Quality Assurance¶

The QA agent independently evaluates every translated segment and assigns a numeric score. Low-scoring segments trigger automatic correction loops before delivery.

What QA Evaluates¶

Criterion	Description
Source fidelity	Does the translation accurately convey the source meaning?
Language quality	Is the target text grammatically correct and natural?
Terminology	Are glossary terms applied correctly? Are forbidden alternatives avoided?
Constraint compliance	Are formal constraints (e.g. `max_chars`) satisfied?

Scoring¶

QA scores each segment on a 0–100 scale. The final qa_score in the API response is the average across all segments.

Score Thresholds¶

Threshold	Value	Applies to
QA pass score	95	All segments — passes without correction
Global minimum	90	Standard segments
Short segment (≤7 words)	85	Relaxed threshold for short segments
Very short segment (≤3 words)	75	Highly relaxed for labels/fragments
Min acceptable score	80	Below this → `needs_review`

Correction Loops¶

If any segment falls below the QA pass score:

Only the failing segments are sent back for re-translation
QA re-evaluates the corrected segments
This repeats up to 2 times

After all correction loops:

If all segments ≥ 80: job status → completed
If any segment < 80: job status → needs_review

Delivery Notes¶

When QA issues are found, the result includes structured delivery_notes describing the problems:

{
  "delivery_notes": [
    {
      "type": "low_confidence",
      "segment_id": 3,
      "message": "QA flagged low confidence in segment 3 after 2 correction loops"
    },
    {
      "type": "terminology",
      "segment_id": 7,
      "message": "Glossary term 'Bluthochdruck' could not be applied as specified"
    }
  ]
}

Delivery Note Types¶

Type	Description
`source_error`	Source text issue detected by Source Review
`low_confidence`	QA flagged persistent low confidence
`length_deviation`	Target text significantly shorter or longer than source
`terminology`	Glossary term could not be applied as specified

Scores in the API¶

QA scores are available at two levels:

Per segment: segments[].qa_score in both GET /v1/jobs/{id} and GET /v1/jobs/{id}/result
Overall: qa_score in GET /v1/jobs/{id}/result (average across segments)