Digitising financial records for controlled AI retrieval

Learn in your browser · Chapter 01 of 08

A useful financial-record index begins with a bounded question, an information owner and an inventory that preserves what each record means. Practise this with the invented Northgate Records Lab before proposing any real archive processing.

Northgate Records Lab is an entirely fictional adult learning environment. Its statements, accounts, policy and analyst note are fabricated. Work locally with the supplied packet: no customer upload, external model, network connection, transaction or credit decision is part of this exercise. The practical problem is precise: an analyst needs the approved June statement for one permitted account, while the archive also contains another account, an older statement, a bad transcription and an unapproved note.

Define the proposed contribution as exact-reference lookup of an approved current statement followed by source checking. Do not begin with 'make all finance searchable'. That phrase leaves the purpose, audience and permission boundary unresolved. An inventory can support a records engineer's preparation without granting that engineer statement-query access. A retrieval result can support a human's inspection without authorising financial interpretation. Keep preparation, retrieval and interpretation as separate decisions, even if a future interface combines them.

Create a context card before copying files. Record organisation and entity type, permitted purpose, subject and product, selected jurisdiction or unresolved status, applicable regulator or policy owner, processing and data locations, information owner, and retention or legal-hold owner. In this packet, organisation is the fictional Northgate Records Lab; subject is synthetic account statements; purpose is the bounded learning lookup. The packet does not establish real regulatory applicability or actual owners for a bank. Leave unsupported fields unresolved rather than filling them with a familiar UK default.

Inventory records and derivatives separately. The printed statement is the reference for its supplied figures; OCR-ST-101 is a fabricated transcription derived for teaching and contains an error. A filename containing ST-101 is insufficient to select the June source. Capture identifier, revision, status, account, currency, period and source locator. Add who supplied the record and who can resolve an uncertain status. For this finite exercise, a page locator can be a labelled teaching reference rather than an invented scanned-page count.

Scroll sideways to see every column.

The financial failure case is an archive importer that merges every ST-101-looking file into one row. It keeps the newest file modification time and reports 121500.00 as the June closing balance. Modification time establishes neither approval nor the period described. The importer also loses the distinction between a statement and its transcription. Repair the inventory before retrieval: keep separate record identities, attach status and context, and mark the derivative discrepancy for review. A polished search result cannot recover metadata discarded at ingestion.

1. Write one question: locate the approved June 2026 statement for AC-101 for permitted learning inspection.
2. List each source revision and derivative independently, preserving account, currency, period and supplied status.
3. Record unresolved ownership or context fields as unresolved and identify the decision needed before real use.
4. Separate eligible current references from historical and unapproved material without destroying the inventory history.
5. Keep a human-readable exclusion reason so a reviewer can distinguish wrong period, wrong account and unresolved transcription.

Work the supplied inventory question. ST-101 revision 2 is the candidate source because it is approved current for AC-101, USD, June 2026. Its supplied opening is 12500.00, debits are 250.00 and 150.00, credit is 50.00, and closing is 12150.00. ST-101 revision 1 cannot answer June: its period is May. The intentionally independent earlier record does not explain the June opening. ST-202 cannot substitute for AC-101 even though its arithmetic reconciles to 5900.00. Eligibility and arithmetic answer different questions.

Text equivalent: the bounded question defines the inventory fields. Separate rows preserve revision and context. A matching approved current source becomes a candidate, while other rows retain an explicit exclusion reason. Candidate selection still leaves the caller's permission to be checked.

Scroll sideways to see every column.

Learn in your browser · Chapter 02 of 08

Digitisation preserves meaning only when amounts remain attached to their account, currency, period and source. Check the supplied Northgate transcription with exact arithmetic and inspect discrepancies before permitting any answer to use it.

OCR can turn a page image into proposed text, but this course has not run an OCR engine. OCR-ST-101 is a supplied fabricated transcription designed to expose a closing-balance error. The exercise oracle is finite: the original printed figures supplied for ST-101 revision 2 and their arithmetic. This permits a reproducible comparison of known fields. It cannot measure the accuracy of a real scanning workflow, its layout recognition or its performance on unseen documents.

Separate capture quality from financial correctness. A cropped page can omit a transaction; a transcription can move a decimal point; a parser can interpret a numeric string incorrectly; a summariser can detach the amount from USD or June 2026. Each stage needs its own evidence. Preserve the original reference, label the transcription and record corrections as reviewed derivatives. Replacing the printed amount silently with a computed value obscures whether the source, extraction or calculation was wrong.

Represent amounts with exact decimal rules. For this USD teaching statement, 12500.00 corresponds to 1250000 cents. Subtract 25000 cents and 15000 cents, then add 5000 cents. The result is 1215000 cents, or USD 12150.00. A decimal implementation can perform the same calculation directly. Avoid binary floating-point rounding as an unexplained reconciliation tolerance. Explicit scale and sign rules also stop 12150.000 from being accepted without examination when the finite fixture expects a two-place representation.

Scroll sideways to see every column.

Work the reconciliation in the stated order: 12500.00 minus 250.00 equals 12250.00; minus 150.00 equals 12100.00; plus 50.00 equals 12150.00. Compare the computed result with the printed closing: difference 0.00. Compare the fabricated transcription closing with the printed closing: 121500.00 minus 12150.00 equals 109350.00. This identifies a closing-field discrepancy; it does not establish that every other field was transcribed correctly. A reconciliation can balance despite two offsetting errors, so inspect each supplied field as well.

A financial failure case accepts any positive closing value and reports that AC-101 has USD 121500.00. The value is plausible as a number but contradicts both the supplied print reference and row arithmetic. Another weak check uses a substring search: 12150 appears inside 121500, so it declares a match. Compare complete typed values, with currency and scale, rather than text fragments. For adversarial parsing practice, 12150.000, the string '12150.00' and the boolean true require explicit validation rules. A boolean is not an amount; a string is not automatically a validated decimal.

1. Locate the approved ST-101 revision 2 reference and record account, currency, period and source locator beside the extracted fields.
2. Compare every supplied amount and transaction direction against the printed reference before calculating.
3. Apply exact decimal or integer-cent arithmetic using the statement's supplied debit and credit convention.
4. Record both the reference reconciliation and the transcription discrepancy; do not edit the preserved source.
5. Hold the erroneous derivative from answering and prepare the exact corrected derivative for accountable human review.

The comparison denominator matters. If you inspect only the supplied closing field, you can say one inspected closing field disagrees. You cannot report an archive-wide OCR error rate. If you compare opening, two debit amounts, credit and closing, that is five specified amount fields in this fixture; report the exact inspected set and differences, including any representation failures. Neither description estimates unseen page errors. Real evaluation would need an authorised, representative sample, independent reference labels and separate measures for omitted rows, context errors and numeric errors.

Text equivalent: compare the printed reference and fabricated transcription field by field. Calculate with the checked amounts and directions. Retain the comparison and arithmetic results together so a reviewer can hold the erroneous derivative or review a labelled correction.

Scroll sideways to see every column.

Learn in your browser · Chapter 03 of 08

A financial citation needs more than an identifier. Preserve the source revision, page or field locator, context and transformation history so a reviewer can determine exactly what supports an answer and whether it is still eligible.

Provenance connects a received record to its page image, transcription, fields, indexed references and proposed answer. Trace either direction: from an answer to its supporting source, or from a withdrawn source to derived material requiring removal from active retrieval. This relationship does not establish the truth of arbitrary source prose.

Keep document identity distinct from revision and status. ST-101 revision 2 is approved current for AC-101, USD, June 2026. Revision 1 is superseded and describes May 2026 with closing 12400.00. A citation saying only 'ST-101' could point to either. Revision number alone also does not establish approval. POL-08 revision 3 is approved current synthetic guidance; its revision 2 is superseded. Store the supplied status and its accountable decision relationship rather than inferring current authority from the largest revision number in a folder.

Use page-level provenance in a real scanning workflow, with field locators where needed. A citation would contain record ID, revision, actual page, field or row, account, currency and period, plus the preserved source reference. The Northgate packet supplies figures rather than physical scanned pages, so do not invent a page number. Here use the labelled locator 'supplied statement figures, closing field'. That is a transparent exercise locator and a clear capability gap: actual image-to-page verification remains necessary before any real archive release.

Scroll sideways to see every column.

A financial failure case cites ST-101 revision 1 for the sentence 'The June closing balance is USD 12400.00'. The number genuinely appears in the supplied archive, but it belongs to May and a superseded record. Correct the proposition rather than merely adding another citation. A defensible sentence is: 'The supplied approved current ST-101 revision 2 gives AC-101's June 2026 closing as USD 12150.00.' Do not calculate a supposed monthly change of 250.00 from the May closing. The packet intentionally treats the earlier record independently and does not establish that continuity.

Chunks must inherit source context. Without account and period metadata, 'closing balance' can mix AC-101 and AC-202. Keep row relationships together or link sufficient context to reconstruct them. Preserve signs, labels and currency. Similarity scores describe matching behaviour; they cannot upgrade a note, superseded policy or wrong-account statement into authority.

1. Create a source-version record with identifier, revision, supplied authority, account, currency, period and preserved reference.
2. Record each transformation's input and output references and the operation performed; distinguish supplied teaching transcription from executed extraction.
3. Attach real page and field locators when available; explicitly record missing physical-page evidence in this fixture.
4. Propagate source-version context and eligibility to every derived chunk and active-reference entry.
5. Bind review to the exact source and output versions, permitted purpose and accountable decision; invalidate that binding when a material version changes.

Content digests help compare bytes. If a preserved reference and a later copy have matching digests under the same algorithm, the comparison supports byte consistency against that reference. It does not authenticate the reviewer, prove lawful possession or approve query access. A self-written approval JSON is still self-written. If corrected derivative bytes change, recompute their comparison value and obtain review of the changed output. Do not relabel the earlier review as covering a different amount because the identifier remained ST-101.

Text equivalent: an exact source version supplies a page or field locator. A labelled derivative retains that provenance. A proposed answer cites the checked evidence and receives exact-version review. Byte comparison supports derivative consistency but does not supply approval or truth.

Scroll sideways to see every column.

Learn in your browser · Chapter 04 of 08

A relevant financial record can still be outside a caller's permission. Apply Northgate's expressly synthetic role, purpose, account, status and epoch gates before constructing a result, and keep untrusted document instructions outside the control plane.

These permission rules are an expressly synthetic local teaching policy. POL-08 revision 3 allows approved-statement lookup after role and account checks, with human review of financial interpretation. It grants no real financial access. The finite boundary distinguishes AC-101-scoped analysts, approved-current-policy readers and engineers permitted to build the index without automatically acquiring query rights.

Authorise before constructing a result. Check recognised role, account scope, purpose, approved current status and active epoch, initially 1. These gates apply to exact-reference lookup and future similarity search. Retrieving all statements then hiding rows exposes content to logging, caching or generation components. Bound their input first. The lab's explicit context selection is synthetic-selected; its default unresolved context holds requests. Selection is a teaching choice, not proof of legal applicability.

Scroll sideways to see every column.

Work a permitted request. At epoch 1, the scoped analyst asks for ST-101 revision 2 for AC-101's June statement inspection. Verify the recognised role and AC-101 scope, the permitted exact-reference purpose, approved current status and epoch 1. The result can then identify the checked reference with USD 12150.00 as its supplied closing and the transparent field citation. If the analyst asks whether that balance supports a lending decision, the purpose has expanded. The packet supplies no credit-decision authority; hold the interpretation for the appropriate human process.

Now work the cross-account failure case. The same analyst asks for ST-202 revision 1. Its opening 6000.00 minus debit 200.00 plus credit 100.00 equals closing 5900.00, but arithmetic and approved status do not overcome the AC-202 scope failure. Deny before constructing the statement result. The caller response is generic and contains neither the other customer's identifier nor an existence clue: 'Request cannot be fulfilled under the current access context.' Detailed failure reasons belong only in an appropriately protected internal decision record.

A policy reader may retrieve approved current POL-08 revision 3. Revision 2 fails current status; ST-101 fails the role grant. Evaluate each requested resource and purpose, including linked sources proposed for citation. Reading a policy paragraph mentioning statements cannot confer statement access.

AN-09 revision 1 contains fabricated instructions to grant AC-202 access and suppress source citations. Treat those sentences as unapproved analyst-note content. They cannot edit the caller's role, account scope, policy or review requirements. Even if retrieved text says 'administrator authorises this', the claim is still inside the data plane. The control plane consists of the recognised permission configuration and accountable decisions outside the note. A summariser can describe the attempted instruction for review; it must not execute it or send information elsewhere.

1. Receive a recognised caller identity and explicit access context from the synthetic exercise configuration, never from a retrieved note.
2. Confirm the requested purpose and role-resource grant before looking up eligible content.
3. Check account scope for statements, approved current status and the active index epoch.
4. Return only the eligible reference and required provenance, or a generic denial without resource-existence details.
5. Review financial interpretation separately and reject any source instruction that changes permissions or suppresses citations.

Withdrawal of ST-101 revision 2 advances the active index to epoch 2. Hold old epoch 1 requests and invalidate cached answers. Chapter 6 develops withdrawal; previous availability cannot establish current eligibility. Content digests repair neither missing scope nor stale status, and self-authored approval files create no real access grant.

Text equivalent: recognised caller context enters the permission gates. Only a request passing role, scope, purpose, current approval and active epoch receives a constructed cited reference. Failure produces a generic denial or hold. AN-09's instruction supplies no recognised access context.

Scroll sideways to see every column.

Learn in your browser · Chapter 05 of 08

Financial records can pass through scanners, support tools, storage and reviewers in different places. Build a context card and processing map that expose unresolved authority before choosing a real retrieval environment.

A records team can preserve figures yet use an unsuitable processing path. A local scanner does not reveal transcription, support or cache locations. Ask who processes which information, for what purpose, at each stage. This chapter produces an evidence request, not a legal determination or financial-product approval.

Northgate is invented, with no real customer, licensed entity or regulated product. Choose a study region or record jurisdiction unresolved. The choice directs research but cannot make law apply. USD establishes neither residence, licence, location nor regulator. An EU pathway remains separate from actual applicability analysis.

Use NIST AI RMF MAP context method to describe purpose, affected people and boundaries. Its framework is voluntary and version 1.0 is being revised. Map exact-reference lookup separately from financial interpretation, transfers or credit decisions. Name the information owner and processing decision owner. A risk description supports their decision; it cannot replace their authority.

Scroll sideways to see every column.

For a real proposal, draw the path before reading supplier assurances. Name the source archive, scanner, transcription process, working folder, index, cache, reviewer interface, support channel and backup/export locations. For every arrow, ask whether content, identifiers or operational metadata leave one controlled environment. A regional hosting label may omit remote support, subcontractors or diagnostic logs. Obtain evidence for those paths rather than filling gaps with the main server's advertised location.

ICO AI guidance is UK guidance under review following the Data (Use and Access) Act. Use its lifecycle consideration of accuracy and minimisation as an educational method, subject to actual applicability. Justify copied fields against the lookup purpose. Account identifiers support this exercise boundary; an imagined address adds no teaching value or jurisdiction evidence.

Keep three legal research questions distinct. GDPR territorial and material scope questions concern whether the relevant processing is within that framework. International transfer questions concern the actual movement or access involved and the relevant mechanism. DORA questions concern whether the actual entity and ICT arrangements fall within its scope and what operational-resilience obligations apply. These sources do not create a universal banking checklist. Bank secrecy, retention, localisation, contractual duties and other rules may require separate research in the selected jurisdiction.

1. Write the entity, product, purpose and data facts that are known. Mark unknown facts explicitly instead of assuming a UK institution.
2. List processing and access locations, including support and derived stores. Separate observed facts from supplier statements awaiting verification.
3. Create an applicability question for each relevant source. Ask an authorised legal, compliance or policy owner to resolve it using actual entity and product facts.
4. Record the permitted learning operation independently. Keep the Northgate lab offline while real-world evidence is missing.
5. Bind any eventual real approval to the named path, purpose and versions. Reopen review when any of those changes.

Worked example: scan ST-101 revision 2 locally, send it to an unspecified transcription service and return text to the index. USD and June 2026 identify statement context, not service locations or support access. Record those paths unresolved and hold external transcription. Continue supplied OCR comparison offline. Reconciliation is 12500.00 minus 250.00 minus 150.00 plus 50.00 equals 12150.00 USD; arithmetic does not resolve processing authority.

Text equivalent: start with a learning-region choice and a separate set of entity and processing facts. Both inform questions, but only actual evidence and an authorised review can support a real path decision. An unresolved arrow stays held. No country selection automatically authorises storage, access or interpretation.

Scroll sideways to see every column.

Learn in your browser · Chapter 06 of 08

Use an epoch transition to withdraw the active statement reference, linked chunks and cached answers together. Reopen retrieval only from intact source evidence.

After active withdrawal of ST-101 revision 2, a cached answer still returns 12150.00 USD for June. The figure matches the original packet but violates eligibility. Withdrawal controls permitted use, not merely incorrect numbers. Removing a visible search reference alone leaves the answer path inconsistent.

Northgate begins at active-index epoch 1. An epoch versions retrieval state, not a source or approval credential. Withdrawal creates epoch 2: remove the exact reference, derived chunks and cached answer. Epoch 1 queries hold. This finite transition does not prove legal erasure, backup/export deletion or production propagation.

Keep four records concepts separate. The source archive preserves material under an authorised records decision. The active index contains references eligible for this retrieval purpose. Derived chunks are copied teaching content linked to those references. Cached answers are earlier results with their own purpose, scope and epoch context. A withdrawal can change active eligibility without authorising destruction of the source archive. Real retention schedules, disputes or legal holds need their actual owner before any real deletion.

Use the NIST AI RMF MANAGE risk-response method: name the withdrawal owner, affected stores and reopening evidence. The framework is voluntary, with version 1.0 under revision. Controlled records disposition remains separate from active working-copy removal. Northgate's epoch and eligibility rules come from the fictional packet, not a regulation.

Scroll sideways to see every column.

A minimal tombstone records the withdrawn identifier, revision, decision and epoch without statement text or balances. An authorised internal owner can inspect the transition evidence. Denied callers receive generic responses without customer or document-existence details. Accountability records and caller responses have different access contexts.

1. Identify the exact source revision and active purpose to withdraw. Do not interpret withdrawal as permission to destroy the whole archive.
2. Prepare the new eligible reference set and remove its linked active chunks and cached answers.
3. Verify the new stores agree, then make epoch 2 the active state. In a real implementation, avoid a window where the new reference list shares old answers.
4. Reject stale epoch requests before result construction. Inspect an authorised query against epoch 2 and a deliberately stale query against epoch 1.
5. Record the decision, affected stores and residual scope. Reopen only from eligible source material and a fresh accountable decision.

Worked example: before withdrawal, an AC-101-scoped analyst with permitted purpose and epoch 1 can obtain the approved current ST-101 revision 2 reference. After withdrawal, a query at epoch 2 cannot produce the June balance from that reference because it is absent from the eligible set. ST-101 revision 1 remains superseded and relates to May 2026, so it is not a fallback for June. ST-202 concerns AC-202 and remains outside the analyst's scope. Recovery therefore means holding the June answer while a valid source and decision are restored, not choosing the nearest number.

Recovery requires an intact eligible source and provenance for rebuilding. Surviving OCR-ST-101 cannot recreate an approved source: arithmetic detects its wrong figure but cannot recover lost authority. If the preserved source exists, verify version/bytes, reconcile derivatives, check status and obtain review. Digests detect byte differences; they confer neither approval nor truth.

If evidence is missing, seek a replacement from the authorised source owner and hold June lookup. Do not reconstruct transactions from cached text, extrapolate from May or broaden access. State what was restored and what remains held; a delay is safer than hidden substitution.

Text equivalent: withdraw the reference, linked chunks and cached result together, then verify epoch 2. Old-state queries hold. Archive disposition remains separate. Recovery begins with an eligible source and review, never a superseded amount.

Scroll sideways to see every column.

Learn in your browser · Chapter 07 of 08

A proposed answer must preserve the account, currency, period, revision and page that support it. Evaluate concrete failures against the supplied packet before treating fluency or a correct total as adequate evidence.

A proposed answer says the current balance is 12150.00 and the analyst may proceed. It omits currency, account, period, revision and the meaning of proceed. A reader could mistake a June figure for a live balance or action authority. Verify the whole claim, not only its number.

POL-08 revision 3 requires role/account checks for approved statement lookup and human review of interpretation. Build a claim ledger with each claim, exact source, context and added inference. For this finite fixture, cite supplied statement figures, closing field; physical page not supplied. Structural eligibility can pass while actual scan-page inspection remains a release gap.

Use NIST AI RMF MEASURE principles to choose methods matched to the task and risks. The framework is voluntary and under revision. Set expected results from printed figures, statuses, roles and epochs before execution. Inspect permitted and denied paths. This finite packet cannot estimate real OCR, model, language or product performance.

ICO AI guidance is UK guidance under review. Its accuracy method encourages inspecting sources and distinguishing fact from opinion where applicable. Keep printed figures separate from fabricated OCR and commentary. AN-09's instructions to grant AC-202 access and suppress citations remain untrusted content; inspect them without obeying them.

Scroll sideways to see every column.

Worked verification: attach AC-101, USD and June 2026 to opening 12500.00, debit 250.00, debit 150.00, credit 50.00 and closing 12150.00. Compute 12500.00 minus 250.00 minus 150.00 plus 50.00 equals 12150.00. Compare printed closing with fabricated OCR closing 121500.00. The difference is 109350.00 USD. That discrepancy establishes this transcription is wrong for the supplied field. It does not justify silently rewriting source bytes or claiming the extraction pipeline has achieved an accuracy rate.

The lab rejects non-string and non-two-decimal money rather than silently rounding. Python Decimal documentation supports constructing decimal values from strings; float construction can retain binary float artefacts. A malformed 12150.000 must hold even if numerically equal to 12150.00. Real source conventions require their own validation.

1. Break the answer into factual claims, context and proposed actions. Identify any inference that goes beyond lookup.
2. Check source status, version, account permission, purpose and epoch before reading or constructing an answer.
3. Match amounts to printed figures and reconciliation. Use the supplied closing-field locator; record physical page not supplied.
4. Run the adversarial cases with expected dispositions fixed in advance. Record what happened, including whether a denial exposed another account.
5. Submit the exact proposed answer and source/output versions for accountable human review. Hold any unresolved amount, citation, interpretation or authority claim.

Use failures diagnostically. If a stale epoch serves a correct answer, fix state enforcement and cache invalidation. If AC-202 appears in denial text, fix the disclosure path. If 121500.00 reaches the answer, fix transcription verification. If AN-09 changes the role, fix the separation between source content and control authority. Repeating the same successful lookup does not address any of those failures. A test suite should explain which boundary failed and what evidence would demonstrate a repair.

Report named executed cases, expected results and fixture/code versions. Do not translate synthetic passes into OCR accuracy, model quality or universal security assurance. A denied query may be the correct task result; fluency may be wrong. Count results against predefined expected behaviour.

Text equivalent: list the answer's claims, then check eligibility before using evidence. Inspect printed amounts and citation context independently. Present both with the exact wording for human review. Any unresolved factual or action claim holds. A digest or successful arithmetic check cannot substitute for that decision.

Scroll sideways to see every column.

Learn in your browser · Chapter 08 of 08

Assemble a reproducible exact-reference teaching prototype and an evidence pack that separates executed behaviour from untested deployment needs. The knowledge assessment and written capstone answer different learning questions.

The capstone solves a bounded problem: preserve supplied figures, enforce purpose/account permissions, cite eligible versions and withdraw active derivatives offline. Its evidence pack makes results inspectable. It establishes no deployed service, real OCR accuracy, vector retrieval quality or authority to process real financial records.

Preserve every Northgate revision and status, including fabricated OCR and unapproved AN-09. Printed figures and declared roles form the finite oracle. The prototype selects known identifiers deterministically; it runs no embedding, external model, transaction, credit decision or customer communication.

Record actual files, code/fixture versions, instructions and observed outputs in a manifest. Digests compare recorded bytes. Purpose, owner, status and accountable review remain separate evidence. A matching AN-09 digest preserves unapproved instructions; it cannot grant their requested access.

Structure evidence using the voluntary NIST AI RMF: MAP context, MEASURE evaluation, MANAGE responses and GOVERN responsibilities. Version 1.0 is being revised. Describe boundaries, predefined/observed results, holds, recovery and decision owners. These headings organise evidence rather than certify it. Keep preserved sources, derivatives, statuses and disposition decisions explicitly linked.

Scroll sideways to see every column.

1. Read the README and keep execution offline. synthetic-selected is a fictional learning context, not legal approval.
2. Run audit, query and scenario against unchanged fixtures. Capture actual commands, versions, observations and failures.
3. Inspect permitted AC-101 lookup and denied engineer, policy-reader statement and cross-account requests. Denials omit existence details.
4. Reconcile printed June figures; reject fabricated OCR and malformed money. Inspect source and field-locator context.
5. Inspect withdrawal across reference, chunks and cache, then test old/new epochs. Explain June holds.
6. Assemble the evidence pack and gap record; real adaptation needs its relevant review.

Worked capstone disposition: the scoped analyst's permitted epoch 1 lookup can cite approved current ST-101 revision 2 for AC-101 and June 2026. The records-engineer can prepare an index but cannot rely on that role for statement queries. The policy-reader can obtain approved current POL-08 revision 3, not balances. After withdrawal, epoch 1 holds and epoch 2 has no eligible June ST-101 revision 2 reference. The demonstration remains useful even though it withholds a requested answer, because withholding is the expected result for that state.

Report each result separately: synthetic reconciliation, wrong-field rejection, specific permission dispositions and active-store withdrawal. The closing-field locator supports finite teaching lookup; physical page and scan provenance remain untested. No result proves backup/export erasure. Name residual paths rather than hiding them in a general success claim.

Preserve failing evidence and repair its cause. Permission failures need permission-path repairs; figure failures need verification repairs. Rerun affected and dependent checks. Do not change the oracle, remove adversarial cases or widen access to obtain a pass.

The forty-question bank draws twenty mapped MCQs per sitting and assesses knowledge only. Strictly above eighty percent means sixteen correct fails and seventeen passes. Written practice and capstone remain ungraded. A knowledge result establishes no professional competence, identity assurance, accreditation or deployment approval. Keep the private bank separate from practice answers.

Date the real-deployment gap: financial-operations teams need auditable access to actual archives; current evidence is a private teaching prototype. A possible offering adds institution-specific ingestion, authority/permission mapping, retention integration and recovery. Dependencies include authorised data, qualified records/compliance review, representative OCR/retrieval evaluation, hosting/support evidence and security/accessibility checks. Prioritise discovery and authority mapping before integration because wrong balances and cross-customer disclosure have material consequences.

Text equivalent: the preserved packet supplies expected facts; the offline prototype supplies observations. Inspect their agreement and failures before writing a conclusion. The conclusion describes this teaching behaviour. A separate gap record states what a real operational service still needs. Neither path creates approval through a digest or assessment score.

Scroll sideways to see every column.