01 · Context
The decision loop
A management question moves through discussion, spreadsheet edits, a model run, email and a meeting. Each hand-off can lose assumptions, ownership and context.
1
Business question
2
Scenario assumptions
3
Indicative impact
4
Governed run
5
Comparison
6
Explanation
7
Approval
8
Decision
9
Audit trail
02 · Architecture
Seven architecture rules
01Python is the source of truth for SCR, VaR, TVaR, standalone capital, allocation and reinsurance.
02Foundry reshapes and presents outputs; it never recreates a competing capital calculation.
03Every governed run is immutable: run_id, versions, seed, hashes and timestamps.
04Simulation rows live in datasets. Only decision-relevant entities become objects.
05Indicative outputs carry a visible "not a governed model result" label.
06Passing technical checks does not approve a run.
07A changed scenario creates a new version and needs fresh governed evidence.
02 · Architecture
Indicative and governed lanes
Indicative
"Which way might capital move, and roughly by how much?"
Stored scenarios, marginal curves, interpolation. Fast, and never changes the approved record.
Shows: method · source runs · valid range · time generated · warning label
Governed
The Python engine with actual dependencies, reinsurance, seeds and controls.
Persists the full output pack, reconciles, and registers a new immutable run.
Shows: versions · hashes · seed and simulation count · risk measure · control and approval state · lineage
Example
A Cyber sensitivity estimates SCR +£12m. The governed run shows +£5m, because extra losses enter a stop-loss layer and change diversification.
02 · Architecture
What goes where in Foundry
STEER concept
Foundry representation
Excel input workbook
Versioned file in a raw dataset
Python capital engine
Python transforms Code Repository
Full simulation matrix
Large dataset, not a business object
Capital model run
Capital Model Run object
Capital breakdown
Capital Breakdown object
Scenario and management decision
Scenario and Decision objects
Approval or request
Ontology Action, logged writeback
Management interface
Workshop module
03 · Data layer · Lessons 0–1
Foundry access and project setup
STEER – Capital Decision Platform
00_Documentation
01_Raw_Inputs
02_Raw_Model_Outputs
03_Pipelines
04_Curated_Data
05_Ontology
06_Applications
07_Run_Operations
08_Evidence_Screenshots
S00 · Foundry navigation
S01 · Project structure
03 · Data layer · Lesson 2
Loading one governed run
First load pack
run_registryRequired
capital_breakdownRequired
business_class_summaryRequired
checksRequired
reinsurance_summaryRecommended
tail_forensicsRecommended
Every file carries the same run_id. Raw values are never edited in Foundry.
S02 · Raw run registry in Dataset Preview
03 · Data layer · Lesson 3
Core curation pipeline
S03 · PB_STEER_Core_Curated in Pipeline Builder
Rename and cast
Snake-case headers, explicit types for dates and capital fields
Join on run_id
Only where needed; no copying metadata onto 50,000 rows
No hidden defaults
A failed cast is fixed at source, not masked
03 · Data layer · Lesson 4
Curated outputs and data expectations
S04 · Successful curated build
Six app_ datasets
Published to 04_Curated_Data to back the Ontology
Primary-key checks
run_id and breakdown_id: unique and non-null
Deterministic keys
Built from run, dimension and metric, never row numbers
03 · Data layer · Lesson 5
Lineage, health and refresh
S05 · Data Lineage, raw to curated
Lineage
app_steer_run_registry back to the raw upload and pipeline
Health check
Build status on the curated run registry
Safe refresh
Manual or parent-update build, not a needless daily schedule
04 · Ontology · Lesson 6
Capital Model Run object type
S06 · Capital Model Run in Ontology Manager
Backed by
app_steer_run_registry
Identity
run_id as primary key and title
Properties
SCR, VaR, TVaR, versions, hashes, seed, control and approval status
04 · Ontology · Lesson 7
Runs, breakdowns and business classes
S07 · Capital Model Run linked to Capital Breakdown
Capital Breakdown
By class, risk type and risk driver
Foreign-key link
Belongs to Run · Has Capital Breakdowns
Simulations stay data
No simulation-row object type
04 · Ontology · Lesson 8
Scenario and decision chain
S08 · Scenario, Decision and Capital Model Run
Decision links
Supported by Run · Selects Scenario
Scenario lifecycle
DRAFTINDICATIVE_READYRUN_REQUESTEDRUNNINGAWAITING_REVIEWAPPROVEDREJECTEDSUPERSEDED
05 · Application · Lessons 9–12
Four-page Workshop application
Executive Overview
Governed Runs
Compare Runs
Decision Record
06 · Governed execution · Lessons 13–18
Governed run and approval
1Request Governed Run action
2Schedule picks up the queued request
3Python transform runs the STEER engine
4Output pack persisted and reconciled
5New immutable run, AWAITING_REVIEW
6Approve Decision action, by a named reviewer
Python transform in Code Repositories
07 · Delivery
Build roadmap
0Access and navigation
1Workspace
2Load one governed run
3Curate and validate
4Business entities
5Management interface
6Compare stored scenarios
7Indicative sensitivities
8Host Python execution
9Review workflow
10Reinsurance showcase
11Portfolio steering
12AIP explanation
13Package the portfolio
RequiredRecommendedOptional
Questions
Place every pound with intent.
STEER · Foundry build walkthrough