AI foundations

When an AI skill drafts a description, proposes an entity, or reviews a model for you, it does not improvise its standards on the spot. It reasons from foundations: curated reference documents the AI cites to ground its reasoning. A foundation lives apart from any one skill. It captures a body of guidance once (how to model the business, how to handle personal data, how to name things), a platform administrator reviews and publishes it, and from then on every skill that cites it works from that same text.

That separation is the point. Your modeling standards live in one governed place instead of being restated, slightly differently, inside every skill. Two people running the same skill on the same catalog get proposals that read like one team wrote them, and when a standard changes you change it once. This page lists the nine foundations that ship with DeltaVault, explains how a foundation reaches a run, and shows platform administrators how to author and publish one.

The shipped foundations

DeltaVault ships nine foundations. House Style is the always-on baseline: it ships already published and applies to every run. The other eight are topical: each ships as a draft, and an administrator publishes it before it grounds any skill.

House Style: the always-on baseline

House Style carries the platform-wide voice and guardrails every skill inherits. It ships already published and is folded into every run automatically, whether or not the skill cites it, so no run is ever ungoverned.

The eight topical foundations

Each topical foundation covers one body of modeling or governance guidance. An administrator reviews and publishes it before it grounds any skill; until then, skills that cite it simply run without it.

• Business Modeling. Model the business, not the source: source tables, files, and pipelines are evidence, never definitions. Every name and definition must read as vocabulary a domain expert with no technical background understands, and nothing is invented that the business context does not support.
• PII Handling. Treat personal data as sensitive by default: identify and tag the columns that carry it, and flag uncertain fields for review rather than assuming they are safe. Prefer masking, tokenization, or aggregation over exposing raw identifiers in analytical layers.
• Medallion Architecture. Refine data through progressive layers: Bronze lands raw source data with fidelity and history preserved, Silver deduplicates, validates, and conforms it, and Gold curates business-level aggregates for consumption. Each layer reads from the one below it.
• Slowly Changing Dimensions. Choose deliberately, per dimension attribute, how change is tracked over time: Type 1 overwrites when only the current state matters, Type 2 adds dated history rows when point-in-time history is required, and Type 3 keeps a limited prior value for previous-versus-current comparison. The chosen type is stated, never implied.
• Data Vault Modeling. Model stable core concepts and instantiate them as Data Vault 2.0: identity becomes a hub, relationships and units of work become links, and time-variant context becomes satellites, loaded insert-only. It is self-contained, carrying the logical modeling rules it depends on, so it needs no second foundation paired with it.
• Dimensional Modeling. Design analytical marts as facts and dimensions with a declared grain: every fact states its grain as one business sentence, every measure states how it aggregates, and shared concepts use conformed dimensions rather than one copy per mart. Like Data Vault Modeling, it is self-contained.
• Data Quality & Profiling. Assess data quality along its named dimensions (completeness, validity, uniqueness, consistency, timeliness, and accuracy) and ground every claim in profiling evidence: row counts, distinct counts, null ratios, ranges, and value patterns. Every finding carries its evidence and a severity, and a definite defect is distinguished from a candidate anomaly.
• Glossary & Naming Standards. Define business terms and propose names as one coherent, governed vocabulary: one-sentence definitions that disambiguate, singular business-language names free of implementation suffixes, and relationships named as present-tense verb phrases. Synonyms are flagged and resolved to one canonical term, never silently duplicated.

How foundations ground a skill

Each skill declares which foundations it cites. The citation is part of the skill itself, so you never pick foundations per run. When you run the skill, the cited foundations that are published are resolved and injected into the skill’s prompt, ahead of the catalog context the run gathers and the inputs you type. House Style joins every run whether the skill cites it or not.

Publishing is the switch. A cited foundation that is still a draft is skipped, and an archived one drops out of every future run: only published text ever grounds a skill. The flip side is just as direct: publish a revision and every skill that cites the foundation reads the new text on its next run, with no change to the skills themselves.

The result is one shared, governed source of reasoning. A generation skill and a validation skill that both cite Data Vault Modeling apply the same hub, link, and satellite rules; a skill that cites Glossary & Naming Standards proposes names your glossary would approve. For every skill you can launch and the stage of work each one serves, see the AI skills catalog.

Author and publish a foundation (administrators)

Foundations are managed on the AI Foundations page in the platform administration area. The page is restricted to platform administrators; if you do not hold that role, you will not see it. It lists every foundation with its status (draft, published, or archived) next to the editor. House Style already exists, seeded and published, so the flow below is for the eight topical foundations and any foundation you add yourself.

1. Create. Select the create control; the Create new foundation dialog asks for a foundation key (lower-case words joined by hyphens, unique across foundations) and a label. Creating saves the foundation as an empty draft and opens it in the editor.
2. Author. Write the body: it is the exact text injected into skill runs, so write direct guidance in plain language (principles and rules, not a long document). Give the foundation a label and a short description, then select Save draft. A draft grounds nothing, so you can save as often as you like.
3. Publish. Select the publish button, which names the version it is about to cut, so it reads Publish v1 the first time. Publishing requires a non-empty body: until the label and body are filled in, the button stays disabled and lists what is missing. Publishing snapshots a numbered version and sets the foundation live; every skill that cites it reads the new text on its next run.

Every publish is versioned. To revise a published foundation, edit and save it (it returns to draft status) and publish again. To roll back, restore a previous published version as a new draft and publish that. Archiving removes a foundation from every future run until you restore and publish it again.