Write for the next decision
A quality report should help a lead decide what needs attention, help an editor make a justified correction and help a tester verify the result. A page of scores rarely serves all three. Start with the decision the report supports: triage, a terminology update, another review pass or a release discussion.
Separate observations from conclusions. “The model flagged this line” describes how a question entered the queue. “A reviewer confirmed that the instruction reverses the required action” describes an assessed defect. If those statements share one count, readers cannot tell how much of the report has been verified.
Keep the summary short but attach enough detail to reproduce the findings. A confident executive headline must not hide that one language, platform or content type was excluded. Unknown coverage is not equivalent to passing coverage.

Give every finding a stable address
A finding should identify the language, string ID, resource location and build or export revision. Add the source text, observed translation and the relevant context. State the expected meaning separately from any proposed wording: several translations may satisfy the same requirement.
Describe player impact in a sentence. “Terminology mismatch” is a category; “the player cannot match the quest’s item name to the inventory label” explains why it matters. Link the evidence, name an owner and record the review state. A finding without an owner is easy to count and difficult to resolve.
ID: TERM-014
Location: quest.vault.014, French, review snapshot A
Observation: the quest uses a different name from the inventory entry.
Assessment: confirmed inconsistency after concept-ID comparison.
Impact: the required item may be difficult to identify.
Next action: language editor updates the quest; tester verifies the vault sequence.
Status: confirmed, awaiting correction.
Keep screenshot and capture references stable, but restrict access when they contain unreleased assets or personal information. A report intended for external discussion should use redacted excerpts or invented examples rather than an uncontrolled copy of the project archive.
Make denominators visible
A count becomes meaningful only when readers know its scope. Report eligible strings, strings actually checked, flags produced, flags reviewed and confirmed findings. Define whether repeated occurrences count as one root issue or several affected strings. Use the same convention throughout the report.
For an invented example, suppose a run checks 1,000 strings, produces 80 flags and reviewers assess 50 of them. If 30 reviewed flags are confirmed, the observed confirmation proportion among reviewed flags is 30 divided by 50, or 60%. It is not 60% translation accuracy, and it does not establish the proportion of errors found across the entire game.
Reviewing only the highest-risk flags introduces selection bias. Label that sample honestly. Estimating missed errors requires a separate review of unflagged content and a sampling method appropriate to the question. Avoid presenting a convenience sample as a statistically representative audit.
Separate category, impact and status
Category tells the reader what kind of issue was found: accuracy, terminology, language mechanics or technical integrity. Severity describes the impact in context. Status describes where the issue sits in the workflow. These fields answer different questions and should not be collapsed into a single colored badge.
MQM offers terminology for classifying translation issues. ISO 5060 provides general guidance on evaluating translation output. Neither replaces a project-specific explanation of scope, sampling and assessment. If you adapt a framework, name the adaptation rather than implying formal certification.
Use visible text labels in tables and dashboards. “Major — awaiting verification” is understandable without color and remains meaningful in a plain-text export. Record rejected flags and accepted exceptions outside the confirmed-error total so editorial decisions remain available without inflating defect counts.

Compare runs without inventing progress
A lower flag count may mean better translations, a narrower check, fewer strings or a different model. Log the checker version, prompt revision, glossary revision and input snapshot. When any of those change, explain the effect on comparability before drawing a trend line.
Distinguish newly introduced findings, persistent findings, resolved findings and reopened findings. Recheck fixes in the build that contains them. If a correction has been approved but not integrated, describe it as approved rather than resolved. If the issue cannot be reproduced, document the environment and the reason for closure.
End with the remaining decision: who must act, which evidence is missing and what has not been reviewed. A report can support a release discussion, but it should not imply that an AI-assisted pass guarantees a defect-free or culturally appropriate game.
Evidence that survives a handoff lets another reviewer challenge a conclusion, repeat a check and understand why a decision changed. That is more durable than a score whose inputs no one can reconstruct.