Capture how experienced people reason, then preserve the source, context, exceptions, disagreements, and evidence needed to test each rule.
GUIDED KNOWLEDGE AUDITHUMAN EXPERTISEREVIEWED 31 AUG 2026
A model cannot inherit someone’s judgment by reading a document. It can help interview, organize, retrieve, and draft. The result is a set of hypotheses and decision aids that people must verify, not a digital copy of an expert.
Use fictional or sanitized examples during extraction. Do not upload real client, employee, customer, competitor, contract, credential, or confidential material. Before any authorized operational use, establish rights or lawful basis, minimization, vendor and retention review, access control, and an accountable human decision owner.
Part 1: What Counts as Domain Knowledge?
01
Definitions and facts: terms, policies, standards, and dated external requirements with authoritative sources.
02
Deterministic rules: checks that yield the same result from the same inputs and can usually live in code or a checklist.
03
Judgment hypotheses: patterns an experienced person believes are useful, with context, evidence, confidence, counterexamples, and exceptions.
04
Escalation knowledge: signals that require a specialist, a second reviewer, more evidence, or a stop.
EXPERIENCE IS EVIDENCE, NOT UNIVERSAL LAW
“This often worked in our 2025 retail campaigns” is useful when its sample, market, and exceptions are visible. Rewriting it as “always do this” removes the information a reviewer needs to know when it may fail.
Part 2: Use a Knowledge Card
Knowledge ID: [ID]
Statement: [ONE TESTABLE CLAIM OR RULE]
Type: FACT | DETERMINISTIC_CHECK | JUDGMENT_HYPOTHESIS | ESCALATION
Source or author: [WHO / DOCUMENT / LINK]
Observed context: [MARKET, CHANNEL, PERIOD, SAMPLE]
Supporting evidence: [WHAT WAS OBSERVED]
Confidence: LOW | MEDIUM | HIGH
Known exceptions: [WHEN IT DID NOT APPLY]
Counterexample: [ONE CASE THAT CHALLENGES IT]
Required inputs: [FIELDS + DEFINITIONS]
Prohibited use: [DECISIONS OR CONTEXTS OUTSIDE SCOPE]
Reviewer: [ACCOUNTABLE OWNER]
Last reviewed: [DATE]
Review trigger: [CHANGE / INCIDENT / EXPIRY]
Status: HYPOTHESIS | VALIDATED_FOR_SCOPE | DEPRECATED
Never let a model merge several people’s views into one confident rule without attribution. Preserve disagreement. A minority exception may be the most important safety signal in the document.
Part 3: Run a Safe Knowledge Interview
Choose one narrow fictional process and name the intended user and decision boundary.
Ask the practitioner for a recent sanitized example, then ask what evidence they saw, what alternatives they considered, and what would have changed the decision.
Capture exceptions, failure cases, regional or language differences, and signals that trigger escalation.
Ask a second practitioner to review the card independently. Record agreement and disagreement; do not force consensus.
Trace external rules to current primary sources. Label personal experience as experience, not regulation or market fact.
Have the author approve the wording and scope. A model may format notes, but it cannot certify them.
You are an interview facilitator, not an expert or decision-maker.
The case is fictional: [CASE]. Ask one question at a time to help me document:
- the decision and required evidence
- the context in which my rule was observed
- exceptions, counterexamples, and uncertainty
- deterministic checks versus judgment
- prohibited uses and escalation signals
After each answer, draft a Knowledge Card using only what I said. Quote my wording where useful, mark unsupported fields UNKNOWN, and never generalize beyond the stated context. Do not request names, personal data, credentials, client files, or confidential examples.
Part 4: Turn Rules into Testable Hypotheses
Claim: “Short copy always converts better in the Gulf.”
Problem: no market, channel, audience, period, metric, sample, comparison, or exception.
Hypothesis: “For the fictional mobile acquisition test A-17, concise Arabic headlines may reduce comprehension errors compared with the long variant.”
Test: predefine comprehension and conversion measures, use properly randomized variants, check translation and accessibility, document sample and duration, and report uncertainty. A qualified owner decides whether evidence supports this context only.
01
Specific: population, environment, intervention, comparator, outcome, and time window are named.
02
Falsifiable: evidence that would weaken or disprove the statement is defined.
03
Bounded: prohibited uses and cases requiring escalation are explicit.
04
Reviewable: source, author, version, date, and accountable reviewer are visible.
Part 5: Test Retrieval and Application
A correct card is not enough. Test whether a candidate system retrieves the right card, preserves its caveats, ignores malicious instructions in sources, and stops when context falls outside the card’s scope.
Representative
Synthetic cases squarely inside the documented context, including Arabic and English where needed.
Expected: cite and apply carefully
Boundary
Cases near a threshold, with missing evidence, conflicting cards, or a changed market and date.
Expected: qualify or escalate
Malformed & adversarial
Wrong field types, secret requests, and source text that tells the model to ignore the knowledge policy.
Expected: refuse and surface
Review these versioned Knowledge Cards against the synthetic cases.
For each case return:
- applicable card IDs and quoted scope
- supplied facts versus interpretation
- missing or conflicting evidence
- exception or counterexample triggered
- DRAFT, NEEDS_REVIEW, OUT_OF_SCOPE, or REFUSED
- checks for the human reviewer
Do not resolve disagreement by majority vote. Do not use unstated knowledge. Do not recommend or perform a live action.
Part 6: Govern the Knowledge Base
Ownership: each card has an author and an accountable reviewer; high-impact material uses separation of duties.
Versioning: edits create a new version with a change log; prior approved versions remain available for rollback and audit.
Access: readers see only cards required for their workspace and tenant. Source documents do not become globally searchable by default.
Lifecycle: expiry, policy change, data drift, an incident, or failed evaluation triggers review or deprecation.
Monitoring: record which version supported a draft, what the reviewer changed, and where retrieval or application failed.