Derive across languages
Start a profile in English from one you trained in Spanish, or the reverse. Keep transferable habits, adapt language-dependent ones and test the result through new training.
Derive is a cross-language workflow. It is not automatic same-language inheritance or a command for merging unrelated voices.
1. Create a provisional profile
/vf-derive divulgacion-es divulgacion-en
The source must exist. The target must be new and end in the other supported language. The helper refuses to overwrite a target. It records derived_from: <source>@<version> and stage: provisional.
The command-line equivalent, from the app checkout:
python3 skills/voice-forge/scripts/vf.py derive --from divulgacion-es --to divulgacion-en
This helper creates a proposal. It does not finish translation, approve rules or prove the new voice. Copied rules initially have status: approved TODO and an adapt field, so the profile fails validation until you complete the review.
2. Understand what travels
| Source label | Derivation behavior |
|---|---|
any |
Copy the rule and translate its values |
adapt |
Keep the idea; review wording, examples, measures and scope |
en-only, es-only |
Leave it out and list it in Notes |
| No label | Use the section default |
Defaults are any for architecture, argument, paragraphs, discourse and evidence. They are adapt for sentences, vocabulary, voice and constraints. An angle-quotes check is Spanish-only.
Only no-dashes and max-paragraph-sentences automatic checks survive copying. Other automatic checks are removed and noted for adaptation. A sentence-length number measured in Spanish is not an English target. Gather English texts before proposing a new number.
3. Approve each adaptation
The agent first shows omitted rules. It then works through the copied rules in small batches:
- Translate the instruction and the values of
applies,exceptandcheck. - Adapt examples, pronouns and conventions that do not travel.
- Keep IDs, field labels, exact section headings and parser keywords in the required format.
- Show the whole block for yes, no or reword.
- On yes, set the real approval date and remove the
adaptline. On no, remove the rule and record why in Notes.
The brief, register, audience, topic descriptions and Notes need review too. Validation should pass before saving the approved result.
python3 skills/voice-forge/scripts/vf.py validate --profile divulgacion-en
4. Confirm through training
/vf-train divulgacion-en
/vf-status divulgacion-en
Copied rules have transferred support. Training prioritizes them so that you can test whether they work in the target language. A confirmed training round becomes evidence only with your approval; that rule's support becomes confirmed.
The agent reminds you once per session that the profile is provisional. Status shows how many transferred rules have been confirmed. You may use it after adaptation, while continuing to train.
5. Make it official deliberately
When you are satisfied, ask the agent to make the profile official. The helper command is:
python3 skills/voice-forge/scripts/vf.py make-official --profile divulgacion-en
It refuses while rules remain only transferred. A force option exists for an explicit user decision, but it supplies no missing evidence. Official is a stage marker, not a certification of voice fidelity.
6. Bring in later source changes
/vf-derive divulgacion-en update
The helper compares source fingerprints and writes new or changed rules to rules.update.md. It preserves the active rules.md. Adapt and approve each proposed addition or replacement, using fresh IDs for new target rules and keeping target IDs for replacements. Remove the proposal file after the changes are applied.
Profiles are still independent. Updating the source never silently edits the target. Read the complete derivation method and language scope contract.