CBE asks what an already defined capability still needs to perform its own responsibility. It separates legitimate internal expansion from work owned by neighboring capabilities, then makes the boundary and handoffs explicit.
A useful nearby feature is not automatically part of the target's responsibility. Every candidate must trace to an internal gap and survive relevance, ownership and identity checks.
Unknown ownership stays unresolved until evidence or an architectural decision can settle it.
| Verdict | Use when | Design result |
|---|---|---|
| KEEP | The responsibility is needed and belongs to the target. | Add it to the target's design. |
| SPLIT | The candidate mixes target-owned work with portions that belong elsewhere or are unnecessary. | Separate meaningful parts and classify each: KEEP the target portion, DELEGATE peer-owned work or REJECT unnecessary scope. |
| DELEGATE | The target needs the result, but a peer owns the operation. | Define a handoff and expected return. |
| REJECT | The candidate does not close an internal gap or changes target identity. | Record the reason for excluding it. |
UNRESOLVED records insufficient evidence or an ownership question. It is not a fifth acceptance verdict. Using a peer's output does not mean owning that peer's operation.
Responsibility invariant: accept documents and produce normalized ingestible records. Preserve the existing steps: accept, extract, normalize and submit downstream. The Semantic Indexer owns embeddings; the Retention Policy Service owns retention decisions; the Malware Scanner owns malicious-content detection.
| Gap / candidate | Decision | Reason or handoff |
|---|---|---|
| Accepted formats are not distinguished: detect document format. | KEEP | Format detection supports the pipeline's own extraction step. |
| An accepted format needs an extraction handler: select its parser. | KEEP | Use the supplied format-to-parser mapping; the pipeline owns handler selection. |
| Repeated submissions are not distinguished: perform ingestion-level duplicate detection. | KEEP | Compare the supplied document identifier with the provided ingestion record set. Record new or duplicate and apply the supplied submission rule. |
| Submission needs a scanner outcome: coordinate the safety gate. | KEEP | Request a scan, preserve the returned outcome and hold downstream submission unless the supplied gate condition is met. |
| Determine whether document content is malicious: implement the malware detection algorithm. | DELEGATE | The Malware Scanner owns malicious-content detection. Coordinating the handoff does not transfer its algorithm to ingestion. |
| Downstream search needs vectors: generate embeddings. | DELEGATE | Send normalized content to the Semantic Indexer; consume its return. |
| Records need retention metadata: determine policy and attach classification. | SPLIT | DELEGATE policy decisions; KEEP attaching the returned classification. |
| Add a conversational assistant for readers. | REJECT | No demonstrated gap in ingestion requires a reader assistant. |
| Add regional access decisions; ownership is not supplied. | UNRESOLVED | Clarify the responsibility boundary before assigning a verdict. |
Owns: attaching the supplied retention classification to a normalized record.
Does not own: defining retention policy or deciding how long records must be retained.
Input: normalized record plus a classification returned by the policy service.
Output: enriched record, or an explicit unresolved classification.
Trigger: a record needs retention classification.
Destination: Retention Policy Service.
Target supplies: the agreed record metadata.
Destination owns: the policy decision.
Expected return: classification or unresolved result. The pipeline attaches the return or preserves the unresolved state.
Trigger: an accepted document needs a safety result before downstream submission.
Destination: Malware Scanner.
Target supplies: the document and its correlation identifier.
Destination owns: the malware detection algorithm.
Expected return in this fixture: clear, flagged or unavailable, tied to that identifier.
Target behavior after return: preserve the result. Only clear permits the safety gate to release submission; flagged or unavailable holds it. A clear result is the scanner's answer, not proof of universal safety.
The fixture supplies document identifiers and an ingestion record set. Its submission rule allows a new identifier to proceed to the safety gate; a duplicate is recorded and held without rewriting the prior record. Ingestion owns comparing these inputs and recording the outcome; it does not invent a retention decision.
The safety gate owns requesting and consuming a scanner result. It does not inspect content to decide whether it is malicious. Duplicate outcomes and the three supplied scanner returns are reviewed in I1; their handling is locally exhausted in this finite fixture.
The identifier comparison, supplied duplicate rule and clear/flagged/unavailable return contract are teaching assumptions added to the source's responsibility example. They are not claimed as prescribed implementation details.
This bounded example also retains parser selection alongside format detection. It is the I1 snapshot replayed in CBE-IX. The next pass examines previously unspecified ambiguous parser matches and conflicting classification returns; it does not rediscover parser selection or the unresolved-classification state.
The candidate verdict changes the design representation. It does not edit the pipeline, call the peer or authorize deployment.
KEEP, SPLIT, DELEGATE and REJECT classify individual candidates. The source's output contract, section L, requires one final verdict for the complete analysis, alongside its verification audit and residual uncertainty.
The source requires the V1-V14 audit. A FAIL on V1 (responsibility invariant), V5 (KEEP owning a peer responsibility) or V10 (target identity) makes the expansion unacceptable as-is.
When unresolved ownership prevents further iteration without peer takeover, CBE's BOUNDARY_UNRESOLVED corresponds to CBE-IX's BLOCKED_BY_BOUNDARY_DECISION. An unresolved candidate alone does not automatically establish that iterative blocker.
The classification lab below assesses one candidate at a time. It does not calculate or certify a final CBE verdict for an entire analysis.
Choose a candidate from the ingestion example, make a prediction and inspect the curated explanation. This exercise uses declared boundaries; it does not infer ownership from arbitrary text.
The target and peer responsibilities stay fixed throughout this exercise.
Capability-Gap and CSR establish what exists and where a need is already covered. CBE then examines the selected target's internal completeness. SORR can resolve ambiguous identities and responsibility claims.
Dialogue Lifecycle controls transitions. 5PP Variation A structures the scoped work and audit. AICS can represent the resulting requirements and decisions. Analysis authorization and implementation authorization remain separate.
Use CBE-IX when an accepted extension may expose further legitimate internal gaps. It adds an iteration ledger, a frontier of unprocessed work and explicit convergence checks.
See bounded capability design in the solution atlas. Practice the authority distinction in the Blueprint reader labs.
Derived from the author-provided Capability Boundary Expander - AI Execution Prompt, updated 24 August 2026. The ingestion example adapts its Example 2; the reader-assistant and unknown-owner cases are additional teaching examples.
Source mapping and reproducibility notes identify the exact supplied source snapshot. The profile explains a design method; the lab demonstrates curated cases and does not establish general semantic correctness, runtime availability or measured effectiveness.