Quality Gates for Captured Knowledge: Preventing Garbage-In, Garbage-Out

The failure mode of a high-volume capture system is not that it captures too little. It is that it captures at a rate that overwhelms the quality assurance process and starts publishing low-quality, inaccurate, or misleading knowledge artifacts. A knowledge base with a 20% error rate on its content will be abandoned faster than one with sparse but accurate content. Users encounter an error, lose confidence in the system, and revert to asking colleagues. The investment in the capture infrastructure is wasted.

Quality gates are the mechanisms that prevent low-quality capture from reaching the live knowledge base. They are designed around a single principle: the cost of publishing a bad knowledge artifact is higher than the cost of delaying a good one. A false negative (blocking a correct artifact) has low cost — the knowledge gets published after revision. A false positive (publishing an incorrect artifact) has high cost — it erodes user trust and may cause someone to make a wrong decision based on bad information.

1. Schema Validation The first gate: does the artifact meet the minimum schema requirements? Required fields completed, naming convention followed, owner assigned, review date set, at least one related-artifact link included. Schema validation is automated — the publication workflow rejects any artifact that fails validation and routes it back to the creator with specific fields flagged. No human judgment required at this gate.
2. Accuracy Review The second gate: is the content factually accurate? For AI-generated or AI-assisted content, accuracy review by a domain expert is mandatory before publication in high-stakes categories (policies, procedures, runbooks, decision records). For human-authored content in lower-stakes categories (FAQ additions, reference summaries), a peer review with a 48-hour window may be sufficient. The accuracy review is the gate that catches hallucinations, factual errors, and outdated information.
3. Usability Review The third gate (applied to procedural knowledge): is the content usable by the target audience? A runbook that is technically accurate but written at the wrong level of abstraction — too high-level to actually guide execution — passes the accuracy gate and fails the usability gate. Usability review is most efficiently done by asking a member of the target audience to read the artifact and identify the first point where they would need to ask for clarification.
4. Overlap Detection The fourth gate: does this artifact duplicate or conflict with existing knowledge? Before publishing a new article about a process, the system checks for existing articles about the same topic. Duplicates degrade retrieval quality by creating multiple versions of the truth. Conflicts are worse — two published articles with contradictory information actively mislead users. Overlap detection is partially automatable (semantic similarity check against existing content) and partially requires human judgment to determine whether a near-match is a duplicate or a distinct case.

The quality gate architecture should be revisited every six months as the capture system matures. In the early stages, when the knowledge base is small and the capture pipelines are new, stricter gates are appropriate — the cost of a bad artifact is higher when the knowledge base is small because users encounter the artifact more frequently. As the knowledge base grows and diversifies, the cost of an individual bad artifact decreases while the cost of over-restrictive gates (slow growth, contributor frustration) increases. The gate calibration should shift accordingly. A mature, large knowledge base can tolerate lighter-touch quality gates on lower-stakes content because accurate articles outnumber erroneous ones by a large enough margin that user trust is maintained.