# Tern Harbor synthetic data

Tern Harbor is a fictional software company for field-service teams. The records, messages and pilot outcomes are authored examples. They do not measure a model's performance.

| Table | Grain and keys |
| --- | --- |
| teams | One delivery team per id. |
| weekly_capacity | Available hours by week (Monday) and team. |
| work_items | One work item per id, linked to week and team, with hours, owner, due date and consequence. Flexible work can move into the next week. |
| accounts | One invented customer account per id. |
| requests | Sample intake messages by id. The dataset includes repeats, new evidence, and two dispatch failures. |
| routing_history | Previously accepted tickets. Match by account and normalized reference. Original intake time anchors the response deadline. |
| gates | One release requirement per id. depends_on lists prerequisites separated by a pipe. |
| agent_fixtures | Five reusable prompts, source excerpts and expected actions for an external agent test. |
| pilot_trials | 36 distinct tickets with policy, authored draft, expected/actual action, simulated reviewer response, severity, and score. |

## Delivery Planner

The dashboard compares committed and flexible hours by team. For the Cedar scenario, Support can lend up to six hours of certified access-mapping coverage, subject to spare capacity. Deferred work carries into the next week for its own team. The dashboard does not schedule individual hours or guarantee that dependent tasks finish before the rehearsal. Open the SQLite database and run starter-queries.sql to inspect the base data.

## Request to Action

Import request-to-action.n8n.json and run the manual trigger. The built-in synthetic batch has separate outcome branches. Inspect first_status, final_status, first_trace, retry_trace, writes and tickets_added for the two timeout examples. The workflow simulates the ticket store in memory. It makes no external calls, and elapsed-hour targets use UTC rather than a business calendar. For an external ticket system, use an atomic unique key or a supported idempotency key, reconcile ambiguous failures, and use the actual SLA calendar.

## Launch Readiness

The opening pilot scenario has 34 passes, an unconfirmed scheduling draft and a critical refund promise caught in review. The 36 tickets are grouped into five policy categories. Change the scenario to see how customer exposure, loss of review, a single replay or broader regression evidence affects the recommendation. Thresholds are illustrative. The tool's Go decision is a recommendation to the program lead.

To test a separate agent, provide the input and reference from agent-fixtures.jsonl. Keep expected_action and expected_point for review. Paste the agent's case ids and action labels into the site to score action selection. Review actual draft content against the reference separately.