A sovereign model is only as trustworthy as the corpus behind it. We assemble, clean, and govern in-nation datasets so every token of training data carries provenance an owner can defend.
Data enters the pipeline only after its legal basis and residency are established, not after the model is already learning from it.
Each source is registered with its licensing basis, consent terms, and permitted-use scope before ingestion. Sources without a defensible legal basis for training are rejected at the gate, not filtered downstream.
Domestic corpora — statute, regulatory filings, public records, institutional archives — are collected and staged inside sovereign infrastructure. Raw data does not leave the residency boundary to be processed elsewhere.
Financial, legal, and administrative sources are deliberately over-represented for models intended to reason over CBDC policy, settlement rules, and tokenized-asset structures. Corpus composition is a decision, recorded and reviewable, not an accident of scraping.
Sources are screened against known evaluation and benchmark sets so test material never leaks into training. This keeps later evaluation honest rather than inflated by memorization.
An owner can trace any behavior in the finished model back to the specific data that shaped it.
Every document and every derived shard is content-hashed and linked into the same tamper-evident hash chain used across our infrastructure. Altering a source after the fact breaks the chain and is detectable.
Cleaning, deduplication, filtering, and tokenization steps are each logged as an immutable record referencing the input and output hashes. The path from raw source to training shard is fully reconstructable.
Each training shard retains a manifest of which sources contributed to it and under what license. This supports takedown, license expiry, and audit questions long after training completes.
A dataset version is a fixed, hash-addressed snapshot. Re-running curation against the same snapshot yields the same corpus, so training runs are reproducible and disputes are settleable against a fixed artifact.
Quality work targets the failure modes that matter in sovereign finance rather than generic web hygiene.
Exact and near-duplicate documents are collapsed to prevent over-weighting boilerplate — repeated disclosures, standard clauses, template filings. This reduces memorization of high-frequency text and frees capacity for genuine signal.
Tables, schedules, and numeric fields in financial and legal documents are parsed with their structure intact rather than flattened into noise. A rate table stays a rate table, so the model learns the relationship, not a word salad.
Documents are scored for language quality, coherence, and domain relevance, and low-quality material is down-weighted or dropped. Thresholds are tuned per corpus and recorded so the filtering itself is auditable.
Content is screened for injected instructions, malformed encodings, and adversarial poisoning attempts before it can influence weights. Suspicious material is quarantined for review rather than silently trained on.
The state or institution that owns the model holds authority over what its data is allowed to become.
Personal and confidential fields are identified and either removed, redacted, or restricted to controlled corpora under explicit policy. Handling rules are defined by the owner and enforced in the pipeline, not left to operator discretion.
Curation runs inside the owner's residency boundary end to end. Provenance records prove that raw data was never processed or copied outside the sovereign environment.
When a license expires or consent is withdrawn, per-source attribution lets the affected data be located and its downstream shards flagged. Revocation is a traceable operation rather than a best-effort promise.
A dataset version is frozen only after a documented review of its composition, provenance, and governance posture. The frozen snapshot, not an informal folder of files, is what proceeds to training.