Source linking

How the Show source panel works, what the colour-coded flags mean, and how reviewers clear them before sign-off.

Source linking is the feature that lets a reviewer trust a draft they did not write. Every AI-written paragraph in FRA Flow has a Show source button. Tap it and a panel slides in from the right showing every observation the paragraph was built from, plus any divergence between what the AI was given and what it cited.

What the panel shows

For each observation linked to the paragraph, the panel renders:

• The full description and recommended action.
• The risk level and the location it was captured at.
• Photos as inline thumbnails (tap to open full-size in a new tab).
• Voice notes with an audio player and the transcript.
• The named assessor who captured it.

The observations are listed in the same order as they appear in the paragraph's input set, so the reviewer can read the paragraph and the source list side by side.

The two divergence flags

At the top of the panel, two flags surface differences between the input set and the AI's cited evidence:

• Yellow: "AI did not cite this observation." The AI received this observation in the input set but did not include it in its cited list. Either the AI judged it not worth mentioning, or the paragraph is missing coverage.
• Red: "AI cited this observation that was not sent." The AI claimed to use an observation that is not in the input set. This is the hallucination guard. Sign-off blocks until the reviewer acknowledges the warning.

Reviewer triage

Three rules of thumb for a reviewer working through the flags:

1. Yellow is informational, not blocking. Most yellow flags are fine. The AI was given five observations and only two needed mentioning; the other three are routine. Tap the observation in the panel, confirm coverage is right, move on. Do open the observation if you do not recognise it.
2. Red is always blocking, but rarely real. Red almost always means the AI mis-cited an ID format, not that it invented a finding. The panel resolves the cited observation if it can find it (often a sibling section in the same assessment), and tells you when it cannot. Either way the reviewer acknowledges, the report records the override, and sign-off continues.
3. The bigger the divergence, the more likely the paragraph needs regeneration. If a paragraph has three yellow flags and a red, it has not synthesised the input properly. Tap Regenerate on the paragraph card, let a fresh draft come in, re-read.

What "cited" actually means

The AI is asked to return the paragraph plus the list of observations it referenced. "Referenced" is the AI's judgement, not a syntactic check. The AI may cite an observation it only alluded to ("a number of door-related issues were noted") or omit one it summarised generically ("housekeeping was generally satisfactory" without listing the housekeeping observations).

The flags are signals, not verdicts. The reviewer is the verdict.

What if a paragraph is human-authored?

The Show source panel still works. A paragraph that has been switched to Human authored mode keeps its observation input set attached. The reviewer can verify the human author drew on the right inputs even though the AI is no longer involved.

In practice yellow flags in human-authored mode are common because a human will write tighter prose than the AI; one yellow flag per observation that the human consciously skipped. Red flags in human-authored mode are rare and usually point at a copy-paste-from-elsewhere accident.

Where to go next

• How AI drafts work covers the generation pipeline that produces the cited list in the first place.
• Hallucination guards walks through every guard that fires before sign-off, including the red flag.
• Audit trail explains what gets embedded in the final PDF for years-later challenges.