Policies and versions

What the platform is allowed to do in Phase 1, which versions are live, and how every change is tested before it reaches production.

Step 23 of 24 · Presenter note
What the AI is allowed to do in Phase 1: new, fuzzy and AI-corrected content is always reviewed, hard-fail rules block release, and every change runs the regression suite.

Review policy for Phase 1

Locked in Phase 1
Unit typeWhat happensHuman review
Approved exact matchReused without AI, after terminology, protected-content and integrity checks. Shown as context.Context only
Exact match in a hazard statement (High criticality)Reused, but shown for confirmationConfirm
Exact match with a conflict or context warningRouted to reviewRequired
Fuzzy matchUsed as evidence for the AI; never approved as outputRequired
New AI translationTranslated with terminology and contextRequired
AI-corrected unitCorrected by the loop, history visibleRequired
Hard fail or unresolved findingMandatory human correctionRequired

Reducing human review needs a calibrated quality score and sign-off by Quality and Legal. That is Phase 3.

AI correction loop

Proposal for Phase 1: 2

Hard-fail rules

§17.4
  • HF-01Changed safety-critical number or unit
  • HF-02Omitted warning, prohibition or condition
  • HF-03Altered protected identifier
  • HF-04Invalid tag or placeholder that breaks the target system
  • HF-05Missing translation unit
  • HF-06Duplicated content that changes instructions
  • HF-07Obsolete source version
  • HF-08Unsupported publication output
  • HF-09Source ambiguity that prevents reliable translation
  • HF-10Critical terminology violation
  • HF-11Failed reconstruction integrity

Live versions

ComponentVersionSinceApproved byRegression
Translation modelprivate-llm 2027-06 (EU tenant)2 Aug 2027Priya RamanPassed · 1,240 units
Prompt templatetr-dita-install v1.416 Aug 2027Priya RamanPassed · 1,240 units
Evaluatorgtl-qe 0.916 Aug 2027Priya RamanPassed · 400 seeded defects
Hard-fail rulesv3 (11 rules)1 Jul 2027QualityPassed
Segmentation rulesacolad-compat-srx 1.0.214 Jun 2027Localization engineerPassed · leverage −0.6 pp
Parserdita-xliff 1.3.014 Jun 2027SISPassed · 212 topics round trip
TerminologyTermWeb 2027.09 + local L-2141 Sep 2027Luca Bernasconi4 conflicts open
Proposed change

Prompt tr-dita-install v1.5

Blocked: 4 regressions

Adds style rules for hazard statements. Submitted by Priya Raman, Tue 14 Sep.

Regression suite
1,240 benchmark units in 9 languages
Passed
1,236
Regressions
de-DE 2, th-TH 1, ar 1: prohibited term reintroduced in 3 units, one tag moved

AI service

  • Private model, EU tenant
    Healthy · 138 ms median per unit
    Live
  • Fallback model
    None, by policy. No unapproved model is ever used silently.
    None
  • On an outage
    Work queues and resumes; GTL is alerted after 15 minutes.