Evidence & Method Appendix

DBS Case File 003 — Evidence & Method Appendix

Version: v1.3.1 · Companion: 01_MAIN_TECHNICAL_CASE_v1.3.1.md · Cutoff: 31 October 2026
Purpose: public provenance, adverse-event triage, and evidence-navigation record; not a Candidate.3 machine-conformant assessment.
Source status terms: DBS-reported = originated with DBS; regulator-documented = Monetary Authority of Singapore (MAS) statement; journalist-reported = independently published reporting that may quote a company forecast; interpretation = our analytical inference. Multiple reports of one DBS-origin figure are not independent confirmations.

01 / SOURCES

A. Source register and provenance

Each locator is intended for manual verification against the original page; dates and origin groups apply to all dependent claim IDs. DBS's published 2025 annual-report subpages share one origin; their number is not independent replication. 2025 Sustainability Report version and verification rule: S08 is the DBS-hosted 2025 report (2026 publication; the SGX record also has an Updated 12 March 2026 filing). The >11,000 workforce-identification claim is confirmed in the report's Highlights, printed page 6. The separate 96% commenced-training statistic appears in web-extracted text but could not be tied to a visually verified printed page 62 and identified PDF filing version during this revision. To satisfy the second-pass T-02 gate conservatively, 96% is omitted from the release-candidate case and atomic claim C12. Its exclusion is a verification-control decision, not an assertion that DBS did not report it. Other source passages remain subject to reviewer spot-checking.

Lineage: v1.1 source S13 and claim C18 remain retired. v1.3 introduced S17–S20 and C20; v1.3.1 adds S21 and C21 while preserving prior ID numbering. Adverse observations do not become model-specific incidents solely because they fall within the reporting year; their cause, denominator and intervention exposure must be established.

02 / CLAIMS

B. Atomic claim ledger

Key: location uses section header or paragraph rather than unverified printed report page; no invented inference from a company's use of the word “resulting” to independent causality.

Known conflict log: (i) DBS Joy timeline in S01/S15/S19—present as different phases and do not combine unique customers, chats and monthly clients; (ii) possible conflation of 2024 “economic impact” with cash savings; (iii) 81% relative monthly average count vs incident rate; (iv) 2 May 2024 outage with unknown cause and 1 Jun 2024 mobile disruption without demonstrated change causation versus 19 Mar 2026 regulator-defined change cause. These are measurement/causal boundaries, not evidence of deception.

03 / CONDITION MAP

C. Public-evidence condition navigation: all 21 Candidate.3 identifiers

Not ratings. This table records which question an external reader would ask and which public claim, if any, is relevant. “Unresolved (informal)” means unresolved from the reviewed public dossier, not that the bank lacks the condition. This is not the Candidate.3 formal U rating. Formal ratings and confidence, independent interpretation and a frozen scope are intentionally absent.

No maturity score, formal 0–3/U/N/A assessments, or percentages of unsupported conditions are calculated.

04 / ADVERSE EVENTS

D. Adverse-event and contradictory-evidence triage

Candidate.3 requires six admissibility components for a formal positive gap: event, bounded population/period, source/comparator, material link, approved materiality/alternatives, and falsification. This nonconforming public illustration cannot satisfy the independently approved materiality and finance/assessment obligations. Nonetheless, observed adverse events must be displayed rather than hidden.

Protected-constraint caution: Safety, availability, customer harm and regulatory obligations cannot be waved away by economic-benefit claims. Missing detailed incident-severity and exception records prevent a full protected-constraint examination. A formal “no material disconnect” clearance is not available here, and no adverse observation is declared immaterial merely because model causality is unknown.

05 / METHOD BOUNDARIES

E. G1, G2, case boundaries and test contract

Essential next tests (not performed; to be conducted only by properly authorized parties):

1. Define source-system records and exact intervention rollout dates; publish risk model accuracy / severity calibration with legitimate comparator.
2. Examine monthly change-request exposure and incident count/severity before and after scoring, excluding and including pause months, with sensitivity by deployment criticality.
3. Link model predictions to actual approval/override/escalation decisions; measure whether workflow changed relative to comparable controls.
4. Attribute effect using staged adoption, matched risk classes or robust time-series analysis accounting for MAS restrictions and other remediation.
5. Financially value attributable avoided incidents, retained manual effort and cost of model development, monitoring and governance under finance-controlled assumptions.
6. Establish the 2 May and 1 June 2024 disruptions' root causes and service-availability impacts before considering any change-risk-model signal; clarify whether vendor-side changes such as the October 2023 Equinix upgrade fall within the reported incident series; triage the March 2026 change-caused incident separately for customer-harm severity and control failures under an appropriately extended period, not by retrofitting the 2024 claim.

Methodology-development question (not a Candidate.3 revision): For this case as scoped, Steps 1–2 cannot be used without an independently approved retrospective materiality/constraint contract; Step 4 cannot be reached on present public intervention-cost evidence. The Step 3 headline would not differentiate DBS from another similarly opaque case, but its subtypes and blockers can locate evidence gaps. Properly authorized, differently registered retrospectives need not share these limits. Separate owner questions remain open regarding G1 coverage observations and post-period sustainment events. These questions are not publication findings or changes to the normative specification.

06 / GLOSSARY

F. Glossary

• AI/ML: Artificial intelligence / machine learning, encompassing statistical and other model approaches; not synonymous with generative AI.
• ADA / ALAN: DBS-named enterprise data and model knowledge/development platforms.
• MtJ: Managing through Journeys, a cross-functional operating approach.
• OMT: Operating Model Transformation, DBS's term for broader workflow and role changes enabled by AI.
• MAS: Monetary Authority of Singapore, DBS's banking regulator.
• L1–L5: Task, workflow, capacity, business and economic links; L6 is a cross-cutting attribution overlay, not a sixth operating endpoint.
• G1 / G2: Operational interpretability / material downstream coverage evidence gates.
• Cost avoidance: Credible counterfactual costs prevented, not necessarily observed expense cuts.
• Net economic contribution: Attributable benefit less relevant retained, implementation, recurring, monitoring and governance costs.
• Unresolved (informal): Public evidence does not settle a condition question here; this is not the formal Candidate.3 U rating and not proof the condition is absent.

07 / REVIEW CONTROLS

G. Independence, review limitations and release controls

1. Research independence: Uses public information only. No professional-access, customer-specific, proprietary, confidential or nonpublic data are introduced or implied; general methods do not rely on nonpublic company hypotheses.
2. Author interest and review characterization: The author designed both The Outcome Disconnect and The Outcome-Back Operating Model. The case has been challenged by separately commissioned AI reviewer sessions; they are not independent human assurance, do not validate these methods, and do not substitute for owner approval.
3. Review sequence: First-pass and sealed second-pass findings dated 2 Nov 2026 are preserved unchanged in the controller archive. v1.3 implemented a targeted response; the same AI reviewer subsequently confirmed T-01 and T-02 resolved and recommended readiness for owner release decision, while disclosing the self-review limitation. v1.3.1 implements the resulting N-01–N-06 and residual editorial corrections, not independent human closure. The second-pass reviewer explicitly reported not being the technical first-pass reviewer; the first-pass technical review came from a different AI reviewer session. Corrected passages require focused re-review.
4. Source/access: DBS's Sustainability Report 2025 Highlights p.6 supports the >11,000 identification statement; PDF p.62/version visual verification of the separate 96% claim was not achieved and the percentage is omitted. A web PDF screenshot request returned a size error; a direct PDF download also failed. MAS 2024/2026 statements were checked; the 2 May 2024 MAS follow-up is sourced to contemporaneous journalism, not a located official MAS publication. The second-pass reviewer accessed the EDB page; the earlier blanket “blocked” note is obsolete. CEO DBS Joy, PMA, capital-allocation and CodeBuddy wording were subsequently checked directly against DBS 2025 CEO reflections S01; no independent measurement verification follows from that access.
5. Causal status: No claim of independent ROI assurance, complete AI-cost accounting, causal isolation of 81% change, or proof that 2026 outage was scored by the model.
6. Release: Technical v1.3.1 not deployable until the owner records a release decision. v1.3 underwent focused AI re-review; the v1.3.1 editorial changes have source and structural checks but no separate independent re-review. The accessible DBS editorial companion is a separate artifact with its own release record; its earlier AI review recommended corrections and human countersignature. Cross-publication consistency is an editorial checkpoint, not implied approval here.

End of Appendix v1.3.1.