Spanish-Language Model Corpus Governance Director

The mandate

The developer’s training mixture combines licensed publications, public web material, partner archives, conversational data and internally generated text, but sampling cannot reliably reproduce which source, permission and processing rules produced a given corpus slice. The defined problem is to establish a controlled, regionally intelligible evidence chain before enterprise customers rely on claims about provenance, balance or removal.

The deliverable is a Spanish Model Corpus Control Book containing a source register, permission and restriction taxonomy, transformation lineage, regional and genre sampling model, duplication and contamination tests, sensitive-data routing, deletion propagation method, dataset version manifest and an operating approval workflow. It is an engineering and governance artefact, not a legal opinion on individual sources.

Work begins 4 January 2027. Milestone one on 29 January is the reconciled source universe, evidence-gap report and signed taxonomy; milestone two on 26 February provides the sampled lineage audit, regional composition analysis and removal rehearsal; milestone three on 26 March delivers the accepted control book, reproducible manifests, exception ledger and ownership transfer.

Acceptance requires the Chief Data Officer’s team to select a statistically and risk-based sample and trace it from source evidence through filtering, deduplication, transformation and training-ready shard; internal audit must reproduce the regional composition measures and one withdrawal path. Unsubstantiated permission, unlocatable derivatives or material unexplained variance will block final acceptance and milestone closure.

The client will provide collection records, licences and restrictions, raw and processed snapshots, pipeline code, content hashes, language classifiers, model-training manifests and secure compute. Named legal and linguistic owners will resolve interpretation rather than delegating it to the consultant; source renegotiation, wholesale dataset replacement, model retraining and legal conclusions remain outside scope.

Why this is external work

Acquisition teams are measured on corpus growth, while training engineers inherit mixtures and counsel sees only sources escalated for review. No internal owner can independently connect rights evidence, transformation lineage and regional composition across the whole pipeline. A bounded external build gives enterprise claims a testable operating basis without pretending technical traceability resolves every legal judgment.

What you will own

• Reconcile acquired, scraped, partnered and generated source populations into a register with evidence, restrictions, geography, language variety and accountable owner.
• Trace filtering, language identification, normalisation, deduplication, quality scoring, safety processing and sharding into reproducible version manifests.
• Design risk-based and statistical samples that test lineage completeness across source class, region, genre, age, permission and transformation path.
• Quantify regional Spanish, code switching, domain, register and publication imbalances without treating country labels as linguistic quality proxies.
• Rehearse correction and withdrawal from source identifier through processed derivative and future training exclusion, recording technically irreducible uncertainty.
• Establish exception gates for ambiguous rights, provenance loss, sensitive material, partner restrictions and mixtures whose composition cannot be recreated.
• Transfer schemas, queries, manifests, sampling logic, approval roles and refresh triggers to the permanent data-governance owners.

Candidate qualifications

• Governed training corpora for a large language or foundation model across acquired, licensed, web-derived and generated source classes.
• Built reproducible data lineage through filtering, deduplication, scoring and sharding where earlier identifiers were incomplete or transformed.
• Analysed regional Spanish or another pluricentric language without collapsing geography, dialect, register and domain into one balance measure.
• Implemented correction, deletion or future-exclusion controls across dataset versions and can state what trained parameters could not reveal.
• Worked jointly with data engineers, linguists, counsel and auditors while preserving their distinct evidence and decision responsibilities.
• Delivered a corpus-control system that supported enterprise or regulatory scrutiny after the external engagement had ended.

Non-negotiables

• The named director must attend both Buenos Aires workshops and the Madrid partner review despite the otherwise remote engagement.
• No corpus sample, licence, partner archive, model artefact or query result may leave the client’s approved environment.
• Will not convert technical provenance into a legal clearance statement or treat public accessibility as documented permission.
• Must disclose present publisher, archive, data-broker, model-provider and rights-holder engagements before accessing source evidence.

1. 49 words maximum. Describe a corpus lineage gap you resolved and the transformation step that had obscured the source.
2. 49 words maximum. How would you measure regional Spanish representation without mistaking geography for linguistic quality or intended use?
3. 49 words maximum. Confirm three-month capacity, Buenos Aires and Madrid travel, and every relevant publisher, archive or model relationship.