IWE reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

The four reports connect within the repository-shipped knowledge and agent-memory system fixed by boundary. They distinguish deterministic document/bookkeeping services from external host judgment and scheduling. No supplied report establishes an observed end-to-end capture, actual host delivery or causal benefit. These shared limits do not negate the implemented persistence and later-consumer paths.

Duplicate supersessions and distinct identities

Amendment: EPI-OBJ-proposal is superseded by MEM-OBJ-proposal. Both identify the same Proposal portion of RT-OBJ-session-record: natural-language title/body/evidence with symbolic key, update and status fields, serialized outside the graph. Their supporting passages from SRC-1, crates/iwe/src/internal/claude/session.rs, describe complementary checks and staging of that same object, not different candidates or storage layers. The epistemic inventory and capture-ledger references therefore concern the memory-declared proposal as well. The supersession changes no finding: nonempty evidence is not authenticated evidence, and durable candidate retention is not user selection or factual acceptance. Both IDs remain declared.

Amendment: EPI-OBJ-policy-body is superseded by MEM-OBJ-policy-body. Both identify the natural-language body of MEMORY, with named capture, writing, dedup, provenance, curation and session-start sections interpreted by humans/models. The former describes this payload within the loaded graph; the latter describes its filesystem storage and section extraction. Neither record identifies a separate policy instance. Identity evidence is their common MEMORY body referent, supported by SRC-1, crates/iwe/templates/claude/enable/typed.md, and the section-extraction passage in the memory record. Epistemic policy and adaptation findings retain their scope through the memory-declared object; neither record establishes actual deployed policy contents or compliance. The graph containment statement and durable storage description remain compatible.

Other overlaps do not establish duplicate identities:

• MEM-OBJ-documents is durable backing content, whereas RT-OBJ-graph is a process-memory representation that can also contain unsaved working copies. Neither replaces the other. The proposal and MEM-OBJ-rejected are separately consumed parts of the session record, not replacements for its valid bookkeeping container.
• EPI-OBJ-injection identifies the injection mapping only. MEM-OBJ-policy-knobs also covers distill settings and environment fallbacks. The injection mapping overlaps a narrower part of that configuration account; it is not identical to the whole. The broader record distinguishes its consumers and remains valid, so no split supersession is warranted. Neither configuration object is the delivery route RT-RTE-session-index or the authority path RT-BAP-session-policy.
• MEM-RTE-inbox-read implements pending-candidate read-back. WF-RTE-inbox-selection additionally traces foreground merging and user disposition. They overlap at grouped output but have different endpoints and responsibility boundaries. Both are parts of RT-RTE-session-capture, not duplicate declarations of the entire capture procedure.
• EPI-RTE-review-prompts covers MCP prompt/resource delivery. MEM-RTE-mcp-context also includes squash, tree and statistics channels. The narrower prompt/resource account does not replace this valid alternatives container. Shared use of DocumentReader does not make either identical to RT-RTE-document-retrieval.
• EPI-RTE-schema-check computes evidence; RT-BAP-schema-admission describes the enforcing consumer. EPI-OBJ-learning-schema is predicate configuration, not either computation or authority. These are distinct referents.
• EPI-RTE-post-tool is the diagnostic part of RT-RTE-editor-reload. The combined route remains valid alongside this part. MEM-RTE-graph-restructure and EPI-RTE-link-cleanup add structural operations rather than replace RT-RTE-mcp-document's create/update/query account. Cleanup also overlaps query deletion, but its transformation is not the query admission route as a whole.

Independent convergence

RT-CLM-capture-exclusivity, the epistemic claim comparison and the workflow account converge on the same distinction: the selected-write rule is a host procedure, while CLI, MCP and editor routes lack a required selection token. This is a comparison between source doctrine and alternate implementation paths, not a disagreement between analysts or evidence of an actual unauthorized write. EPI-CLM-shared-context is a separate continuity/access claim and is not superseded by the exclusivity claim.

The schema account similarly converges without collapsing statuses. Checked writes can reject structurally inadmissible content; code that computes a report is distinct from code that consumes it as a veto. EPI-RTE-post-tool and the memory/workflow annotations describe after-write advice, not pre-write rejection. Normalize's unchecked persistence in MEM-RTE-graph-restructure does not contradict the runtime's explicitly limited create/update/query checks.

RT-RTE-policy-maintenance, MEM-OBJ-rejected and the epistemic bounded conclusion connect a rejection-ledger-to-policy-to-later-consumer pathway. Requested human-assisted adaptation, symbolic settings consumed by code, and unobserved model compliance remain separate. The epistemic ledger's doctrine only adaptation assessment concerns the requested revision behavior; it does not deny the wired consumers of installed settings. No report's uncertainty about criticism, learning or exercised self-improvement becomes an absence finding by reconciliation.

Integration-issue dispositions

The memory integration issues identify these consequential qualifications:

1. No-groups guard. The annotation on RT-RTE-session-index supplies the SRC-1, crates/iwe/src/internal/claude/hook/index.rs, passage returning no output when injection groups are empty. Memory calls this an additional correction to runtime wording. The supplied runtime already says output is conditional, can be absent when store/index conditions fail, and does not assert that MEMORY existence guarantees delivery. Thus the handed findings are compatible; the specific guard remains in the memory annotation rather than being silently added to the runtime declaration. Policy existence alone cannot support receipt, and neither report establishes host delivery.
2. Brief sampling. Memory distinguishes the embedded prompt's ten-document description from implementation-selected injection groups, with starter default 20 and custom limits. RT-RTE-policy-maintenance and the boundary loop do not assert a fixed ten-document implementation. This is a doctrine/implementation qualification, not an anchored inter-report conflict. It prevents a universal sample-size conclusion.
3. Provenance projection. Durable documents may hold provenance, but the memory annotation on RT-RTE-document-retrieval records body rendering without arbitrary frontmatter. MCP resources offer a distinct path. This qualifies what reaches the consumer without contradicting runtime's body/metadata account. Neither stored evidence nor the broader resource capability proves that ordinary retrieval delivered provenance.
4. Feedback interpretation. Completion can sweep unasked candidates into the rejection ledger; MEM-RTE-session-adopt advances progress while dropping pending entries without reading. The workflow independently preserves both distinctions. These qualify the interpretation of capture bookkeeping, not the existence of capture or rejection-feedback affordances.
5. Failure boundary. The workflow annotation on RT-RTE-session-capture adds hub-linking before record save and a save-failure return. It complements runtime's nontransactional recovery account. Written documents, hub effects and settled records must not be treated as one atomic delivered result.

No unresolved inter-report disagreement was established in the handed versions, and no necessary missing operative part was identified that requires superseding a combined record. No analyst correction answers were supplied, so no declined objection is treated as an established defect. Remaining unknowns—host wiring, deployed content and authorship, lower-level transformation preservation, other maintenance variants, concurrency, exercised selection and measured benefit—remain scoped limitations in their declaring reports. They prevent broader end-to-end, universal-check and benefit conclusions, not the bounded connections above. This reconciliation required no additional source reads.