← Glossary

Devancore Post-Trade Glossary

Conversational Books and Records

Conversational books and records are permissioned natural-language workflows for querying regulated records while preserving source citations, retention, supervision, and audit evidence.

Definition

Conversational books and records are controlled natural-language workflows for asking questions over regulated operating records. The user can ask about orders, trades, blotters, positions, cash, confirmations, allocations, settlement state, exceptions, reports, approvals, and audit trail. The answer should be grounded in governed records, not free-form memory.

The control problem is access with evidence. A conversational layer can make regulated records easier to retrieve, but it should not weaken entitlement, retention, supervision, source lineage, or books-and-records discipline. The question, retrieved sources, response, citations, action, and final state may all matter later.

Conversational record layer

Conversational record layer

The interface is useful only when every answer can point back to governed records.

Layer Record or control Operating question
User request Prompt, user, role, account scope, portfolio scope, timestamp, and session context Who asked and what were they allowed to ask?
Entitlement Role, desk, entity, account, portfolio, market, client, and record-level permissions Which records can the workflow retrieve?
Retrieval Trade, order, blotter, ledger, position, cash, settlement, confirmation, allocation, and exception records Which source records support the answer?
Answer Cited response, source IDs, as-of time, confidence boundary, missing records, unresolved ambiguity, and citation validation Can the user verify that the answer and cited records are real?
Action Draft instruction, report, escalation, exception note, correction request, or approval workflow Did the conversation create an operational next step?
Retention Prompt, retrieved records, response, citations, user decision, approval, downstream event, and final state Can the interaction be reconstructed later?

The system of record remains the authority. A conversational interface can retrieve, summarize, compare, explain, or draft workflow actions against records, but the underlying order book, trade blotter, ledger, position record, cash record, confirmation file, settlement status, or exception record remains the source.

Permissions are part of the answer. A user should not receive a summary of records they are not entitled to view. Scope can depend on role, desk, entity, account, portfolio, client, market, instrument, workflow state, or supervisory responsibility. If a response excludes records because of permission boundaries, the workflow should preserve that scope decision.

Citations convert an answer into usable evidence. A statement such as "three trades failed" is not enough for institutional operations. The answer should point to the trade IDs, account, security, amount, settlement date, fail reason, source system, and as-of time that support the response. Citation hallucination is a separate control risk: before an answer is shown, deterministic validation should confirm that referenced IDs, rows, documents, and calculations actually exist in the governed record set.

The conversation itself can create a record. A query may lead to an exported report, draft exception note, settlement escalation, correction request, approval, rejection, or supervisory review. If a natural-language prompt instructs, modifies, approves, escalates, or queries transaction-related activity, the prompt and response may need to be retained as business-communication evidence under policies shaped by SEC Rule 17a-4 and FINRA records obligations. The prompt, retrieved context, response, user decision, and downstream outcome should be retained according to the firm's policy and workflow requirements.

How it works

Conversational books and records work by placing a permissioned query layer over governed records. A user asks a question. The workflow scopes access, retrieves source records, generates a cited answer, and records any action that follows.

Query to evidence

Query to evidence

Natural language should retrieve records through controls, not bypass them.

Step Workflow action Required evidence
Ask User asks a question about trades, accounts, positions, cash, exceptions, reports, or controls Prompt, user, role, timestamp, and context
Scope System applies entitlement, record-level permissions, portfolio scope, and data-domain limits Permission result and excluded scope where relevant
Retrieve Workflow queries governed records and supporting files with as-of time and source lineage Record IDs, source system, version, and timestamp
Answer Model produces a concise response with citations, gaps, and uncertainty where needed Cited rows, documents, calculations, unresolved items, and lineage checks
Act User drafts, approves, rejects, escalates, exports, or opens a workflow item Action type, actor, approval state, and affected records
Retain Prompt, retrieved context, answer, citations, action, and final state are stored under policy Replayable query and decision trail

The workflow starts with a question. The question may be simple, such as "show open fails for this account," or more complex, such as "explain why cash changed after yesterday's close." The original wording should be retained because it explains the user's intent.

Entitlement runs before retrieval. The system checks whether the user may access the relevant accounts, portfolios, entities, clients, markets, instruments, workflows, and record types. Natural language should not become a shortcut around role-based access or supervisory separation.

Retrieval gathers the record set. The workflow may query orders, executions, allocations, confirmations, settlement instructions, custodian files, bank statements, ledger entries, position records, cash balances, exception cases, approvals, reports, and supporting documents. Each source should keep version, timestamp, and lineage.

The answer should cite its basis. If the response explains a break, it should point to the internal row and external record. If it summarizes positions, it should state the exact as-of timestamp for the position, cash, settlement, or ledger view. If it calculates an amount, it should identify the inputs. If the records are incomplete, stale, or conflicting, the answer should say so.

Action remains controlled. A user may ask the system to draft a message, open an exception, prepare a report, route an approval, export a file, or generate a correction request. Those actions should create workflow events with user, role, timestamp, source records, approval state, and downstream response.

Retention closes the loop. The firm should be able to replay what was asked, what records were retrieved, what answer was shown, what citations supported it, what action followed, and what final state resulted. For regulated workflows, the chat log is not just interface exhaust; it can be part of the supervised evidence trail.

Conversational records — query controls

Devancore · message matrix

Rail Message Purpose Record
Prompt user question request user, role, session, account scope, portfolio scope, timestamp, and requested record domain
Permission entitlement check scope allowed records, blocked records, data domain, desk, account, entity, and supervisory view
Retrieval source grounding evidence order, trade, ledger, cash, position, confirmation, settlement, exception, report, and file citations
Response cited answer explain answer, source IDs, as-of time, missing data, unresolved ambiguity, and calculation basis
Action workflow event control draft, approval, rejection, escalation, export, downstream request, and retained query trail

In Devancore™

Devancore supports conversational books and records as a controlled access and evidence layer over institutional operating records. The platform should be framed as infrastructure for permissioned questions, cited retrieval, workflow actions, approval state, and audit trail.

Devancore should not be framed as legal counsel, compliance owner, accounting authority, broker, custodian, clearing broker, execution venue, or recordkeeping guarantee. Its role is to help teams connect questions, records, evidence, and workflow outcomes.

In a Devancore-style workflow, a user question becomes an operating event. The event is tied to user, role, permission scope, record domain, retrieved sources, cited answer, and any follow-on action. If the user opens an exception, drafts a correction, requests a report, or routes an approval, that action remains linked to the source records used in the conversation.

This page is the books-and-records counterpart to the broader conversational finance cluster. Conversational finance defines the interface. Conversational finance ops applies it to operating teams. AI agent post-trade operations prepare work. Conversational books and records define the retention and evidence discipline that makes those workflows usable in regulated environments.

The practical test is whether a supervisor, auditor, examiner, operator, or engineer can reconstruct the answer later. What did the user ask? What records were they allowed to see? Which records were retrieved? What answer was shown? What action followed? What final state was created?