Module 3 Build a trustworthy knowledge pack

Time plan: 25 minutes teaching, 30 minutes example, 45 minutes practice and 20 minutes checks. Keep your source manifest and version comparison. Prerequisite: a stable configuration and source-based checking skills.

What you will be able to do

Choose authoritative sources, prepare them for reuse, record scope and effective dates, manage a policy change and explain why an assistant citation does not prove correctness or permission.

A folder of files is not a knowledge policy

Uploading every available document can increase confusion. The assistant may encounter an obsolete guide, a draft with a later date and an approved policy that applies only to one month. Before thinking about retrieval, decide which sources are allowed to support a decision.

A source manifest is a small inventory that records document ID, title, owner, version, approval state, effective period, audience, sensitivity and replacement relationship. Its job is to make source selection reproducible. “Latest file” is not enough. A draft uploaded today does not override an approved policy that is still in force. A historical question may need the earlier policy rather than the current one.

For Pack A, P1 and F1 are approved for October. N0 is archived and unapproved. Exclude N0 from active knowledge. Keeping it beside the active documents with a hopeful warning is a weaker design than leaving it out. If you need it to test conflicting information, use a separate test copy and remove it afterwards.

Watch a manifest row get filled

A manifest makes a source-selection decision visible. Here is a complete row expressed as a card so every value has a reason:

- ID and title: P1, Loan policy.
- Owner and version: Equipment desk owner; version 1.
- Approval state: approved training policy, copied from the source header rather than inferred from its filename.
- Effective start and end: 2026-10-01 through 2026-10-31 inclusive.
- Audience: desk staff and adult employees.
- Sensitivity: course-public synthetic.
- Replacement relationship: no earlier approved policy is identified in Pack A; do not invent one. P2 later supersedes P1 from 2026-11-01.
- Active for the 2026-10-14 exercise: yes.
- Decision reason: the approved policy covers this audience and date, and no unresolved equal-authority conflict is supplied.

Contrast N0. Its ID is N0, title Old desk note, owner unknown, version not stated, approval unapproved/archived, recorded date 2026-09-15, effective period not established, audience not stated, and classification course-public synthetic as part of this training pack. It identifies no authoritative replacement relationship. Active on 2026-10-14: no. The exclusion reason is lack of approval and effective scope, not merely that its seven-day rule differs from our preferred answer.

Try a guided row: F1, Desk FAQ, owner Equipment desk owner, version 1, approval ______, effective period ______, audience/classification ______, replacement relationship not specified in Pack A, active on 14 October ______, reason ______. Fill the blanks from F1's header before checking the solutions. Then independently make a complete P2 row for 2 November. Explain what happens to the historical P1 row; deleting its history would make October run records harder to understand.

Prepare references for the reader and the model

Use clear text with a title, document ID, version, effective dates and numbered clauses. Keep exceptions near the rule they modify. Give tables explicit headers and make units visible. A scan with unreadable text or a row whose labels are lost can make retrieval unreliable. Inspect the extracted content in the actual tool you use; an upload success message only proves that a file was accepted.

Break a long source into coherent sections, not arbitrary fragments that separate a rule from its exception. You do not need to implement a search engine for this course. The useful principle is that a retrieved passage should preserve enough context to support a decision. Supply the complete small synthetic pack here.

A citation is a pointer for review. The model might cite the wrong clause, omit a condition or invent a document label. Check the actual clause and its effective date. The phrase “according to the policy” is not a sufficient reference when two policies exist.

Permissions travel with the source

A person having access to one document does not mean every assistant user should receive its contents. In a real deployment, confirm that the tool and its users are allowed to process and see the material. Sharing an assistant can expose instructions, references or outputs depending on the product. Private appearance and access control are different questions.

The course avoids this problem by using only synthetic, course-public references. Do not replace them with an employer handbook, real support tickets or customer messages for convenience. A realistic extension needs an owner-approved data inventory, a permission check, a retention decision and an inspection of logs as well as visible output. Secrets do not belong in instructions or reference files.

Worked change from October to November

On 14 October, a five-day webcam request is within P1.2's standard maximum. On 2 November, a new five-day webcam request exceeds P2.2's three-day maximum and requires human review. Projector and speaker limits are unchanged. P2 explicitly supersedes P1 from 1 November.

The correct update procedure is: verify P2's owner and approval state; record its effective date and supersession in the manifest; replace the active October pack for new November requests; retain a clearly labelled historical version if historical questions are in scope; test the changed webcam case and an unchanged projector case; record the release version. Do not silently rewrite old run logs. Their source version explains the answer given at the time.

D2 says six days and is dated 2 November. It remains an unapproved draft, so it does not override P2. A different situation is two contradictory, approved policies with equal authority and overlapping effective periods. The assistant cannot fix that by choosing the more convenient value. It must stop the policy conclusion and ask the owner to resolve the conflict.

Read the knowledge diagram

The visual 02-knowledge-lifecycle.svg shows candidate sources passing through owner, approval, date and permission checks before becoming the active pack. A change creates a new version and triggers regression tests. Old versions remain identifiable in historical run records. The diagram is not a claim that a vendor automatically performs these checks; you define and verify them.

Practice

1. Create a manifest row for P1, F1 and N0. Include which are active on 14 October and why. Do not use “newest” as the sole reason.
2. Prepare a small active reference file containing P1 and F1. Preserve IDs, clauses, exceptions and dates. No need to upload it if you have no approved tool.
3. Answer this request using Pack A: R110, staff, webcam, five days. Save the exact source reference supporting the duration conclusion.
4. Apply Pack B to a new request with the same equipment and duration on 2 November. Explain the changed route. Test a two-day projector request as an unchanged case.
5. Introduce the separate conflicting-approved-source situation described in the capstone. Write the assistant's safe response and the owner's next action. Do not invent a supersession rule.
6. Complete a data-flow inventory: what enters the assistant, where a reference might be stored, what an output or log might contain and who could see each. For this exercise all entries should be synthetic.

Expected artifact: source manifest, two answer records, one unchanged regression case, a conflict response and a data-flow inventory. If tests are rehearsed rather than live, retain that label.

Check your understanding

M3 Q1. An unapproved draft is newer than an approved current policy. Which should govern the answer, and what additional evidence could change that?

M3 Q2. A model cites P1.2 accurately but uses it for a request dated 2 November. What failed?

M3 Q3. A teammate says an assistant is safe to share because its chat is private. What source-access question still needs to be answered?

Before you continue

You should be able to identify the active source for a date without asking the model to guess. You will now map one request into an explicit workflow record.
