# Starting-state fidelity explained | SourceX

Canonical: https://refer.sourcex.si/guides/concept-starting-state-fidelity
Last updated: 2026-11-04

Starting-state fidelity is how accurately a record captures the situation at the moment a task began, not just how it ended. In an ordinary order desk, it means knowing which lines had shipped, which were picked and which were still open when a customer's change request arrived. For a referral, what matters is whether the company's shipment and order-change events are timestamped and kept, so that moment can be rebuilt, and who can authorize a licensing discussion.

## What does starting-state fidelity mean in an ordinary business workflow?

Put simply, the starting point decides what a correct answer is. The same request can be easy or impossible depending on that state, so a task that is graded against the wrong starting point teaches the wrong lesson.

The idea comes from how AI agents are trained and tested. In the Gymnasium toolkit, every episode begins with `reset()`, which puts the environment into an initial state and returns the first observation. Data companies apply the same discipline to business simulations: Toloka lists versioned environments and deterministic resets so that agent runs can be reproduced exactly. The research benchmark τ-bench keeps a database as hidden state that changes only through tool calls, and grades by comparing the final database with an annotated goal. Every one of those designs depends on knowing the starting state precisely.

Order changes show why. Shopify's help center notes that merchants can edit unfulfilled items but cannot remove fulfilled items or change their quantity. Most order systems work the same way, so the right answer to "please cut the quantity" depends entirely on what had already left the warehouse when the request came in.

## The order at 10:42 (Illustrative)

The company is fictional: a 410-person industrial supplies distributor. A customer ordered three lines. At 10:42 on a Tuesday, its buyer emails a change request: cut the gloves from 200 cartons to 120, add 20 first-aid kits, and send whatever has not shipped to a new site.

Here is the same order seen two ways: as it truly stood at 10:42, and as the order record reads today.

| Line | State at 10:42 (starting state) | Order record today | What a correct response needed to know |
|---|---|---|---|
| Gloves, 200 cartons | 150 shipped at 08:15 on shipment 1; 50 open | 150 shipped, line closed, quantity changed to 150 | 150 had already shipped, so 120 was impossible; the honest answer was to cancel the open 50 and offer a return for 30 |
| Safety glasses, 120 packs | Fully shipped at 08:15 | Shipped, invoiced | Nothing to change; the new address does not apply |
| First-aid kits, 40 | Not yet picked | 60 shipped at 15:30 to the new site | Open line, so the extra 20 and the new address could both apply |
| Delivery address | Original site on all lines | New site on the kits shipment only | The address change applied only to open lines |
| Invoice | Not yet issued | Issued for final quantities | Pricing could still change at 10:42 |

The record today is accurate. It is just the wrong starting point. Someone reading it would see a closed glove line at 150 and could not tell whether the customer asked for 120, got 150 and accepted a return offer, or never asked at all. A model trained on the final record would learn that the change request "produced" 150 cartons, which is backward. This is one reason [AI agents fail](/questions/why-ai-agents-fail) on order tasks that look routine: the hard part is reading the current state correctly before acting.

## Evidence fields that rebuild the starting state

The 10:42 column above cannot be read off any single screen. It has to be rebuilt from events that each carry their own time. These are the fields that make the rebuild possible.

| Field | What it restores | In the example (Illustrative) |
|---|---|---|
| Original order lines with quantities | What was asked for before any change | Gloves 200, glasses 120, kits 40 |
| Shipment events with timestamps and line quantities | What had left the warehouse, and when | Shipment 1 at 08:15: gloves 150, glasses 120 |
| Pick and pack status changes | What was in progress but not yet shipped | Kits not yet picked at 10:42 |
| Change request with arrival time | The exact moment the task began | Buyer's email at 10:42 |
| Order edit history with before and after values | How the record was changed in response | Gloves 200 to 150; kits 40 to 60 |
| Address and ship-to changes by line | Which lines a change could still affect | New site on the kits shipment only |
| Invoice issue time | Whether pricing was still open | Not issued at 10:42 |

The edit history row is what most systems lack. Many order screens show only current values, so the starting quantity survives only if the system logs each change with its old value. Warehouse and carrier systems often keep better timestamps than the order system itself, which is why the shipment events usually do most of the work.

One example cannot show how many change requests can be matched to a precise moment, or whether timestamps survived a system migration. SourceX measures that in its review with the company.

## Does the company qualify, and what stays unknown?

Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English. The company also needs 50 or more full-time employees at peak (contractors excluded), documented operating history, rights to license the records, and an authorized contact.

Distributors often run separate order, warehouse and carrier systems, and sometimes a third-party logistics provider holds the shipment events. Who owns those logs is the first rights question. What stays unknown until SourceX reviews it: how many change requests can be matched to a precise starting state, whether timestamps survived system migrations, how customer data will be handled, and whether any buyer wants this task type. Partners never request order exports or sample histories. Real partial-shipment cases are hard to invent convincingly, which is part of why [synthetic data rarely replaces real records](/questions/will-synthetic-data-replace-real-data). More concepts sit in the [data types hub](/data-types), and [inventory and fulfillment](/data-types/inventory-and-fulfillment) covers the category.

## Go deeper on starting-state fidelity

- [Questions to ask a client about starting-state fidelity](/questions/does-starting-state-fidelity-data-qualify): a discovery list for an operations or fulfillment lead, kept at metadata level.
- [Record gaps that make starting-state fidelity hard to assess](/questions/starting-state-fidelity-evidence-gaps): the overwritten fields and missing timestamps that most often prevent a rebuild, and how to spot them early.
- [Starting-state fidelity review packet](/resources/starting-state-fidelity-review-packet): what an evidence packet for SourceX's review should contain, item by item.

## Next step

> **Refer a company. No cap on rewards.** Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.

If the operations lead confirms timestamped shipment events and kept order-change history, [register as a partner](/register), check the company against [who qualifies](/who-qualifies), and read the [terms](/terms) before you introduce the decision-maker.

## Sources

- https://gymnasium.farama.org/api/env/
- https://toloka.ai/agent-data
- https://arxiv.org/abs/2406.12045
- https://help.shopify.com/en/manual/fulfillment/managing-orders/editing-orders/considerations
- https://refer.sourcex.si/terms
