An AI-generated explanation can sound plausible while quietly changing what the source actually says. A chart without a unit becomes a temperature measurement. Two metrics moving together become a causal story. A process description loses the exception that changes the next action.
One useful review artifact is a source-to-scene map: a small record that connects each important statement to its source, the point where it appears in the video, and the limits on its interpretation.
Disclosure: I build KPainter. An AI agent drafted this article using our public, fictional demonstrations. The map below is a general review template, not a KPainter API format or an automated fact-checking guarantee.
Start with a statement that can be checked
“Explain the spreadsheet” is too broad for a reviewer. “The final completion rate is three percentage points above the example target” is checkable.
For each statement, record:
- The exact source version and location.
- The value, unit, period and comparison baseline.
- The proposed narration and its video timestamp.
- What the source does not establish.
- The person responsible for reviewing it.
The video does not need to show this entire record. Keep it beside the source so another reviewer can retrace the statement.
Example: preserve a percentage-point comparison
Our fictional workbook has a latest completion rate of 95% and a target of 92%. The difference is 3 percentage points. Calling that “3% growth” changes the calculation and introduces a growth claim that the target comparison does not support.
Here is a compact JSON template for the review record:
{
"source": {
"title": "Operating KPI workbook",
"version": "revision-1",
"location": "KPI Summary!B5:D5",
"kind": "fictional demonstration"
},
"statement": {
"latest": 95,
"target": 92,
"unit": "percent",
"period": "W12",
"difference": 3,
"differenceUnit": "percentage points"
},
"proposedNarration": "The latest completion rate is three percentage points above the example target.",
"limits": [
"Week labels are not calendar dates.",
"The file does not supply raw event records or denominator definitions.",
"This comparison does not explain the cause of a change."
],
"videoSegment": null,
"reviewer": null,
"status": "needs_review"
}
The narration in this JSON is a proposed line, not a quotation from the existing video. Leave the timestamp and reviewer empty until they have actually been assigned. A populated JSON object is not evidence that someone approved the content.
The public workbook case includes the source file, generated video and a source-to-scene map. Its Week Six anomaly is a useful review prompt; the workbook alone cannot establish a root cause.
Example: keep missing information missing
Our six-slide urban-heat demonstration contains a cooling chart with illustrative values, but the deck does not specify a temperature unit. A reviewer should not silently add °C because it seems likely.
The correct review action is to preserve the illustrative status, identify the missing unit, and ask the source owner before making a measured-temperature claim. The course example maps slides 1–6 to a 2:17 video, so readers can inspect where the explanation draws on each part of the deck.
Use the map when the source changes
When a number, threshold or process owner changes, search the map for the affected source locations. Review those statements and their video segments. Then check adjacent scenes: a changed value can also invalidate a comparison, conclusion or next step.
Store the source version and video revision together. Record who reviewed the replacement and which earlier version it supersedes. For operational training, keep the current controlled procedure available to the viewer; a video should not become an unversioned copy of the only authoritative instructions.
Before publishing, ask three questions:
- Can a second person locate the source behind each important claim?
- Does the explanation preserve units, uncertainty and exceptions?
- Can the team identify what must be reviewed after the source changes?
These checks work whether the video is recorded by a person, exported from slides or generated with AI. Which source detail is hardest to keep intact in your own explainer workflow?
Top comments (1)
The percentage-point vs percent growth distinction in your example is exactly the kind of thing that slips through even careful human review — it just sounds right until someone does the math. Encoding that in a structured JSON review record where the unit and difference are explicit fields rather than buried in prose is a smart forcing function.
The "keep missing information missing" principle is underused. The instinct to fill gaps (adding °C "because it seems likely") is strong and it's where a lot of AI explainer drift happens. Making the absence a first-class field that requires a source owner sign-off before it can be populated is a real structural fix.
The source version + video revision pairing for operational training materials is something most teams skip until something breaks badly. The idea that a video can become "an unversioned copy of the only authoritative instructions" and nobody notices — that's a real failure mode.
At Black Label we're thinking about similar provenance problems for AI-generated sales and technical content. The three-question checklist at the end is a good lightweight gate for teams that won't adopt a full review system.