Showcase: From Case to Verdict (Inspire)

This showcase walks every layer of the engine — from the Gold contract to confirmed verdicts — using a real build of the nuMetrix pack against the Inspire tenant. The numbers, signals, and theses below are not mocked: they come straight from numetrix_inspire_kls.duckdb, the same KLS our internal JinDesk reads.

Inspire is the team’s synthetic but realistically-shaped hospital tenant — the dataset is generated, but the shape mirrors what a mid-sized Swiss hospital looks like in production. We use it to validate every release and as the canonical demo for the nuMetrix pack.

The tenant

Inspire models a hospital that uses OPALE as its source system: cases, billing, materials, suppliers, prices, and the supply organisations (pharmacy, logistics) that own them. The numbers:

Entity	Rows
Cases	29,925
Billing events	1,258,632
Case material usage	1,258,632
Materials	10,000
Article catalogues	30,000
Material movements	60,000
Suppliers	2,500
Cost centers	498
Departments	498
Price observations	30,000
Price palette	100,000
Price series	1,135,887
I/O coefficient (weekly)	496,001
Taxonomy nodes	551

These are the Gold contract — the consumption layer. Signals operate exclusively on Gold; nothing below is visible to the analytical instruments.

Step 1: Signals — detect anomalies

The pack ships 48 signals for nuMetrix. On the Inspire tenant, 22 of them have findings (the rest are inactive against this dataset). Here are the top by finding count:

Signal	Findings	Worst	What it detects
signal_negative_margin	402,830	high	Billing rows where the billed price is below the internal acquisition price
signal_billing_below_internal_price	393,196	high	The same pattern, scoped per material rather than per row
signal_controlled_substance_traceability	36,382	high	Controlled-substance dispensings missing required traceability metadata
signal_cc_producteur_mismatch	29,078	high	Cost-center-of-record disagrees with the producing cost center
signal_cross_site_billing	26,404	high	Material flow crosses site boundaries without a service-mandate justification
signal_duplicate_billing	10,297	high	The same material dispensed and billed twice for one case
signal_active_materials	10,000	low	Catalogue inventory baseline
signal_migel_reimbursement_gap	10,000	high	Materials whose MiGeL reimbursement is below the billed amount
signal_price_stability	10,000	high	Materials whose internal price has drifted beyond stability thresholds
signal_io_coefficient	5,916	high	Single-period I/O imbalance (inputs ≠ outputs) per material-cost-center
signal_io_coefficient_weekly	5,916	high	Weekly-rollup variant of the same
signal_active_single_source	5,029	high	Active materials sourced from a single supplier
signal_single_source_dependency	5,029	high	The same concentration risk, framed as a supplier risk
signal_io_coefficient_trend	816	high	I/O imbalance trending worse over time
signal_drg_cost_outlier	400	high	DRG cases whose cost falls outside the expected band

The rest (small counts) cover catalogue completeness, billing amount outliers, and weekly trend variants.

Step 2: Perspectives — aggregate

Four perspectives roll signal findings up to entity-level health:

Perspective	Entities	What it measures
perspective_material_health	9,080 materials	Overall material risk score combining pricing, sourcing, and compliance signals
perspective_material_compliance	10,000 materials	Regulatory + classification compliance (MiGeL, VESKA classification, etc.)
perspective_catalog_inventory	10,000 materials	Catalogue freshness and inventory balance
perspective_billing_quality	29,312 cases	Billing integrity across cases (margin, duplication, cross-site)

The entity_signal_summary table holds 49,617 entity-signal pairs — every entity that any signal touched, with severity, score, and aggregated risk.

Step 3: Theses — confirmed

The build produced verdicts on 12 theses. Six are confirmed:

Thesis	Status	Findings	Evidence score	Money at risk
thesis_overpriced_generics	confirmed	411,910	0.71	CHF 1.50 B
thesis_supplier_concentration_risk	confirmed	14,109	0.83	CHF 438.6 M
thesis_cross_site_leakage	confirmed	84,794	0.75	CHF 227.5 M
thesis_io_imbalance	confirmed	21,757	0.97	CHF 68.9 M
thesis_controlled_substance_gap	confirmed	46,382	0.71	CHF 9.3 M
thesis_duplicate_billing	confirmed	39,609	0.67	CHF 6.9 M

Six others are not_observed against this synthetic build — the engine evaluated them but didn’t find the pattern (catalogue maintenance, revenue leakage, shelf-life waste, cost-center misallocation, phantom cases, stale catalogue).

A note on the numbers: Inspire is a synthetic tenant, so the absolute money_at_risk values are notional. What’s faithful is the shape — which theses light up, how the evidence scores stack, and which signals carry the weight. On a real hospital build the multipliers shrink but the structure stays.

Step 4: The full chain

For one confirmed thesis, the chain reads top to bottom:

Every box above is a row in the KLS. Click through them in JinDesk and you walk the same path the engine just walked — finding to thesis to verdict, with full provenance.

What this showcase teaches

• The Gold contract is the interface — every analytical instrument sees only Gold; nothing else
• Signals are typed observations with severity, polarity, and money_at_risk — counted, not curated
• Perspectives aggregate signal findings into entity-level health scores
• Theses test business questions against signal evidence and produce a verdict
• The whole chain is declarative YAML compiled to deterministic SQL — the same chain runs on every tenant in the nuMetrix pack

Compare

See the same engine at work on a completely different domain: From Person to Verdict (Val d’Oria) — the hrcentral workforce pack against the Val d’Oria Academic Institutions tenant. Different entities, different signals, same instrument chain.