Enterprise & institutions is Naevum's Tier 3 stack for organizations that must explain their decisions, so that every answer shows the rule, the date and the person behind it.
Append source-connected claims, episodes, lessons and measured failures as distinct records.
Attributed original or revised commitment.
Append source-connected claims, episodes, lessons and measured failures as distinct records.
Source, holder, observation time and record identity.
Supersession, contest and retraction retain the earlier record.
Revision, contest and retraction preserve the earlier source history.
Resolve validity, knowledge time, scope and conflicting authority for the requested question.
Scope, validity time and knowledge cutoff.
Claims and operations with source ancestry.
Resolve validity, knowledge time, scope and conflicting authority for the requested question.
Event validity and recorded-at cutoff.
Authority ancestry and unresolved conflicting claims.
Both the validity question and the knowledge cutoff govern which records apply.
Find relevant evidence and assemble a bounded source-bearing answer.
Applicable, contested or unavailable evidence.
Rebuilt index/packs and retained failure context.
Find relevant evidence and assemble a bounded source-bearing answer.
Canonical records remain the historical basis.
Item/byte bounds, source pointers and abstention reasons.
The answer stays within the selected scope and may abstain when evidence is inadequate.
Integrate eligible change, retain standing causes and distinguish blindness from resolution.
Eligible changed source, freshness and provenance.
Integrate eligible change, retain standing causes and distinguish blindness from resolution.
Charge, surprise, decay and evidence trust.
New, standing, stale and cleared states.
Stale or missing sensing does not clear a standing issue.
Compare a stored estimate with a later identified outcome and retain uncertainty.
Defined target and context known before execution.
Independent outcome under the stored task identity.
Compare a stored estimate with a later identified outcome and retain uncertainty.
Defined target, task identity and resource estimate.
Independent outcome, unknown endings and eligible-choice advice.
Outcome advice remains separate from permission to act.
Rebuild derived context and retain useful failures without replacing source history.
Canonical records and their integrity state.
Attributed retrieval use, not proof of future utility.
Rebuild derived context and retain useful failures without replacing source history.
Duplicate/conflict annotations and derived packs.
Measured harms and reasons for caution or release.
Derived indexes can be rebuilt while canonical history remains intact.
Attributed original or revised commitment.
Append source-connected claims, episodes, lessons and measured failures as distinct records.
Source, holder, observation time and record identity.
Supersession, contest and retraction retain the earlier record.
Revision, contest and retraction preserve the earlier source history.
Scope, validity time and knowledge cutoff.
Claims and operations with source ancestry.
Resolve validity, knowledge time, scope and conflicting authority for the requested question.
Event validity and recorded-at cutoff.
Authority ancestry and unresolved conflicting claims.
Both the validity question and the knowledge cutoff govern which records apply.
Applicable, contested or unavailable evidence.
Rebuilt index/packs and retained failure context.
Find relevant evidence and assemble a bounded source-bearing answer.
Canonical records remain the historical basis.
Item/byte bounds, source pointers and abstention reasons.
The answer stays within the selected scope and may abstain when evidence is inadequate.
Eligible changed source, freshness and provenance.
Integrate eligible change, retain standing causes and distinguish blindness from resolution.
Charge, surprise, decay and evidence trust.
New, standing, stale and cleared states.
Stale or missing sensing does not clear a standing issue.
Defined target and context known before execution.
Independent outcome under the stored task identity.
Compare a stored estimate with a later identified outcome and retain uncertainty.
Defined target, task identity and resource estimate.
Independent outcome, unknown endings and eligible-choice advice.
Outcome advice remains separate from permission to act.
Canonical records and their integrity state.
Attributed retrieval use, not proof of future utility.
Rebuild derived context and retain useful failures without replacing source history.
Duplicate/conflict annotations and derived packs.
Measured harms and reasons for caution or release.
Derived indexes can be rebuilt while canonical history remains intact.
An illustrative process for source-grounded recall, attention and outcome comparison.
Additional mechanisms are described in the technical reference.
What the research says
The AI Act gives people affected by certain high-risk decisions a right to an explanation, and the Dutch fine shows what an unexplained risk rule costs an institution. Assistants also lose recall over long interactions, so Vagus keeps each decision with its date, authority and sources, and the explanation is read from the record.
An institution asks why a commitment changed, which policy applied then and who owns the next decision.
A source-connected memory can retain current and historical claims, disagreements and the evidence behind them. An attention layer can surface new conditions that make an earlier decision worth revisiting.
Useful institutional intelligence depends on roles, custody and the actual workflow.
Specify who may see which sources and how current those sources need to be.
Select the context needed for a task and account for its destination and history.
Present the evidence, uncertainty and permitted next action to the responsible person.
An agency director decides whether to reopen a decision after its rule changed; memory supplies what was true then and what changed since. Illustrative.
Systems behind it
Institutional memory & attention
Vagus finds a past decision with its date and authority, and flags it when a later change touches its basis.
Private computation & model orchestration
Synapse lets staff use outside models on case work while each provider receives only what the task needs.
Governed application foundation
Agency Engine gives the institution's app its roles, approvals and record, so every answer has an owner.
Start with recurring institutional questions and decisions. Test correct source retrieval, temporal accuracy, useful attention and successful completion of the intended workflow.
Map access. Specify who may see which sources and how current those sources need to be.
Scope model context. Select the context needed for a task and account for its destination and history.
Return an accountable answer. Present the evidence, uncertainty and permitted next action to the responsible person.
Map access. Specify who may see which sources and how current those sources need to be.
Scope model context. Select the context needed for a task and account for its destination and history.
Return an accountable answer. Present the evidence, uncertainty and permitted next action to the responsible person.
An illustrative institutional question that needs both history and controlled model context.
Follow the storyWhich decision would your institution struggle to explain if the person it affected asked tomorrow?
Discuss an institutional system