How does context engineering use the documents behind a business decision?
It uses them to define which inputs a correct decision depended on, so an agent can be given those inputs and judged against the outcome. Anthropic describes context engineering as managing the whole context state an agent works with across many steps, and notes that a model's precision on retrieval and long-range reasoning can degrade as that context grows. More is not better. The task is selecting the few documents that matter.
Business decisions show why selection is hard. A credit approval, a refund exception or a change request was decided under a specific rule, using specific evidence, by someone with specific authority. Months later the rule has changed, the evidence sits in three systems and the approver has moved on. An agent handed today's policy and a pile of files will often make a confident, wrong decision.
A decision record that names its policy version and evidence answers three questions an agent builder needs: what should have been retrieved, what should have been ignored, and what the correct outcome was given those inputs.
A field map: one decision and its context bundle
Illustrative. A fictional 300-person industrial distributor raises a customer's credit limit. All details are invented.
| Field | Value in the record | System | Role in context |
|---|---|---|---|
| Decision | Credit limit raised from $60,000 to $120,000 | ERP customer master | The outcome to predict or check |
| Decision date | 14 March | ERP change log | Selects which policy applied |
| Policy in force | Credit policy v4.2, effective 1 February | Document management | The rule the decision followed |
| Threshold that applied | Limits above $110,000 need CFO approval (v4.2, section 3) | Policy text | Explains the approval route |
| Evidence 1 | 24 months of payment history, average 31 days to pay | AR aging report | Fact the policy tests |
| Evidence 2 | Two trade references | Credit application file | Fact the policy tests |
| Evidence 3 | Latest year-end financial statements | Credit application file | Fact the policy tests |
| Approval | CFO, 14 March, note: "strong payment record, reviewed statements" | Approval workflow | Authority and reasoning |
| Later outcome | No balance past 60 days in the following year | AR aging report | Shows the decision held |
Now the trap. The company issued credit policy v5.0 in August, lowering the CFO threshold to $75,000 and adding a credit insurance check. An agent given v5.0 to judge the March decision would flag a missing insurance check that was never required. The record is only useful because the decision date and the policy version are both kept, and the old version was not overwritten.
| Context choice | What the agent concludes | Correct? |
|---|---|---|
| v4.2 plus the three evidence documents | Approval route and decision consistent with policy | Yes |
| v5.0 plus the same evidence | Decision missing a required insurance check | No: wrong version |
| v4.2 with no evidence | Cannot test whether the policy criteria were met | Incomplete |
| Every credit file for the customer since 2015 | Slower, and older notes may mislead | Too much context |
Why do decision records matter to AI developers?
They turn "use the right documents" into something that can be checked. Evaluation vendors describe review loops in which a person compares a model's whole attempt, including the steps and tool calls, against a rubric; micro1, for example, describes its flow platform that way and says it builds realistic multi-step task environments. In a business setting, the decision record with its linked policy and evidence is close to that rubric already written: it says which inputs were relevant and what the outcome was.
The same structure appears across functions. The credit limit approval page covers the example above in more depth. Decision records and operational decision logs cover the general pattern, and standard operating procedures are the policy layer when they are versioned and dated. For a plain-language overview to share with a client, use the client explainer on licensing operational data.
Company qualification questions
Discovery stays at metadata level. You are checking whether the links exist and who owns them, never the content.
| Ask | Strong answer | Why it matters |
|---|---|---|
| Are policies versioned with effective dates, and are old versions kept? | Yes, in a document system with history | Lets the right version be matched to each decision |
| Does each approval record the policy or threshold it relied on? | Often, through a workflow field or note | The link between rule and decision |
| Is supporting evidence attached to the decision record? | Attached or linked by ID | Lets the evidence set be rebuilt |
| How many years of decisions exist, and in which systems? | Five or more years, named tools | Depth |
| Who can authorize a licensing discussion? | The CEO, usually the owner in an owner-run company | The sponsor to introduce |
The introduced company needs 50+ full-time employees at peak (contractors excluded), a documented operating history, rights to license the records and an authorized contact. 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.
What stays unknown until SourceX reviews it: SourceX's current data needs for this record type, whether customer or employee data in the files can be redacted adequately, the rights review result, and any offer.
Limits and unanswered questions
This page does not claim that any AI developer has licensed decision records for context engineering, or publish figures on their value; SourceX has not released such data. The sources cited describe context engineering and evaluation practice in general terms.
Use type also matters. A company may be comfortable licensing records for model training or evaluation but not for live retrieval inside a deployed product, or the reverse; the training vs retrieval license comparison explains the difference. Policies themselves can carry third-party content, such as an insurer's underwriting guide, that the company cannot license.
Company compensation and referral rewards are separate. The company is paid under its own agreement with SourceX. The partner's reward is calculated on what SourceX actually pays the company, is funded separately and never reduces that payment. Illustrative only: a $400,000 payout would mean a $70,000 reward. See how rewards work.
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 a client keeps dated policy versions and links its approvals to them, register as a partner, check the company against who qualifies and read the program terms before you introduce its decision-maker. More agent data topics are in the guides hub.