family propagation logic.

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 update

mean += α × (observation − mean)

Variance update

var = (1 − α) × (var + α × diff²)

Z-score

z = (observation − mean) / max(std, MIN_STD_FLOOR)

Detection thresholds

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
The PE cascade in numbers: SC degradation → QS propagation in 4–6 days (similarity 0.88). QS → OI in 2–4 days (0.82). OI → FS in 4–6 days (0.76). Total cascade time: 10–16 days from initiating supply chain event to financial sentinel flag. GP is informed within hours of FS crossing the alert threshold — 13–21 days before the portco's next quarterly report.

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

  1. IngestRunner.run_once()
    Fetches all portco data from configured sources — webhooks, REST endpoints, NetSuite, CSV drops.
  2. PortcoPipeline.run()
    Per-company, per-family pipeline execution across all 7 AIP stages. Parallel execution per portco; families within a company execute in cascade order.
  3. signal_aggregator.aggregate()
    Normalises across families, applies worst-state-wins logic at the portfolio company level, and generates the bridge estimate for reserve monitoring.
  4. FiduciaryMonitor.evaluate()
    Maps contractual obligations triggered by state transitions. Sources: LP side letters, co-investor agreements, fund governance documents, regulatory filing calendars.
  5. TaskCreator.create_tasks()
    Generates draft notifications for each mapped obligation, computes statutory deadlines, assigns priority, and queues for GP review. Drafts include pre-populated counterparty, subject, and body text.
  6. ReserveMonitor.snapshot()
    Tracks follow-on pool utilization and flags bridge demand. Surfaced in the GP dashboard as reserve runway alongside portco health verdicts.
  7. portfolio_report()
    Aggregates the GP-level fund health view: portco verdicts, open obligations, cascade events, reserve utilization, and queued actions.
  8. Review queue
    GP approve / reject / defer interface. Approved tasks proceed to delivery. Rejected tasks are logged with reason. Deferred tasks re-surface at the next cycle with escalating priority.
  9. delivery.deliver_approved()
    File-drop to designated secure folder plus optional SMTP delivery. Nothing is sent without explicit GP approval. Delivery receipts are logged and surfaced in the review queue.

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.

Task de-duplication: A company crossing from watch to alert and back again within a reporting cycle does not generate duplicate obligations. The signal_store tracks state history per company. The FiduciaryMonitor suppresses re-notification on the same state within the de-dup window — configurable per obligation type, defaulting to one reporting period. Re-escalation on a sustained state is a distinct event with its own de-dup window.

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.