Make enterprise information usable by AI.Keep authority with the enterprise.

Collections make existing enterprise information available to cognition as governed information domains—without moving source-of-record authority into a model, retrieval layer, or new system of record.

Your data estate remains authoritative.

Indwel works with the systems the enterprise already trusts. It does not ask them to become an AI platform.

Across operational systems, analytical platforms, content repositories, catalogs, and search infrastructure, the enterprise already has sources it trusts for different kinds of truth. Collections add a governed boundary between that information estate and machine cognition.

A Collection publishes a bounded information domain: which sources govern which propositions, what the domain contains, how it is structured, which operations are valid, how it may relate to other domains, and what must happen before its contents can support a claim, answer, or decision.

The original system continues to own its records and their meaning, even when Indwel builds indexes, embeddings, extracted structures, or semantic representations to make those records useful to cognition. Those supporting mechanisms can be rebuilt or replaced without quietly becoming another system of record.

That separation is central to Indwel. Storage can change, retrieval can improve, and models can be replaced while the meaning, ownership, provenance, and institutional standing of enterprise information remain explicit.

A Collection publishes a governed domain.

Every published Collection has a durable identity, a responsible owner, an organizational and access scope, a source history, and an immutable publication version. Its Sovereign Collection Contract specifies the governed information domain.

That specification establishes the domain's native units—the records, transactions, document sections, events, observations, or other structures that carry meaning—and gives them stable identities that survive retrieval, Evidence, presentation, replay, and audit.

It also declares what the domain can do. Exact selection, bounded ranges, complete enumeration, comparison, aggregation, temporal analysis, semantic retrieval, and predefined derived views can be published where appropriate. Exact operations remain exact; derived views have declared meaning rather than becoming free-form model transformations.

Those capabilities are compiled, validated, and granted before execution. A model may choose among capabilities authorized for the current Work, but it does not invent selectors, transformations, joins, access rights, or Evidence authority at runtime.

The Sovereign Collection Contract governs the information domain, not the undertaking that uses it. Authority over consequential Work remains in the applicable Cognitive Contract. This separation turns enterprise information into an executable resource while keeping domain definition and Work authority outside inference.

From source to admitted evidence

Collections and EvidenceSources and Library

supports

Collections and EvidenceSovereign Collections

exposes

Collections and EvidenceNative Selector Families

supplies candidates

Collections and EvidenceEvidence Firewall

admits evidence

Cognitive MachineryCognition Kernel

Inspect the Collections and Evidence boundary

Govern everyday information without creating a data-engineering project.

Collections are intended to make governed information practical at enterprise scale. Documents, presentations, spreadsheets, delimited data, text, and other bounded business material can enter through a low-friction authoring path. Authorized users and developers bind sources, review the structure Indwel discovers, refine it where necessary, qualify the result, and publish an immutable version.

Indwel can assist with structure, field meaning, retrieval hints, and other authoring decisions, but assistance does not create authority. Suggested meaning becomes part of the Collection only after it has been reviewed and validated.

Specialized Collections extend the same architecture for domains that need richer native models, purpose-built selection and analysis, predefined derived views, specialized source or storage bindings, explicit correspondence with other Collections, domain-specific presentation, or stronger certification requirements.

The standard does not weaken as authoring becomes easier. A simple Collection is simple to create, not lightly governed. Identity, provenance, access, Evidence, publication, lifecycle, and operational control remain part of the same Sovereign Collection Contract.

Nor does a Collection require one storage pattern. Information may remain in its authoritative system, be materialized into an Indwel-managed representation, or participate through another supported binding. The implementation can change without changing which source owns the underlying truth.

Retrieval is not Evidence.

Finding information and establishing that the institution may rely on it are different acts.

A Collection can use the retrieval method that fits the domain—deterministic selection, traditional search, semantic or vector retrieval, hybrid strategies, or another qualified method—without allowing the retrieval mechanism to redefine the request.

Similarity may be right for finding relevant material. It is not a substitute for an exact record, a complete population, a defined time window, or a known denominator. Where exactness or completeness matters, the Collection preserves it. Candidate discovery, Evidence qualification, and model context remain separate concerns; a context window is a computational limit, not the boundary of institutional knowledge.

When retrieved material is proposed as a factual basis for reasoning, it crosses the Indwel Evidence boundary under the governing Cognitive Contract. Source identity, Collection publication, lineage, rights, scope, freshness, structural standing, and applicable policy remain available to the Evidence Firewall when admission is decided.

A citation can show where material came from. Evidence must also establish why the institution is entitled to rely on it for the proposition being made. Collections preserve that distinction so admitted Evidence remains inspectable, attributable, replayable, and auditable after an answer or Work outcome has been produced.

Use information together without erasing its boundaries.

Important business questions rarely live inside one system. A decision may depend on transactions, contracts, correspondence, operating policy, analytical data, and prior Work, each governed by a different owner and carrying a different kind of authority.

Collections can participate together without collapsing those domains into a synthetic source of truth. Cross-Collection correspondence declares how domains relate, what may correspond, whether the relationship is one-to-one, one-to-many, or otherwise bounded, what happens when either side is absent or ambiguous, which source remains authoritative for which proposition, and what claims the combined view may support.

The relationship is governed just as the underlying domains are governed. A model can use correspondence that has been declared and authorized; it cannot create an ad hoc join and then promote the association into an institutional fact.

Results preserve provenance back to the relevant Collection publications and native units. The enterprise can therefore federate cognition across a heterogeneous information estate while retaining the boundaries that made each domain trustworthy in the first place.

Collections are production assets.

A Collection is not finished when content has been indexed. Before publication, Indwel can qualify the domain definition, source custody, native structure, supported selectors and derived views, Evidence lineage, access boundaries, security posture, and operating behavior the publication claims.

Certification attaches to the exact revision and source state that were qualified, and published versions are immutable. Material changes to sources, interpretation, operations, Evidence requirements, or meaning-bearing machinery can require requalification rather than inheriting an earlier certification by assumption.

A new publication can supersede an earlier one without destroying history. New uses resolve against the publication currently in force. Existing Work remains attributable to the exact Collection publication it actually used; moving that Work to a later publication is a governed dependency change, never silent substitution. Withdrawal or revocation can change what is available for new Work without making prior Evidence unintelligible.

Production operation continues through access control, source rights, usage limits, audit, security, freshness and drift, service and performance expectations, and diagnostics. Operators can see which publication is active, what sources it covers, whether it is current, which capabilities are certified, and where a capability has degraded or failed.

Refresh and reingestion are resumable and idempotent. Migration, backup, restore, supersession, withdrawal, rollback, and forward repair are governed operations, preserving continuity as the underlying estate changes.

If an operation cannot be satisfied within the Collection's declared authority or operating limits, Indwel can clarify the request, use a truthful supported path, or decline it. The system does not widen the domain, weaken the Evidence standard, or manufacture authority to make a request succeed.

One governed identity everywhere it is used.

A Collection does not become a different object when it moves from one Indwel surface to another. The same platform-owned identity, publication history, capabilities, source bindings, and Evidence lineage serve Unity, durable Work, the Platform API, and applications built through the Business Application Engine.

People can work with a Collection directly, long-running Work can depend on it, developers can address it through stable platform resources, and business applications can inherit its capabilities without building a separate retrieval, provenance, and Evidence regime.

The Sovereign Collection Contract also supplies the stable technical surface around that identity: capability discovery, API representation, generated types, application bindings, certification status, and governed presentation. Different teams can work at different layers of the information estate without creating different definitions of the domain.

They are all operating against the same governed information domain.

Keep the systems that run the business.

Collections let AI use enterprise information without becoming the authority over it. Systems of record keep ownership; the published Collection governs use; Evidence retains lineage; and people, Work, APIs, and applications operate against the same institutional definition. Models and retrieval systems can change without changing who owns the information, what it means, or what the institution is entitled to treat as true.

Discuss your information estate See the architecture