AI Review

AI review is an optional quality pass that ROPARC runs over the requirements in a review. It reads each work item and flags common requirement-quality problems, posting every issue as a comment for you to triage. The AI only proposes — a human always decides. Nothing is changed automatically; findings are just comments you keep or dismiss.

Running an AI review

  1. Open an In Progress review in the document (the review side panel must be open).
  2. In the panel, click Run AI review (the violet button near the top).
  3. When it finishes, a short summary tells you how many items were reviewed and how many issues were flagged.

AI review is available to review participants only, and only while the review is In Progress. It runs over the requirement items of every document in the review’s scope; structural items (headings, the document itself) and empty items are skipped.

What it checks

The built-in rubric covers classic requirement smells that a model catches well and a human finds tedious at scale:

  • Atomicity — the requirement bundles several obligations that can’t be verified or traced individually
  • Ambiguity — vague or weak wording (“user-friendly”, “fast”, “as appropriate”, weak “should/may”)
  • Testability — no objective, measurable way to confirm the requirement is met
  • Passive voice — the responsible actor is hidden (low severity — style, not correctness)

Each finding carries a severity: info, warn, or block. A finding may also be marked low confidence when the model was unsure.

What AI findings look like

Findings are posted as ordinary comments authored by AI review — badged with a sparkle icon and shown on a violet card so they stand out from human comments. A finding names the rule it cites and its severity, explains why the rule fails, and where possible quotes the exact flagged text and suggests a rewrite.

In the review side panel’s Findings list, AI-flagged items appear with an AI · {severity} badge, and the AI filter chip narrows the list to just AI findings.

Triaging findings

On a finding’s card (in the item’s Review thread) you have two first-class actions:

  • Keep — converts the AI finding into your own comment (preserving the rule, rationale, and quote) and resolves the AI proposal, so it becomes a normal, human-owned review comment.
  • Dismiss — resolves the AI finding without keeping it.

From the Findings list, Dismiss all info clears every low-severity (info) AI finding in one click. Kept or dismissed, resolving a finding also removes it from the panel’s open-finding and progress counts.

Re-running

You can run AI review again as the document evolves. A re-run is idempotent: it clears the previous run’s still-open AI findings and posts a fresh set. Anything you already dismissed stays dismissed — the re-run skips those items/rules and notes how many it skipped — so it never nags you about issues you’ve already judged. (Findings are cleared only after the model responds, so a failed run leaves the previous findings intact.)

AI credits

AI review is a paid AI feature and spends your plan’s AI credits. Usage is metered per run against your organisation’s monthly AI-credit quota. If you have run out of credits the run is refused with an out-of-credits message, and on plans where AI is turned off entirely the feature is unavailable. Everything else about reviews — opinions, comments, checklists, the protocol — works with no AI at all.