Changelog: Design Decisions
Over Time

Autonomous Enterprise Platform positioning established. The four process pillars — Quote-to-Cash, Finance Operations, Procure-to-Pay, Supply Chain — are formally documented as Platform members, each sharing the ClaraVis anchor client, GCP data fabric, HITL framework, and XAI/SHAP audit layer. The Autonomous Enterprise is the cross-cutting governance layer, not a fifth pillar.

Shared data fabric contract formalised. All four pillar platforms write to a shared BigQuery dataset (ae_governance) via Pub/Sub event streams. The governance layer subscribes to these streams via the topics ae.governance.dashboard.refresh, ae.governance.xai_audit, and ae.governance.hitl.events. Pillar platforms are decoupled from the governance layer by the event bus (ADR-006).

Suite navigation band added to all reference documents. A persistent navigation element on Changelog, Glossary, and ADR Index surfaces direct links to all four pillar sites, reinforcing the Suite relationship without modifying the core AE module page architecture.

Five-panel architecture specified at pillar resolution. Each Strategy Dashboard panel now has an explicit data source assignment: Revenue Pipeline (Quote-to-Cash), Fleet Health (Supply Chain), Financial Risk (Finance Operations), Carbon Emissions (GreenOps, all pillars), Compliance Posture (all). Materialised BigQuery view refresh cycle: 15 minutes, triggered by Pub/Sub.

Drill-through contract formalised. Every dashboard panel cell navigates to the source module's HITL queue or XAI explanation viewer. The Strategy Dashboard is explicitly a navigation layer — not a destination — preventing the common failure mode of dashboards becoming reporting dead-ends.

München cross-module thread made visible at Strategy Dashboard level. Unit MRI-7T-MCH-0042 (Universitätsklinikum München) is the demo anchor for the fleet health panel, with drill-through linking back to the Asset IQ (M-05) HITL queue. This makes the five-module cross-pillar narrative observable in a single interface.

H1 gate constraint formalised. Data Governance (M-08) is the sole Horizon 1 module. ADR-016 (phased adoption) is extended with an explicit dependency: no H2 ML module (ae.ml.scoring.* Cloud Run endpoints) may reach production until M-08 schema validation pipelines pass with a ≥99.5% SLA for 30 consecutive days. This is enforced as a Binary Authorization attestation requirement (ADR-015).

Quarantine-then-review principle extended to all pillar data feeds. The TFX ExampleValidator quarantine pipeline now handles schema-violating records from all four pillar platforms, not only the Asset IQ telemetry feed. Quarantine bucket partitioned by pillar_id, source_system, quarantine_timestamp. Steward reinstatement HITL (HITL-11) SLA: 48 hours.

Feature lineage graph extended to cover Procure-to-Pay and Quote-to-Cash feature groups. Vertex AI Feature Store feature groups fg_qtc_opportunity and fg_p2p_supplier added to the lineage graph. Data Governance dashboard now renders cross-pillar feature-to-model dependency paths.

Hard-deadline bypass enforced at GKE admission controller. The workload annotation ae.greenops/deferral-eligible: "false" is required on all batch jobs that must not be deferred: the daily RUL scoring job (Asset IQ, M-05) and the FinRisk Sentinel model refresh (M-04). The GreenOps scheduler rejects deferral requests for annotated workloads at admission time — not as an application-layer check.

Scope 3 attribution extended to Procure-to-Pay and Supply Chain pillar events. GreenOps now accounts for Scope 3 emissions from supplier logistics events (Procure-to-Pay) and installed unit lifecycle events (Supply Chain), satisfying EU CSRD §29a Category 11 (use of sold products) and Category 1 (purchased goods and services).

CSRD disclosure report generation added. GreenOps generates a downloadable EU CSRD §29a–compliant disclosure on demand from the governance dashboard. Report is rendered as a PDF by a Cloud Run job, signed by Cloud KMS, and written to gs://ae-governance-reports/csrd/ with CMEK encryption.

README.md rebuilt to Google OSS standards. Structured with 14 numbered sections following Google's open-source README conventions: Suite positioning, four-pillar table, governance platform capabilities, anchor client, architecture overview, governance dashboards, module index, navigation, ADR summary, design principles, regulatory posture, München narrative, infrastructure, and about. Semantic versioning badges added.

Changelog upgraded to Keep a Changelog + SemVer format. v2.x releases distinguish Suite-level (MAJOR) from governance layer additions (MINOR) from module-level fixes (PATCH). All prior releases retroactively classified under v1.x.

Glossary extended with 28 Suite-level and governance-specific terms. New sections: Autonomous Suite, Governance Layer, Compliance & Regulatory, Data Platform. All cross-references updated to include pillar site links.

M-07 Strategy Dashboard — Five-panel Looker Studio–backed executive interface: Revenue Pipeline, Fleet Health, Financial Risk, Carbon Emissions, Compliance Posture. BigQuery materialised views. Drill-through to HITL queues. No HITL checkpoint on the dashboard itself — it is a read-only governance surface. München thread anchor: MRI-7T-MCH-0042 visible in fleet panel.

M-08 Data Governance — TFX ExampleValidator pipeline. Vertex AI Feature Store lineage graph. Quarantine-then-review pattern for schema-violating records. Steward reinstatement HITL (HITL-11). The sole H1 module — contractual gate for all H2 ML deployments. ADRs DG-01 through DG-04 covering TFX vs Great Expectations, Feature Store vs manual feature engineering, quarantine vs discard, and steward SLA.

M-06 GreenOps Platform — Carbon Footprint API integration. GKE intelligent job scheduling with hard-deadline bypass for RUL and FinRisk jobs. CSRD Scope 1/2/3 accounting. EU CSRD §29a disclosure generation. No HITL checkpoint — carbon accounting is fully automated; human review reserved for annual CSRD filing sign-off outside this system.

M-04 FinRisk Sentinel — Isolation Forest anomaly detection on BigQuery streaming payment events. Dual-reviewer HITL-08 (CFO + Finance Controller — both must approve for a payment event to be cleared). TreeExplainer-generated feature contribution cards presented at HITL. ADR-011 (Isolation Forest over autoencoder) justified by SHAP TreeExplainer compatibility.

M-05 Asset IQ — RUL regression model (XGBoost) + anomaly detection (Isolation Forest) on 12,000+ unit telemetry. HITL-06 (Field Service Manager, 4h SLA) and HITL-07 (Regional Service Director, 2h SLA for imminent failure). ISO 13485 Device History Record integration. München anchor unit MRI-7T-MCH-0042 reaches 8.4d RUL in the demo pathway.

M-03 RevRec AI — XGBoost ASC 606 / IFRS 15 classification (SINGLE-ELEMENT, MULTI-ELEMENT, LEASE, SERVICE). SHAP explanation with top-5 features. HITL-04 (Finance Controller, 4h) and HITL-05 (VP Finance, 2h for >€500K contracts). SAP write guard enforcing hitl_id parameter. München contract classified as MULTI-ELEMENT, €2.84M.

M-01 CCAI Sales Agent and M-02 ContractGuard — See v1.3.0 and v1.4.0.

ADR Index — All 16 core ADRs + module-specific ADRs (M-01 through M-08). Each ADR with context, decision, alternatives considered, and consequences.

Glossary — One-sentence definitions for all technical terms. Alphabetically ordered, cross-referenced to modules and ADRs.

Changelog — This document. Full version history from concept to portfolio completion.

Seven-section module page template locked: §01 System Context (C4 L1, light bg) → §02 Architecture (dark bg) → §03 State Machine(s) → §04 Data Flow/Sequence → §05 HITL UI mockup → §06 ADRs → §07 Rebuttals (6 stakeholder cards) → §08 Demo Pathway.

Rebuttal format established: six cards per module, one per stakeholder persona, each with a quoted objection and a structured architectural response with evidence. Personas: CFO, CTO, Legal, CISO, Data Science Lead, Field Operations.

Accessibility rules locked: dark text on light background for all pills and badges, WCAG AA minimum contrast, role="img" + aria-label on all functional SVG diagrams.

Quote-to-Cash, Page 06 — ML Engineering & MLOps — Five model cards (RevRec AI, ContractGuard, Asset IQ RUL, Asset IQ Anomaly, FinRisk Sentinel). Vertex AI Pipelines CI/CD. SHAP XAI layer design. Feature Store partitioning. Drift detection thresholds. Eight MLE design decisions covering data augmentation, Isolation Forest contamination parameter, feature store partitioning strategy, SHAP computation placement, and model versioning.

Quote-to-Cash, Page 07 — Infrastructure & GCP — Terraform IaC. Shared VPC (ADR-013). VPC-SC perimeter design. GKE Autopilot + Cloud Run separation of concerns (ADR-003 vs ADR-015). CMEK. BeyondCorp + Workload Identity Federation. Binary Authorization. Chaos Engineering programme (four game days). Runbook for P0 incidents. Seven infrastructure design decisions including Terraform state IAM, state locking, secret rotation cadence, VPC-SC and Salesforce webhook pattern, local apply prevention, module versioning constraints, and Binary Authorization attestation chain.

Page 08 — Go-to-Market Strategy — EU AI Act enforcement deadline as the commercial forcing function. Buyer map (5 personas × 5 columns: pain, trigger, message, proof, risk). Value by Horizon (H1/H2/H3 with "what makes the next horizon easier" logic). Four adoption risk cards. ADR-016 (phased adoption over big-bang).

Page 03 — TOGAF ADM: Phases A–F — Architecture Vision (Phase A), Business Architecture (Phase B), Information Architecture (Phase C), Application Architecture (Phase C), Technology Architecture (Phase D), Migration Planning (Phase E/F). Work products per phase. ADRs 001–006 documented. Horizon roadmap (H1/H2/H3) with migration sequence and EU AI Act deadline as forcing function for H1/H2 sequencing.

Page 04 — Delivery & Product Design — SAFe ART structure (Commercial ART, Financial ART, Operations/Platform ART). PI-1 through PI-8 delivery plan. Nine stakeholder personas with pain points, current state metrics, and success criteria. FRD with 60+ functional requirements. Eleven HITL checkpoints defined: HITL-01 (CCAI turn-11 escalation) through HITL-11 (Data Governance steward reinstatement) — each with actor, SLA, trigger, timeout behaviour, and audit record format.

Quote-to-Cash, Page 05 — Agent Swarm Architecture — ADK multi-agent orchestration. A2A protocol specification (ADR-008). Five agent specifications with formal state machines: CCAI Sales Agent (11-state), ContractGuard (8-state), RevRec AI (7-state), Asset IQ (6-state), FinRisk Sentinel (5-state). Circuit breaker patterns (closed/open/half-open). Firestore state schema design. ADRs 007–009.

HITL-01 through HITL-11 named and formally specified — this framework is the structural spine of the AE Platform's EU AI Act Article 14 compliance posture. HITL checkpoints are state machine nodes, not process steps: they have formal entry conditions, presentation contracts, decision interfaces, timeout behaviours, and immutable audit records.

EU AI Act Article 14 (human oversight) compliance approach defined. All Annex III high-risk modules require a HITL checkpoint. Dual-reviewer requirement (HITL-08) applies to financial anomaly decisions where unilateral approval creates a segregation of duties risk under SOX.

Page 01 — The Autonomous Enterprise — Concept introduction. The four-layer architecture (Presentation, Agent Orchestration, Data/ML Platform, Infrastructure/Governance). Positioning statement: "Salesforce Einstein delivers intelligence within the Salesforce data model. The Autonomous Enterprise delivers intelligence across the enterprise."

Page 02 — ClaraVis: Client Brief & Requirements — Anchor client defined: ClaraVis Medical Systems, €1.2B German MRI/CT OEM, 12,000+ installed units, 34 countries, 4,200 employees. Three pain points established (47-day CPQ cycle, EU AI Act compliance gap, €40M warranty over-reserve). Nine stakeholder personas mapped with role, pain, and success metric.

Anchor client selection. ClaraVis provides a credible enterprise context spanning all three major AI use case categories (commercial, financial, operational/IoT) in a single company, enabling cross-module narrative continuity. Medical device regulatory context (EU AI Act Annex III, ISO 13485, FDA 21 CFR 820) sets the design bar at FAANG-portfolio level.

GCP as sole cloud platform. Simplifies infrastructure design, maximises native integration (ADK, CCAI, Vertex AI, BigQuery), produces a coherent IAM and VPC-SC story. Multi-cloud was considered and rejected — added complexity does not serve the portfolio's demonstration objectives, and a single-cloud architecture is a more credible production design for a regulated EU enterprise of this size.

Static GitHub Pages deployment. The portfolio IS the project — deploying it as a live site makes every design decision immediately verifiable. No build step, no framework, no backend. Deployment is a proof of simplicity, not a limitation.

ClaraVis as a fictional entity. Avoids NDA, competitive intelligence, and client confidentiality constraints. All numbers (€1.2B revenue, 12,000 units, €40M warranty reserve, 47-day CPQ cycle) are calibrated to be plausible for an enterprise of the described scale and sector. No real company's data is represented.