Release v1.3.3

CHANGELOG.md

@@ -6,6 +6,31 @@ Versioning follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ---
 
+## [v1.3.3] — 2026-04-05
+
+### Added
+
+- **Assignment Score Telemetry (`src/ticket_assigner.py`)**: The ticket assigner now records per-assignment scoring breakdowns (skill, stress, centrality, composite) into a MongoDB `assignment_scores` collection for downstream analysis.
+- **Department Expertise Defaults (`src/utils/persona_utils.py`)**: Added a `DEPARTMENT_EXPERTISE_DEFAULTS` map so personas without explicit expertise lists fall back to sensible department-level defaults rather than a generic "general engineering" placeholder.
+- **Departure Department Field (`config/config.yaml`)**: `org_lifecycle` departure entries now include a `dept` field, making ownership handoff tracking more precise during engineer exits.
+- **Selective Artifact Regeneration (`eval/export_to_hf.py`)**: Added an `--only` CLI flag to selectively regenerate specific artifact types (`nps`, `invoices`, `datadog`) without a full corpus rebuild.
+
+### Changed
+
+- **On-Call Capacity Propagation (`src/ticket_assigner.py`)**: On-call status is now passed dynamically into `_compute_capacity` via an `on_call` parameter instead of being read from a static config key, keeping capacity calculations consistent with the runtime rotation.
+- **Persona Fallback Logic (`src/utils/persona_utils.py`)**: Switched persona lookup from `PERSONAS.get(name, DEFAULT_PERSONA)` to `PERSONAS.get(name) or DEFAULT_PERSONA` to correctly handle personas with falsy but present entries.
+
+### Removed
+
+- **Agentic Eval Harness (`eval/agentic_eval_harness.py`)**: Removed the standalone agentic evaluation harness (2,676 lines). Evaluation logic is now consolidated elsewhere.
+- **Standalone Eval Docs (`EVAL.md`, `README.md`)**: Removed the `EVAL.md` file and the corresponding Evaluation & Benchmarking section from `README.md`. Evaluation documentation will be maintained separately.
+
+### Fixed
+
+- **External Email Drop Test (`tests/test_external_email.py`)**: Refactored the customer email drop probability test to use a fully-specified source fixture and a mocked `_derive_customer_email_signals`, replacing the brittle direct `_sources` injection.
+
+---
+
 ## [v1.3.2] — 2026-03-31
 
 ### Added

README.md

@@ -8,6 +8,8 @@
 
 ### A deterministic corporate simulator for generating ground-truth ecosystems and evaluating enterprise AI agents
 
+![OrgForge corpus overview](orgforge_hero.png)
+
 OrgForge simulates weeks of realistic enterprise activity — Confluence pages, JIRA tickets, Slack threads, Git PRs, Zoom transcripts, Zendesk tickets, Salesforce records, emails, and server telemetry — grounded in an event-driven state machine so LLMs can't hallucinate facts out of sequence.
 
 The dataset is the exhaust of a living simulation. Engineers leave mid-sprint, forcing deterministic incident handoffs, ticket reassignments, and CRM ownership lapses. Knowledge gaps surface when under-documented systems break. New hires build their internal network through simulated collaboration. Stress propagates through a live, weighted social graph. Every artifact reflects the exact state of the org at the moment it was written.
@@ -35,7 +37,6 @@ The dataset is the exhaust of a living simulation. Engineers leave mid-sprint, f
 - [How the Event Bus Works](#how-the-event-bus-works)
 - [Memory Requirements](#memory-requirements)
 - [Project Structure](#project-structure)
-- [Evaluation & Benchmarking](#-evaluation--benchmarking)
 - [Roadmap](#roadmap)
 - [Adding a New Artifact Type](#adding-a-new-artifact-type)
 - [Contributing](#contributing)
@@ -380,18 +381,6 @@ orgforge/
 
 ---
 
-### 🧪 Evaluation & Benchmarking
-
-OrgForge includes a full-stack evaluation harness to measure how well AI agents retrieve and reason over the generated corporate data.
-
-- **Deterministic Ground Truth**: All answers are derived from the simulation’s state machine, not LLM hallucinations.
-- **Multi-Hop Reasoning**: Test agents on causal, temporal, and gap-detection questions.
-- **End-to-End Testing**: Use `eval_e2e.py` to run full RAG pipelines against providers like AWS Bedrock, OpenAI, and Cohere.
-
-For detailed instructions on generating eval sets, running benchmarks, and interpreting scores, see **[EVAL.md](#EVAL.md)**.
-
----
-
 ## Roadmap
 
 - [x] Native integrations for Zoom, Zendesk, and Salesforce CRM

config/config.yaml

@@ -154,6 +154,7 @@ org_lifecycle:
       day: 12
       reason: "voluntary" # voluntary | layoff | performance
       role: "Senior Backend Engineer"
+      dept: "Engineering_Backend"
       knowledge_domains:
         - "auth-service"
         - "redis-cache"
@@ -164,6 +165,7 @@ org_lifecycle:
       day: 24
       reason: "layoff"
       role: "DevOps Engineer"
+      dept: "Engineering_Backend"
       knowledge_domains:
         - "kubernetes-deploy"
         - "terraform-infra"