Positions and modeling assumptions
Principal, book value, dates, coupon, currency, indexation, payment frequencies, market yield, liquidity buffer, and modeled outflows.
6 Deposits6 Mortgages5 Current Accounts
Basel Jabour / Automation & Engineering
A risk number is only useful if you can trace it.
A reproducible daily workflow that turns portfolio files into verified results, visible exceptions, and a reviewable leadership handoff.
Prepare→ETL→Report
August 29 / Recorded exception
91.8%Liquidity coverage against a 100% requirement
5 material warningsCRITICAL
Input / config / codePINNED
Human approvalPENDING
Dates3
Daily positions17
Stages3
CurrencyUSD
Source contract
Validate once. Preserve exactly what was used.
Nine source files cover August 27–29, 2026. Each date contains three small portfolios.
Bad inputs stop the workflow
Required fields, finite values, cross-file IDs, known currencies, permitted frequencies, date ordering, and the ten-position limit.
File and field diagnosticsFailure journalNo silent coercion
Expect exactly the configured inputs.
Reject malformed or inconsistent rows.
Copy inputs and pin configuration.
Bind data, config, code, and predecessor.
System design
A run can resume because its state is explicit.
Python orchestrates stages. Persistent checkpoints make prerequisites and evidence verifiable.
Daily CSVs
+ versioned config
→+ versioned config
PREPARE
→Contract checks
Input + config snapshot
ETL
→Verified input hashes
Deterministic proxies
REPORT
Verified result hashes
Brief + scenario export
Run identityCheckpointsSQLite registryJSONL journalSHA-256 evidence
Same evidence, same run identity
Completed stages are reused. Changed input, configuration, code, or predecessor result produces a different identity.
Check first, then continue
Downstream stages verify their prerequisites. Altered snapshots are rejected before a result is consumed.
Decision signal
The final snapshot exposes a liquidity shortfall.
The reporting sequence progresses from CLEAR to REVIEW to CRITICAL. The dashboard links warnings to evidence.
Bounded assistant
Explanations are anchored to recorded facts.
The briefing layer follows registered tools, verified facts, and explicit reporting rules.
Inspect → cite → explain
Read the verified summary, resolve warnings and source references, then produce a leadership brief and successor questions.
Tool prerequisitesExact metric claimsVisible evidence paths
Review stays with the owner
The reporting layer cannot change results, remove warnings, approve a run, or initiate a financial action.
Pending human reviewSource notes are dataNo hidden provider
Example / from the verified August 29 snapshot
“The modeled liquidity buffer covers 91.8% of net outflow, below the configured 100% requirement.”
result.json → metrics.lcr_pct / metrics.lcr_requirement_pctVerification evidence
Test the failure paths, not just the happy path.
The exported checks exercise the implementation with controlled faults and compare outputs against the accepted baseline.
✓
All v1 USD metrics and portfolio metrics match the archived baseline to six decimals.
✓
Structured values, all warnings and pending approval must match; arbitrary prose stays unverified.
✓
A change to operational source changes the run identity and blocks an old prepared checkpoint.
✓
Live config drift blocks downstream reuse; preparing again preserves the old snapshot as a new run starts.
✓
IDs must be globally unique across all three files, not just unique within each file.
✓
Batch prepare rejects missing predecessor checkpoints before creating any partial runs.
✓
Repeated all reuses exact run identities, artifact bytes and log events.
✓
A concurrent rejected command cannot change the active command's RUNNING state or event history.
Eight selected checks. Full suite in the control tower.
The USD financial proxies reproduce the original scenario. Improvements are in controls, evidence, and interaction.
Checks cover calculation consistency, data integrity, version changes, recovery, and reporting boundaries.
Successor handoff
One project, two ways to inspect the work.
Explore the evidence in the browser or reproduce it locally from the included source.
Understand the operating decision
Compare dates, inspect instruments, explore a liquidity shock, replay actual logs, and question the bounded assistant.
Reproduce the engineering
Run prepare, run-etl, report, or all. Inspect plan, status, and verify outputs. Execute the control suite.
Process the dated synthetic inputs.
Check identity and evidence.
Reconcile exceptions with an owner.
Replace proxies with governed models.
Extension points: approved data adapters, governed valuation and liquidity methods, authenticated review, monitoring, and model-service evaluation.