Architecture Decision Records

Formal audit of Build 025 before Build 026 feature work. Identified RISK-4 (ingestion_runs not persisted) and RISK-6 (hydration mismatch). Both resolved in Build 025.5.

Introduces the architecture governance layer: domain registry, protected systems list, migration registry, build registry, ADR index, dependency map, and governance dashboard. Establishes release-state model and protected-system policy.

Skills extraction (80+ taxonomy), vertical classifier (11 verticals), freshness decay, dedup engine, trend snapshots, company hiring profiles, and admin source panel.

Live ATS ingestion from Greenhouse, Lever, and Ashby via typed provider adapters. Ingestion orchestrator, source health tracking, and manual trigger API.

Establishes the /jobs route as a dedicated robotics jobs board. Initial 60-record seed. WorkforceScout brand.

Establishes Supabase as the sole production database for EAR, providing PostgreSQL, real-time, and auth in one hosted platform.

Adopts n8n for scheduling and orchestrating ingestion workflows, allowing no-code workflow authoring over the API layer.

Defines the signal approval workflow: all AI-generated content passes through human operator review before publication.

Keyword-based forum signal classification taxonomy for robotics topics.

Establishes the EAR ontology as a protected layer. Defines the process for adding, modifying, or deprecating canonical terms.

All AI-generated content on EAR surfaces must carry a disclosure label. Defines disclosure UI patterns.

Numeric credibility scoring model for news and data sources. Score decays on failures and resets on successful ingestion.

RBAC model: visitor / member / operator / admin tiers with gated admin routes.

Append-only entity_timeline_events table for all company and market events. Events are never deleted.

Defines the operator session mode: elevated admin capabilities with audit logging for all destructive actions.

Foundational ADR for the robotics jobs intelligence pillar. Establishes the WorkforceScout brand and data model.

Ingestion can operate in demo (seed data) or live (API) modes. Mode is controlled by environment flag.

Multiple AI providers contribute independent confidence scores to program ratings. Final rating is an ensemble.

Per-vertical specialist intelligence agents provide deeper analysis within each robotics category.

EAR is the OS layer. Multiple brand skins (EAR, TDB, etc.) run on top of shared intelligence infrastructure.

Canonical entity model for companies, people, and products. Entity resolution and dedup strategy.

Revised entity intelligence model with expanded relationship graph and signal-to-entity attribution.

Self-scheduling ingestion pipeline with adaptive frequency based on source activity levels.

Consolidates multiple ingestion pipelines into a single orchestrator. Introduces the operator review queue.