verbatim-citation-gate reconciliation

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

Reconciliation

The reports share the subsystem boundary and SRC-1 revision in the boundary and Source register. They connect a returning citation auditor, its shipped example consumers and its evaluation machinery, not an enclosing application's retrieval-to-answer loop. This reconciliation uses the supplied members only; no source reread was needed. No correction answers or refusal were supplied. No supersession is established, and no disagreement between reports was identified. The overlapping records below retain distinct scopes rather than acquiring the strongest status found in another account.

Audit, checks and returned objects

EPI-RTE-quote-check explicitly declares the deterministic check as part of RT-RTE-audit-citation. The former owns normalization and occurrence classification; the latter also owns callback selection and the final return. Their identity is ruled out: the check is also directly callable, while the combined route conditionally proceeds to judgment. The parent remains a valid route container. Its findings do not become wrong when the check is separately described, so no split supersession is needed.

Likewise, EPI-RTE-verdict-admission is a declared part of RT-RTE-judge-support, not a replacement for it. The parser admits a representation after the callback proposes a judgment; the parent constructs messages and invokes that callback as well. The retained parser and delivery passages in the runtime record support this distinction. Checking the source-relative relationship and admitting a label remain separate functions. Neither record supplants the other.

The output and input objects do not duplicate their producing or consuming routes. EPI-OBJ-docs names current supplied text, not the audit or judge computation. EPI-OBJ-gate-result names classification strings produced by quote or span checking, not RT-RTE-span-gate itself. The span and quote routes have different predicates despite sharing found: valid contentful spans need not satisfy full-document quote boundaries, and optional chunk checking belongs to the span route. Their identity is ruled out.

EPI-OBJ-verdict and EPI-OBJ-reasoning separate symbolic label/confidence fields from a natural-language explanation. They do not duplicate each other or RT-OBJ-judge-system, which is static input policy. The memory annotations on RT-RTE-audit-citation and RT-RTE-judge-support agree with the epistemic inventory: the direct call returns structured fields, the dispatcher exports only the label, and neither persists a judgment. There is no missing combined result record requiring supersession; the operative fields are already distinguished.

Memory and evaluation overlaps

MEM-OBJ-normalized and MEM-RTE-normalization-reuse add the retained cache object and its exact-key reuse mechanism. They do not duplicate the audit, span or epistemic quote-check routes. The cache object is not a route, and its helper route crosses quote and span consumers rather than naming either complete computation. The epistemic quote-check record also covers normalization and acknowledges memoization, but its occurrence-check scope differs from the cache route's later reuse. No identity or supersession follows from this shared implementation. The records agree on 4096 raw-string keys, process-local retention and no retained citation decision; no performance benefit is promoted from wiring to observation.

MEM-OBJ-corpus is a persisted evaluation input, not a duplicate of the generic current-request EPI-OBJ-docs. The benchmark supplies corpus values to deterministic checking, but the corpus record owns filesystem retention and acquisition lineage rather than all caller-supplied citation inputs. MEM-OBJ-results is the lossy aggregate output artifact, not a current-call verdict or reasoning object. Its failure sample and counts are distinct from the corpus's abstract texts. These object identities are ruled out.

MEM-RTE-corpus-reload and EPI-RTE-benchmark-check overlap on acquisition, corpus persistence and reload. They are not identical routes: the memory route follows durable input into later use, whereas the epistemic route additionally owns constructed expected labels, exact comparisons and measurement warrant. Both retain the distinction between code-defined generation and an authenticated generating execution. The broader evaluation account remains valid alongside the narrower persistence account; no supersession is needed. The memory analyst's seven-entry corpus sample and the epistemic analyst's uninspected corpus contents are different inspection coverage, not competing findings about corpus authenticity.

Example consumers and workflow

WF-RTE-demo-batch and EPI-RTE-example-display share the demo's audit-to-print path. Their identity is ruled out by scope: the workflow record traces the four-item demo batch, notes, counter-derived cost labels and batch completion; the epistemic record covers display in both demo and quickstart, assessing its lack of answer-admission force. Neither says the other's broader or narrower route is invalid. They remain overlapping accounts without supersession.

Both are distinct from RT-RTE-audit-citation, the nested library computation. The workflow annotations add termination, continuation and delivery roles without redefining the runtime record. The epistemic display record supplies the quickstart's constant callback passage; the workflow retains the demo's constant callback and loop passages. These independently inspected callers converge on offline stand-ins and informational printing, not observed model judgment or enforced downstream answer suppression. A cost label derived from callback counts is not provider usage measurement.

Authority, claims and integration dispositions

RT-CMP-llm-call, RT-OBJ-judge-system and RT-BAP-judge-instruction are respectively the callback boundary, delivered instruction and its consumer/channel/force account, not duplicate records. Memory and workflow agree that the original full cited document crosses that boundary, not normalized cache output, batch notes or an evidence summary. Missing provider history, truncation and model identity remain limitations; they do not contradict the inspected absence of library context trimming.

RT-CLM-closed-judge and EPI-RTE-verdict-admission converge on the missing-field mismatch: a recognized supports label can survive absent reasoning and confidence. Workflow carries the same qualification. This is an agreed discrepancy between source doctrine and implementation, not a conflict between analysts. It does not license either semantic reliability or a blanket claim that all malformed callback values fail safely.

EPI-CLM-citation-support is distinct from the parser-failure claim: it concerns citation protection and full-strength evidential support. EPI-CLM-benchmark-detection concerns deterministic-suite detection and reported normalization improvement. Their identities are ruled out. Runtime, memory and epistemic findings consistently restrict benchmark conclusions to constructed, filtered cases and stage one; stored results and reported rates establish neither independent reproduction nor a causal component effect. The epistemic normalization counterexample qualifies semantic-preservation readings without contradicting the runtime's normalized-occurrence guarantee.

RT-RTE-release-admission remains separate from audit-result disposition, corpus replacement and benchmark reporting. Human merge/release authority and conditional publication do not supply a trace-fed policy-update consumer. The memory integration issues therefore require no correction or supersession: cache reuse expands the runtime account; corpus and results add distinct artifacts; provider state and artifact provenance stay unresolved. The members establish no exercised learning or improvement from those paths, while leaving provider changes and human development as independent questions.

All declared route records retain their own implementation and operation statuses. The epistemic ledger's implemented assessments concern architectural routes, not observed execution. The reports' shared lack of executed target code, live-model traces and controlled experiments limits activation, accuracy and benefit conclusions without creating an unresolved conflict.