# Fixed starting data

[中文](data-guide.zh-CN.md) · [Run the two tasks](quickstart.en.md) · [Read reports](report-guide.en.md)

Open [draft_for_approval.json](../data/fixed-v0.2/baseline/draft_for_approval.json) and [execute_approved.json](../data/fixed-v0.2/baseline/execute_approved.json) before your first run. These are original synthetic starting states for maintainer inspection. Each run initializes a fresh World from the selected runtime seed; editing an exported JSON does not configure the runner.

## Versions and source

Code `0.5.0`, business task `prototype-v0.1`, language resources `task-language-v0.4.0`, data `fixed-v0.2`, JSON schema `fixed-data-v0.2`. The [manifest](../data/fixed-v0.2/manifest.json) records file hashes and row counts. `provenance` records source paths and hashes. Export uses the same `World` initialization as execution, excluding the duplicate maintainer `initial_state` field. There is no separately maintained seed or newly computed reference audience.

Top-level metadata includes `dataset_id`, `task_id`, `fixed_clock`, `units`, `provenance`, and `agent_access`; `policy` contains the fixed public English policy, and `state` contains initial business records. The JSON is not translated when display language changes. An agent's `read_policy` returns the selected task language.

## Three data instances, unchanged policy

| Instance | Change from baseline | Purpose | Starting files |
|---|---|---|---|
| `baseline` | Original records | Preserve original meaning | [Draft](../data/fixed-v0.2/baseline/draft_for_approval.json) / [Approved](../data/fixed-v0.2/baseline/execute_approved.json) |
| `partial_refund` | O009 refund changes from 10000 to 5000 cents, with payment still 10000 and completion one day before the clock | A positive-net partial refund remains a valid purchase | [Draft](../data/fixed-v0.2/partial_refund/draft_for_approval.json) / [Approved](../data/fixed-v0.2/partial_refund/execute_approved.json) |
| `contact_kinds` | M003's six-day-old wecom contact becomes service; add M001's one-day-old sms service contact and M008's six-day-old wecom marketing contact; retain M005's exact seven-day-old sms marketing contact | Only marketing contacts count, across all channels; service contacts do not exclude a member | [Draft](../data/fixed-v0.2/contact_kinds/draft_for_approval.json) / [Approved](../data/fixed-v0.2/contact_kinds/execute_approved.json) |

The contact instance has four contact rows. All three have 8 members, 9 orders, 8 consent rows and 1 ticket. Each approved starting state has its own consistent campaign and approval; each draft state has neither. The budget field remains 3000 cents even when fewer people qualify. The old fixed-v0.1 files and 0.4.1 archive retain their original content.

There is one business template, two task states and three data instances: six task/data combinations. Maintainer expectations are defined separately from runtime seeds and the reference agent; they must not be returned through normal tools. Data variants do not change policy or add identity-merging scenarios.

## Baseline rows and task-state differences

| State table | Draft task | Approved task | Fields and meaning |
|---|---:|---:|---|
| `members` | 8 | 8 | `member_id`: unified member ID |
| `orders` | 9 | 9 | `order_id`, `member_id`, `status`, `paid_cents`, `refunded_cents`, `completed_at` |
| `consents` | 8 | 8 | `member_id`, `channel`, `valid`; membership is not consent |
| `contacts` | 2 | 2 | `member_id`, `channel`, `kind`, `occurred_at`; includes different marketing channels |
| `tickets` | 1 | 1 | `ticket_id`, `member_id`, `status`; an open support case |
| `campaigns` | 0 | 1 | ID, audience IDs, channel, coupon value, budget, status, content digest |
| `approvals` | 0 | 1 | Approval/campaign IDs, bound content digest, status, approving seed identity |
| `sends` / `coupons` | 0 / 0 | 0 / 0 | Empty initial effect ledgers; these are not missing data |

The approved task begins with campaign `C001` and approval `A001`; the draft task has neither. Existing approved campaign content is a queryable business fact. No task starts with completed sends or coupon issuance.

## Units and time

- All money is integer **CNY minor units**: 100 = CNY 1. Every order's `paid_cents` is 10000; on baseline and contact_kinds one order has `refunded_cents=10000`, the others 0; on partial_refund that order has `refunded_cents=5000`. Coupon face value is 1000 per person; total face-value budget is 3000. The budget is an upper limit, not an amount that must be spent. These amounts are not USD or redemption costs.
- Historical design prose used 5000 for order payment. The runtime seed and this export use 10000. The current explanation follows the runtime; the seed has not been changed to match old prose.
- Time is UTC and fixed at `2026-09-19T00:00:00Z`. One day is 24 hours. Completed orders count only if paid minus refunded is positive. Full refunds do not reset latest valid purchase; positive partial refunds do.
- Dormancy is `[90,180)` days. The all-channel marketing contact exclusion window is `(as_of-7 days, as_of]`. Exact 90 days is included; exact 180 days excluded; exact seven-day-old contact is allowed. Consent, unclosed support tickets, budget and approval rules still apply.

## Inspect without modifying

From the package root after creating the environment in the quick start:

```sh
.venv/bin/python -m json.tool data/fixed-v0.2/baseline/draft_for_approval.json
.venv/bin/python -m json.tool data/fixed-v0.2/baseline/execute_approved.json
.venv/bin/python scripts/export_fixed_data.py --output data/fixed-v0.2 --check
```

`--check` compares the saved data with a fresh runtime initialization; it does not run an agent or scorer. A mismatch needs maintainer review, not manual edits to pass the check. This guide does not require regeneration or replacement.

## Agent boundary

Full exports are maintainer material. Normal agents query members, orders, consents, contacts, tickets and policy through their existing read tools, and campaigns through list/get tools. Full approval, send and coupon ledgers in the export are maintainer-only; this does not hide the lawful approved campaign itself. Reference audiences, scores, and scoring internals are not exported as new answer fields or exposed through normal tools.

The `agent-input` specifications describe tasks and tools in the selected language and data instance; their file count is not a business-scenario count or a set of raw-data tables. Do not feed a full source archive, export, or prior report to the tested agent. This interface boundary is not a hostile-code sandbox. The data is not a training dataset or a new HF dataset publication. Historical model results remain code 0.3.0; current examples are synthetic scripts, not additional model results.