← Back to catalogue
Published

Context Assembly

vr.wm-mem-013 · wm-mem-013-context-assembly

Record the selected, ordered and transformed inputs prepared for one model invocation, their source lineage, budget accounting and visibility limits.

World Models Information and virtual systems INF.MEM.CTX

Bundle → Layer → Finding → Questions Filled

6 bundles · 12 layers · 12 findings · 48 questions

Identity and governing context Make the aggregate, invocation and external authority distinguishable.

Assembly identity and capture boundary

Identify a single invocation-bound aggregate and its revision. A prepared payload is not proof of dispatch or of the runtime effective input. Missing hidden material is recorded as unknown, never as empty.

Assembly identity and capture boundary

Identify a single invocation-bound aggregate and its revision. A prepared payload is not proof of dispatch or of the runtime effective input. Missing hidden material is recorded as unknown, never as empty.

  1. Which master identifier and revision distinguish this assembly from its invocation and earlier assemblies? identity
  2. At which capture boundary and observation level is the context represented? classification
  3. Which intended invocation or retry is bound to this assembly revision? relationship
  4. Which accountable role maintains the assembly and its disclosure decisions? ownership

Policy and authority bindings

Record version-pinned decision references for permitted input roles, purposes and recipients. Content priority is advisory; trusted status requires external authority. A conflicting or expired mandatory decision blocks sealing.

Policy and authority bindings

Record version-pinned decision references for permitted input roles, purposes and recipients. Content priority is advisory; trusted status requires external authority. A conflicting or expired mandatory decision blocks sealing.

  1. Which policy revision authorizes this assembly for its recipient and purpose? authority
  2. Which current grant permits each selected source representation to cross the declared boundary? access
  3. How is untrusted source content kept distinct from authorized instructions in this assembly? security
  4. What happens when mandatory policy bindings conflict or cannot be resolved? exception
Source provenance and time Keep the included source representation and its temporal qualifications traceable.

Source identity and representation

Bind every member to a source version or an explicitly unavailable provenance state, including direct user input and local templates. A URI is a locator, not immutable content. A digest identifies bytes under a declared method, not their trustworthiness.

Source identity and representation

Bind every member to a source version or an explicitly unavailable provenance state, including direct user input and local templates. A URI is a locator, not immutable content. A digest identifies bytes under a declared method, not their trustworthiness.

  1. Which source version or direct-input event produced each candidate representation? provenance
  2. Which exact portion and representation of a source was made available to the assembler? composition
  3. What evidence binds a captured representation to its claimed source? evidence
  4. How are unresolved selectors or changed source versions represented without silent substitution? validation

Temporal validity and freshness

Separate source effective time, source modification time, retrieval time and assembly observation time. Freshness is a policy judgment with a reference, not an inference from the latest timestamp. Unknown clock precision remains visible.

Temporal validity and freshness

Separate source effective time, source modification time, retrieval time and assembly observation time. Freshness is a policy judgment with a reference, not an inference from the latest timestamp. Unknown clock precision remains visible.

  1. Which source effective and observation times apply at the assembly cutoff? temporal
  2. Which external freshness decision permits a source whose age or validity is uncertain? quality
  3. Which source updates or grant revocations occurred between selection and sealing? event
  4. How is an unavailable or stale source handled when its content is required? exception
Selection and derivation Explain membership decisions and information changes without owning retrieval or durable memory.

Candidate inclusion and omission

Record the considered candidate set and its known completeness, then distinguish inclusion, exclusion, deferral and unavailable content. Ranking and deduplication reasons are inputs from a declared policy or retrieval event; they do not prove truth. Omission logs must not leak denied payloads.

Candidate inclusion and omission

Record the considered candidate set and its known completeness, then distinguish inclusion, exclusion, deferral and unavailable content. Ranking and deduplication reasons are inputs from a declared policy or retrieval event; they do not prove truth. Omission logs must not leak denied payloads.

  1. Why was each considered candidate included or excluded from this assembly? decision
  2. Which recall event or manual selection supplied the candidate set and ranking evidence? relationship
  3. Which required members and dependency groups must survive selection together? constraint
  4. Which omission details can be retained without revealing restricted source content? privacy

Transformation lineage and losses

Record assembly-local derivations such as excerpting, summarization, normalization and authorized redaction results. Preserve input-to-output lineage and disclosed loss. Summary content remains derived evidence and cannot silently replace its sources or acquire higher authority.

Transformation lineage and losses

Record assembly-local derivations such as excerpting, summarization, normalization and authorized redaction results. Preserve input-to-output lineage and disclosed loss. Summary content remains derived evidence and cannot silently replace its sources or acquire higher authority.

  1. Which ordered transformations produced the representation selected for inclusion? process
  2. Which source qualifications or contrary assertions were lost or preserved by a transformation? quality
  3. Which external decision authorizes using a redacted or quarantined derivative? security
  4. Can each included derivative be traced to all of its inputs without claiming deterministic regeneration? provenance
Ordered input representation Preserve semantic order, content forms and limits of visibility.

Order roles and dependency preservation

Record logical sequence and role mapping separately from the final serialization. Order is semantically significant. Tool declarations, calls and results are inert input records here; preserve their linkage and required grouping under the selected interface profile.

Order roles and dependency preservation

Record logical sequence and role mapping separately from the final serialization. Order is semantically significant. Tool declarations, calls and results are inert input records here; preserve their linkage and required grouping under the selected interface profile.

  1. What is the exact sequence of included members at the capture boundary? composition
  2. Which message or tool-related dependencies must the assembly preserve? requirement
  3. How does the target interface map logical roles and ordering into its representation? interoperability
  4. Which placement changes require a new revision and renewed authority checks? validation

Content forms and opaque state

Represent text, structured data, images, audio and externally referenced state with their own media and measurement basis. A description or transcript is a separate derivative. Hidden runtime additions and opaque cache handles constrain what the assembly can claim to reconstruct.

Content forms and opaque state

Represent text, structured data, images, audio and externally referenced state with their own media and measurement basis. A description or transcript is a separate derivative. Hidden runtime additions and opaque cache handles constrain what the assembly can claim to reconstruct.

  1. Which content forms and encodings occur in the assembled input? classification
  2. How are lengths or extents measured for each content form? measurement
  3. Which cached or remotely retained context components are referenced but not locally visible? state
  4. What limits prevent reconstructing the runtime input from this visible assembly? exception
Budget and admission Expose the counting basis and reasons for acceptance or refusal.

Budget profile and accounting basis

Use an adopting profile to define capacity, counting units and reserved output allowance. Record tokenizer and serializer versions, modality accounting and measurement uncertainty. Generation limits, input estimates, runtime reported usage and billed units are different quantities; no universal conversion is asserted.

Budget profile and accounting basis

Use an adopting profile to define capacity, counting units and reserved output allowance. Record tokenizer and serializer versions, modality accounting and measurement uncertainty. Generation limits, input estimates, runtime reported usage and billed units are different quantities; no universal conversion is asserted.

  1. Which capacity and output reserve apply to the declared target profile? constraint
  2. Which counting method produced each input estimate or measured count? measurement
  3. Which total reconciles member counts with wrappers and shared overhead without double counting? validation
  4. Which counts are estimated, runtime reported or billing-derived rather than directly measured? quality

Overflow and admission outcome

For compatible additive units under a declared profile, compare final input plus reserved output plus any overhead not already counted with capacity. Do not sum incomparable modalities or subtract cached tokens merely because billing differs. Recount after any transformation; unknown mandatory budgets block a fit claim.

Overflow and admission outcome

For compatible additive units under a declared profile, compare final input plus reserved output plus any overhead not already counted with capacity. Do not sum incomparable modalities or subtract cached tokens merely because billing differs. Recount after any transformation; unknown mandatory budgets block a fit claim.

  1. Does the final serialized representation fit every applicable budget constraint? validation
  2. Which authorized reduction or refusal follows an overflow? decision
  3. Which protected instructions or dependencies prohibit otherwise possible truncation? requirement
  4. Was the modified assembly recounted and rechecked before being marked ready? process
Sealing and controlled continuity Bind the final visible representation to evidence and govern its local copies.

Sealed revision and delivery evidence

Seal the observable assembly with a defined digest scope and immutable revision reference. Record dispatch and receipt only from external evidence. Runtime acceptance or token usage cannot prove all included members were attended to or used in an answer. A failed attempt remains a record.

Sealed revision and delivery evidence

Seal the observable assembly with a defined digest scope and immutable revision reference. Record dispatch and receipt only from external evidence. Runtime acceptance or token usage cannot prove all included members were attended to or used in an answer. A failed attempt remains a record.

  1. Which checks allow the assembly to move from prepared to sealed or rejected? lifecycle
  2. Which serialized artifact and integrity method identify the sealed input representation? evidence
  3. What external evidence distinguishes dispatch, acceptance and failure for this revision? event
  4. How is a retry or reused context linked without claiming equivalent runtime behavior? relationship

Disclosure retention and acceptance limits

Provide recipient-scoped local projections and assembly-copy disposal evidence without changing source masters. Reproducibility is bounded by available representations, pinned methods and permissions. Tombstones and digests can themselves be sensitive and require a retention basis.

Disclosure retention and acceptance limits

Provide recipient-scoped local projections and assembly-copy disposal evidence without changing source masters. Reproducibility is bounded by available representations, pinned methods and permissions. Tombstones and digests can themselves be sensitive and require a retention basis.

  1. Which policy governs retention and erasure of assembly payloads and their derived metadata? retention
  2. Which recipient projection can disclose lineage without exposing denied payloads or identifiers? access
  3. Which mapping and loss report accompany an exported assembly representation? interoperability
  4. Which adversarial acceptance cases and unresolved implementation gaps limit operational use? validation

Classifiers Filled

Family
World Models
Category
Information and virtual systems
Entry kind
aggregate
Navigation path
NAV.INF.MEM.CTX
Domain
INF.MEM.CTX
Industry
Cross-industry
Tags
contextassemblyinf.mem.ctx

What it is Filled

A versioned input aggregate with locally owned membership, placement, transformation and accounting records. A sealed revision describes the observable assembly at a declared capture boundary for one intended invocation. Preparation, dispatch and runtime-reported effective context are distinct assertions. An assembly may be rejected, empty under a permitted profile, partially observable or never dispatched. All hierarchy and operation details are proposed local design informed by the cited specifications, not requirements imposed by those specifications.

In scope

  • Assembly and revision identity, invocation binding, custody and referenced policy decisions
  • Candidate dispositions, source states, permitted transformations, ordered content blocks and representation details
  • Budget estimates and measurements, overflow outcomes, sealing evidence, recipient projections and retention of assembly copies

Out of scope

  • Memory store and item lifecycle, retrieval algorithms and index construction, durable memory consolidation
  • Policy definition or enforcement, trust scoring, access grants, redaction execution and incident management
  • Model execution, output quality, answer grounding, tool execution, hidden internal state reconstruction and universal token accounting

Why it exists Filled

Record the selected, ordered and transformed inputs prepared for one model invocation, their source lineage, budget accounting and visibility limits.

Distinguishing features Derived, awaiting review

  • Unlike WM-AI-004 AI Inference / Agent Run: Reference the intended invocation and observed dispatch receipt. Registry parent is not ratified subtype inheritance: an assembly is an input aggregate, not the execution, output, tool loop or cost ledger.
  • Unlike WM-MEM-002 Memory Item: Source memory keeps its identity and lifecycle. Assembly membership captures a version-qualified projection; copying it does not create new authoritative memory.
  • Unlike WM-MEM-007 Memory Recall Event: Reference retrieval candidate batches and ranks when present. This model records inclusion decisions for available candidates, not querying or ranking the corpus.
  • Unlike WM-MEM-008 Memory Consolidation: Transient assembly transformation records do not own durable merging, memory retirement or consolidation policy.
  • Unlike WM-MEM-009 Memory Conflict and Precedence Rule: Reference the rule and recorded conflict disposition. A similarity score, recency or slot position cannot establish authority or truth.
  • Unlike WM-MEM-010 Memory Retention and Forgetting Policy: Reference an applicable disposal policy for assembly copies; source-memory erasure remains externally executed.
  • Unlike WM-MEM-011 Memory Scope and Sharing Grant: Reference recipient, purpose and tenant authorization decisions; this model cannot grant access or infer it from readability.
  • Unlike WM-MEM-014 Retrieval Index: Index configuration and refresh remain external; snapshot and retrieval references may qualify a candidate batch.
  • Unlike WM-MEM-015 Retrieval Passage: Passage masters retain segmentation and source lineage. The assembly records only the included representation, selector and placement.
  • Unlike WM-MEM-016 Answer Grounding: Inclusion is not evidence of causal use or answer support. Claim-to-source grounding and answer evaluation remain external.
  • Unlike WM-MEM-017 Context Policy: Reference policy revision and decision evidence. Rule definition, precedence governance and enforcement remain external.
  • Unlike WM-MEM-018 Memory Trust Assessment: Reference a qualified assessment with its scope and validity, without generating a new trust score.
  • Unlike WM-MEM-020 Memory Quarantine and Redaction: Reference authorized redaction or quarantine outcomes and resulting representations. The assembly does not release quarantined memory or implement the redaction service.

Note: Derived from boundary notes against neighbouring models.

What robots and AI may and may not do Derived, awaiting review

Must not

  • Default deny across tenants, roles and destinations; source readability alone does not authorize inclusion or reuse.
  • Deny unless the current tenant, purpose, recipient and content-specific grant permit the requested view; recheck at sealing and later reuse.

May

  • Register an assembly revision: Proposed operation, not implemented. Record a new local aggregate for an intended invocation and capture boundary.
  • Record membership dispositions: Proposed operation, not implemented. Attach source-qualified candidate inclusion and omission evidence with minimum permitted metadata.
  • Record an assembly derivation: Proposed operation, not implemented. Link supplied transformation results to inputs with ordering and loss evidence.
  • Assess a declared assembly budget: Proposed operation, not implemented. Apply only a pinned local accounting profile to supplied final-representation measurements; refuse a fit claim for missing or incomparable mandatory values.
  • Seal an observable assembly: Proposed operation, not implemented. Validate local completeness, dependencies, current authorization evidence and budget assessment before binding a revision to supplied integrity evidence.
  • Prepare a recipient projection: Proposed operation, not implemented. Create a local authorized view with redaction, mapping losses and reconstruction limits.

Note: Derived from functions, policies, CRUD and access rules; prohibitions were not authored for agents as such.

Moral aspects Missing, in the backlog

Not described yet. This gap is in the card backlog.

Owners Filled

Steward

Accountable assembly custodian role and tenant namespace

Roles

Assembly custodian
Maintains identity, declared scope and retention decisions within delegated authority.
Assembly preparer
Records permitted source membership, placement and supplied derivation evidence.
Policy decision authority
Owns external authorization and exception decisions; cannot infer authority from rank or context position.
Validation reviewer
Reviews budget accounting, lineage and sealed revision evidence; reports limits and blocked checks.
Authorized consumer
Uses only the permitted projection and preserves uncertainty and source-master boundaries.

Links to other meta-models Filled

references

  • WM-AI-004 - Candidate binding, not a mandatory runtime dependency. Reference the intended invocation and observed dispatch receipt. Registry parent is not ratified subtype inheritance: an assembly is an input aggregate, not the execution, output, tool loop or cost ledger.
  • WM-MEM-002 - Candidate binding, not a mandatory runtime dependency. Source memory keeps its identity and lifecycle. Assembly membership captures a version-qualified projection; copying it does not create new authoritative memory.
  • WM-MEM-007 - Candidate binding, not a mandatory runtime dependency. Reference retrieval candidate batches and ranks when present. This model records inclusion decisions for available candidates, not querying or ranking the corpus.
  • WM-MEM-008 - Candidate binding, not a mandatory runtime dependency. Transient assembly transformation records do not own durable merging, memory retirement or consolidation policy.
  • WM-MEM-009 - Candidate binding, not a mandatory runtime dependency. Reference the rule and recorded conflict disposition. A similarity score, recency or slot position cannot establish authority or truth.
  • WM-MEM-010 - Candidate binding, not a mandatory runtime dependency. Reference an applicable disposal policy for assembly copies; source-memory erasure remains externally executed.
  • WM-MEM-011 - Candidate binding, not a mandatory runtime dependency. Reference recipient, purpose and tenant authorization decisions; this model cannot grant access or infer it from readability.
  • WM-MEM-014 - Candidate binding, not a mandatory runtime dependency. Index configuration and refresh remain external; snapshot and retrieval references may qualify a candidate batch.
  • WM-MEM-015 - Candidate binding, not a mandatory runtime dependency. Passage masters retain segmentation and source lineage. The assembly records only the included representation, selector and placement.
  • WM-MEM-016 - Candidate binding, not a mandatory runtime dependency. Inclusion is not evidence of causal use or answer support. Claim-to-source grounding and answer evaluation remain external.
  • WM-MEM-017 - Candidate binding, not a mandatory runtime dependency. Reference policy revision and decision evidence. Rule definition, precedence governance and enforcement remain external.
  • WM-MEM-018 - Candidate binding, not a mandatory runtime dependency. Reference a qualified assessment with its scope and validity, without generating a new trust score.
  • WM-MEM-020 - Candidate binding, not a mandatory runtime dependency. Reference authorized redaction or quarantine outcomes and resulting representations. The assembly does not release quarantined memory or implement the redaction service.

aligned

  • MCP 2025-06-18 - Conceptual resources and sampling mapping; not a required transport or complete context audit format.
  • PROV-O 2013-04-30 - Conceptual derivation and attribution mapping; no truth or authorization inference.
  • Web Annotation Data Model 2017-02-23 - Optional source selectors with version-qualified representation states.
  • RFC 8785 - Optional JSON digest profile only when its data constraints are met; other bindings declare their own byte rules.
  • OpenTelemetry GenAI client inference - Development telemetry alignment; pin an immutable revision before implementing a mapping.

neighbor

  • WM-AI-004 AI Inference / Agent Run - Reference the intended invocation and observed dispatch receipt. Registry parent is not ratified subtype inheritance: an assembly is an input aggregate, not the execution, output, tool loop or cost ledger.
  • WM-MEM-002 Memory Item - Source memory keeps its identity and lifecycle. Assembly membership captures a version-qualified projection; copying it does not create new authoritative memory.
  • WM-MEM-007 Memory Recall Event - Reference retrieval candidate batches and ranks when present. This model records inclusion decisions for available candidates, not querying or ranking the corpus.
  • WM-MEM-008 Memory Consolidation - Transient assembly transformation records do not own durable merging, memory retirement or consolidation policy.
  • WM-MEM-009 Memory Conflict and Precedence Rule - Reference the rule and recorded conflict disposition. A similarity score, recency or slot position cannot establish authority or truth.
  • WM-MEM-010 Memory Retention and Forgetting Policy - Reference an applicable disposal policy for assembly copies; source-memory erasure remains externally executed.
  • WM-MEM-011 Memory Scope and Sharing Grant - Reference recipient, purpose and tenant authorization decisions; this model cannot grant access or infer it from readability.
  • WM-MEM-014 Retrieval Index - Index configuration and refresh remain external; snapshot and retrieval references may qualify a candidate batch.
  • WM-MEM-015 Retrieval Passage - Passage masters retain segmentation and source lineage. The assembly records only the included representation, selector and placement.
  • WM-MEM-016 Answer Grounding - Inclusion is not evidence of causal use or answer support. Claim-to-source grounding and answer evaluation remain external.
  • WM-MEM-017 Context Policy - Reference policy revision and decision evidence. Rule definition, precedence governance and enforcement remain external.
  • WM-MEM-018 Memory Trust Assessment - Reference a qualified assessment with its scope and validity, without generating a new trust score.
  • WM-MEM-020 Memory Quarantine and Redaction - Reference authorized redaction or quarantine outcomes and resulting representations. The assembly does not release quarantined memory or implement the redaction service.

parent

  • WM-AI-004

What else AI and robots need to interact with it Incomplete

Identity and identifiers required Filled

  • Authoritative master-system identifier
  • Governed global identifier or IRI
  • UUID or ULID allocated by the adopting Dimension

Direct properties not applicable Not applicable

Not applicable

Institutional or informational subject: no invented physical properties.

Recognition optional Missing, in the backlog

Not described yet. This gap is in the card backlog.

Capabilities and actions required Filled

  • Register an assembly revision: Proposed operation, not implemented. Record a new local aggregate for an intended invocation and capture boundary.
  • Record membership dispositions: Proposed operation, not implemented. Attach source-qualified candidate inclusion and omission evidence with minimum permitted metadata.
  • Record an assembly derivation: Proposed operation, not implemented. Link supplied transformation results to inputs with ordering and loss evidence.
  • Assess a declared assembly budget: Proposed operation, not implemented. Apply only a pinned local accounting profile to supplied final-representation measurements; refuse a fit claim for missing or incomparable mandatory values.
  • Seal an observable assembly: Proposed operation, not implemented. Validate local completeness, dependencies, current authorization evidence and budget assessment before binding a revision to supplied integrity evidence.
  • Prepare a recipient projection: Proposed operation, not implemented. Create a local authorized view with redaction, mapping losses and reconstruction limits.

Hazards and failure modes optional Missing, in the backlog

Not described yet. This gap is in the card backlog.

Standards and interfaces required Derived, awaiting review

  • PROV-O: The PROV Ontology
  • RFC 3339: Date and Time on the Internet: Timestamps
  • RFC 8785: JSON Canonicalization Scheme

Context of use required Filled

  • No universal retention period, jurisdictional entitlement, language encoding or model capacity is assumed; adopting profiles must be reviewed.
  • This is a technical context record, not a compliance certificate or permission to process sensitive data.

Sources Filled

  1. MCP Resources - MCP
  2. MCP Sampling - MCP
  3. PROV-O: The PROV Ontology - W3C
  4. Web Annotation Data Model - W3C
  5. LLM Prompt Injection Prevention Cheat Sheet - OWASP
  6. RFC 3339: Date and Time on the Internet: Timestamps - IETF
  7. RFC 8785: JSON Canonicalization Scheme - RFC Editor
  8. OpenTelemetry semantic conventions for generative AI client inference - OpenTelemetry

Open questions

  • Run the coordinator source checker outside the sandbox, verify relevant document sections and errata, and pin immutable revisions for mutable guidance before implementing mappings.
  • Develop executable nested schemas and profile fixtures for stale grants, changed sources, hidden context, multimodal counts, cached subsets, unsafe truncation, permitted empty assemblies and copy disposal.
  • Restore independent external review and qualify adopting profiles before any canonical or publishable-draft promotion.
  • Executable nested instance schemas, target-specific tokenizers and multimodal counting adapters
  • Immutable revision pins for living OWASP and OpenTelemetry documents and comprehensive errata review
  • Operational acceptance fixtures for opaque state, changing grants, empty context, truncation and nondeterministic transformations
  • Hidden runtime context and causal use of inputs cannot be reconstructed from this local record alone

Machine files

Provenance

world-models research · reviewable-draft

Built from: models/wm-mem-013-context-assembly/spec.yaml