# Implementation Plan: Localization work orders

## Approach

The layer between knowing something changed and doing anything about it. Detection
(spec 001) says which units moved; this says what that costs and what about it is
risky, so the answer arrives as a plan rather than an alert.

Everything it reports already exists in pieces and is assembled rather than recomputed:
coverage from the approved memory, the keep-in-source list for segments legitimately
identical, the termbase for agreed renderings. A work order that worked out its own
numbers would be a second opinion that could disagree with the run it describes.

Four risks are worth flagging because each one has already cost something here:

a term with no approved rendering and no termbase entry is where a fresh invention
enters the corpus unnoticed;

an approved rendering that disagrees with the termbase is how retired wording gets
reused — two such pairs exist today, both carrying "identidade de pagamento" after
the termbase moved to "identidade do pagamento";

a segment identical to its source is ambiguous without the keep-in-source list, and
guessing either way is wrong: "Polling" is correct and "Bank Name" was not;

an instruction naming a control is a cross-reference, and translating the instruction
while the control's label stays in the source language, or the reverse, produces text
that tells a reader to click something that is not there. This is the class that
produced the "Leave as Not applicable" case.

Ambiguity escalates rather than resolving (FR-006). The work order's job is to put a
decision in front of someone, not to make it: every escalation carries a recommendation
so the decision is cheap, and none of them is acted on automatically.

## Rollback

Reads and reports. No content is touched.

## Appendix

### Alternatives considered

**Estimating effort in hours.** Rejected, and recorded as a deferral in the spec. There
is no measured throughput baseline, so any figure would be invented and then trusted.
Segment counts and coverage are measured and sufficient to sequence work.

**Resolving conflicts automatically in favour of the termbase.** Rejected. The termbase
usually wins, but not always: the conflict is sometimes evidence the termbase entry is
the stale one. Reporting both renderings costs a reviewer seconds and silently picking
one costs a corpus-wide inconsistency.