01
Pin the instruction version
Every check runs against one approved release of the client's instructions, so a later change never silently rewrites an earlier verdict.
QA evaluation
Check the deliverable against the instructions the client actually approved.
Deterministic checks find the mechanical misses, people make the calls, and every decision stays on the record. The corrected batch is checked again with the same rules, so you can show a client exactly what changed between delivery one and delivery two.
Runs today as a pilot on one of your batches, together with us. No self-serve sign-up yet.
02
Run the deterministic checks
Placeholders, markup, forbidden phrases, required glossary terms, empty targets: the mechanical rules a reviewer should never have to spot by eye. A pass is not a quality verdict; a finding is a candidate for a person.
03
Decide, on the record
Reviewers uphold, dismiss, request a correction or ask for a rule clarification. Decisions are append-only, disagreements stay visible, and a stale edit is refused instead of overwriting someone else's call.
04
Recheck the corrected batch
The corrected delivery is checked again with the same rules: what was resolved, what remains, what is new, and what could not be checked, never counted as a pass.
What stays true in every run
- Evidence export keeps IDs, hashes, rule versions and decisions; it never carries the source or target text.
- Pairs with Knowledge Companion: the instruction release that was published there is the one the checks are pinned to.
- Built for localization and AI-data deliverables first: MTPE batches, annotation exports, evaluation sheets.