A model driver or line-item behaviour
Its exact impact on financial statement line items, cash flows and balance-sheet integrity is formally specified.
Institutional knowledge architecture for deterministic corporate finance modelling
01Purpose
Generative AI is probabilistic. Institutional corporate finance requires deterministic modelling logic, verifiable provenance and reproducible audit trails.
The COASE Knowledge Layer (KL) is the versioned engineering repository underpinning all COASE financial modelling, advisory and recurring Fractional FP&A engagements. It operates as an indirect control plane: it codifies the firm’s authoritative accounting, valuation, financial and procedural logic without autonomously populating client models or issuing deliverables. Numerical execution remains exclusively within controlled calculation engines and financial models; professional judgement and sign-off remain with the responsible Partner.
Primary sources
COASE expertise
Knowledge LayerTyped knowledge graph · immutable releases
Execution · client-isolated
A knowledge object enters the KL only if it maps directly to one of two node types.
Its exact impact on financial statement line items, cash flows and balance-sheet integrity is formally specified.
Its control or deliverable criterion is unambiguously testable.
Advisory mandates and recurring Fractional FP&A engagements draw on the same knowledge base.
PracticesPlanning & ControlDebt & Equity FinancingTransactions & ValuationRestructuringModel & Plan Review
02Assurance
Five risk dimensions and the mechanism that controls each. The sections that follow document how every control is implemented.
Financial computations, ratio reconciliations and balance-sheet checks execute exclusively within controlled financial models and external deterministic calculation services. AI is strictly bounded to retrieval, comparison and structured drafting.
05 MCP interface →Every modelling rule and technical treatment is bound to an indexed primary source, with exact edition, effective date and paragraph anchors.
03 Architecture →Client financials, credit files and contractual terms reside in isolated, access-controlled Engagement Workspaces that reference immutable KL snapshots through read-only pointers.
03 Architecture →Every rule tracks both the legal or accounting regime in force at the valuation date and the approved KL release. Any historical business plan, valuation or debt sustainability analysis can be reconstructed.
04 Ontology →Draft content carries zero evidential weight and is invisible to engagement workflows. Production release requires automated schema validation, non-contradiction checks and audit-logged Partner approval.
06 Governance →03Architecture
The ecosystem enforces a strict architectural separation between the shared institutional knowledge base and client-specific execution environments.
Catalogues institutional primary sources across three tiers. Each source is indexed with edition metadata, effective dates and paragraph-level anchors. Full-text indexed sources support verbatim passage binding; citation-only sources are flagged as non-binding until full-text ingestion is completed.
Stores atomic, reusable definitions, modelling rules, contextual accounting treatments, management-versus-contractual metric variants, review procedures and audit checklists across ten domains.
Maintains structured numerical observations tagged with source provenance, observation period, geography, peer population, units and licensing or permitted use. Inclusion in any engagement requires an explicit comparability and relevance sign-off.
Exposes curated graph traversals for each of the 29 services, assembling the exact sequence of required data requests, analytical methodologies, modelling rules, quality checks and deliverable specifications.
CatalogueBrowse the 29 services
Read-only pointers to immutable KL release snapshots
No client data ingested, indexed or written back
Physically and logically isolated, access-controlled workspaces holding client enterprise data, contractual definitions, engagement assumptions and deliverable drafts. Workspaces reference immutable KL release snapshots through read-only pointers.
Client data is never ingested into, indexed by, or written back to the KL.
Every modelling rule and technical treatment is bound to its primary source by edition, effective date and paragraph anchor.
Civil, tax and insolvency law applicable to the specific Cell. Examples include the Italian Civil Code, the CCII and the TUIR.
SourcesTax and regulationBusiness crisis and restructuringGovernance and controls
The national or international standards governing the calculation. Examples include OIC 9, OIC 15 and OIC 19, or IFRS 9 and IFRS 16, as applicable.
SourcesAccounting standards and reportingValuation standards
The empirical sources used to calibrate assumptions. Examples include ECB yield curves, FRED time series and Damodaran datasets.
SourcesData, markets and benchmarks
04Ontology
The KL is structured as a typed knowledge graph governed by a formal ontology. Entities are linked through directed, typed relationships, each carrying provenance metadata and review status.
Institutional disputes frequently arise from definitional drift between financial statements, management plans and credit agreements. The KL addresses it by engineering knowledge in Cells: the atomic intersection of four coordinates.
ResultCodifying by Cell guarantees cross-domain consistency: a single financial aggregate’s accounting treatment, cash-flow impact, covenant sensitivity and audit check are resolved simultaneously rather than in silos.
Every node and relationship maintains two time dimensions. Together they support two operational guarantees.
The period during which a statute, accounting standard or market rule is legally or technically in force.
The immutable KL release tag under which the node was approved.
Any past valuation, restructuring plan or model audit can be reconstructed using the exact KL release and statutory regime in force at the original valuation date.
When a primary source is revised or superseded, graph traversal immediately flags every dependent Modelling Logic Node, Service View and active Engagement Workspace affected by the change.
05MCP interface
The KL operates as a versioned knowledge service exposed to authorised host applications through a Model Context Protocol (MCP) server. Interoperability is standardised across three primitives.
Stable, versioned URIs exposing approved nodes, paragraph-level source excerpts, metric definitions and Service Views.
Schema-validated endpoints for hybrid retrieval, impact analysis, citation verification and delegation to external deterministic calculation services: model checks, ratio reconciliations and formula validation.
Pre-configured, auditable prompt templates for comparative treatment analysis, gap identification and technical drafting within approved boundaries.
Security and governance are never delegated to prompt instructions. Hard filters are enforced at the server layer before any token enters the language model context window.
APPROVED only for production engagements06Governance
Nodes follow a strict state machine. Draft nodes carry zero evidential weight, are invisible to engagement queries and cannot support client deliverables.
Zero evidential weight. Excluded from production queries.
Automated checks cleared. Excluded from production queries.
The only state that production engagements can query.
Excluded from production queries.
A node is promoted to APPROVED only when four mandatory gates are cleared.
For Modelling Logic Nodes, the exact impact on financial statement line items, cash flows and balance-sheet integrity is formally specified. For Procedural Nodes, the control or deliverable criterion is unambiguously testable.
Every normative or technical claim is bound to a verified source identifier and paragraph anchor.
Automated checks verify ontology compliance (schema and relationship constraints), bitemporal metadata completeness and non-contradiction with existing approved nodes.
Direct, audit-logged approval by a responsible Partner.
LLM usage within the KL engineering pipeline is strictly tiered and bounded, to exclude full-document hallucination and uncontrolled drift.
Automated metadata tagging, taxonomy classification and structural segmentation during ingestion.
Initial drafting of atomic nodes and relationship proposals, strictly from paragraph-level extracts.
Complex ontology alignment, cross-framework comparative checks such as OIC versus IFRS deltas, and pre-review consistency audits.
RuleLLMs operate exclusively on segmented, paragraph-level extracts, never on unparsed full documents. All AI-generated artefacts remain non-authoritative until cleared by deterministic validation and Partner approval.
Every MCP request is authenticated against engagement-level permissions. Immutable audit logs record what backs each engagement output, ensuring full inspectability for internal review, statutory auditors and institutional counterparties.
07Use cases
The critical work sits between disciplines: a treatment that is valid in one context can distort another. These composite mandate scenarios illustrate the dependency structures the KL is designed to support.
OIC–IFRS conversion, purchase price allocation and acquisition financing.
Accounting · Valuation · Tax · Debt
An IFRS buyer acquires an OIC industrial group with capitalised development expenditure, material leases and a financing agreement using its own EBITDA and net-debt definitions.
OIC–IFRS bridge · IAS 38 development criteria · IFRS 16 lease measurement · IFRS 3 allocation · IAS 12 tax bases · valuation methodology · SPA and financing definitions.
Link the conversion bridge to comparable EBITDA and enterprise value, identifiable net assets and goodwill, deferred tax, the equity purchase-price bridge and post-deal covenant capacity.
A lease treatment must remain coherent across earnings, multiples, debt deductions and contractual ratios. Acquisition accounting, price mechanics and covenant definitions require separate bindings.
Decision outputReconciled acquisition model with traceable valuation, purchase-price and financing sensitivities.
Conditional debt relief, repayment restructuring and downside viability.
Liquidity · Insolvency · Tax · Accounting
A distressed group negotiates tax-debt relief and lender rescheduling while its short-term liquidity depends on delayed supplier payments and uncertain collections. The plan also supports asset carrying values.
CCII procedural conditions · OIC 19 debt accounting · applicable tax treatment · OIC 9 impairment · corporate-capital constraints · creditor terms · treasury and covenant methodology.
Bind legal effectiveness and creditor conditions to debt recognition, tax effects and repayment schedules. Reconcile the 13-week forecast with the integrated plan, debt-service coverage, impairment sensitivities and equity movements.
Negotiated relief, accounting recognition and cash settlement have distinct triggers. Viability must remain testable when relief is delayed, collections deteriorate or deferred supplier payments unwind.
Decision outputConditional restructuring scenarios connecting liquidity, capital requirements and financing feasibility.
Inventory absorption, capitalised revamping, separation costs and debt capacity.
Operations · Accounting · Tax · Transactions
An industrial division shows margin expansion despite flat sales. The perimeter includes accumulated inventory, capitalised revamping, a claimed tax incentive, shared infrastructure and a post-separation investment programme.
OIC 13 inventory measurement · OIC 16 or IAS 16 asset criteria · incentive eligibility evidence · production and cost-allocation methods · carve-out perimeter · valuation and financing definitions.
Connect production volumes and absorbed costs to earnings quality; capitalisation and incentive evidence to asset and cash adjustments; separation assumptions to standalone margins, working capital, capex and debt service.
The same adjustment can affect EBITDA, inventory, tax, investment needs and the equity bridge. Every effect must be assigned to the correct perimeter and period without duplication.
Decision outputStandalone operating and financing model linking earnings quality to valuation and capital requirements.
Composite scenarios for illustration. They are not results of actual engagements.
08Engine Tools
Our proprietary tools automate time-consuming tasks and high-volume workloads across data preparation, model construction and reporting.
09FAQ
For IT security, compliance and model risk management teams.
No. The architecture enforces a strict physical and logical boundary between the shared KL and client execution environments.
Client financials, credit files and contractual agreements reside exclusively in segregated, access-controlled Engagement Workspaces. They are never ingested into, indexed by, or written back to the KL.
Engagement Workspaces query immutable KL release snapshots strictly through read-only pointers.
LLM inference operates under enterprise zero-data-retention terms. Models within the KL pipeline receive only segmented, paragraph-level excerpts of approved public or institutional sources, never full proprietary documents or unredacted client ledgers.
By decoupling language processing from numerical execution. The KL enforces a Zero AI Arithmetic policy.
All financial computations, cash-flow reconciliations, balance-sheet integrity checks and ratio calculations execute exclusively inside controlled financial models and external deterministic calculation services invoked through typed MCP Tools.
LLMs are restricted to structured retrieval, cross-framework comparison such as OIC versus IFRS deltas, and technical drafting strictly within pre-filtered, approved context.
Every output generated through the MCP server is required to return a structured evidence payload containing node IDs, KL release tags, exact paragraph-level source citations and explicitly surfaced assumptions. Claims without a bound source anchor are flagged as unsupported.
Through hard pre-retrieval filtering at the server layer. Security and governance are never delegated to prompt instructions.
Every MCP request is authenticated against engagement-level permissions before any query reaches the knowledge graph or vector store.
Production engagements can exclusively query nodes in the APPROVED state. DRAFT, SCHEMA_VALIDATED and DEPRECATED nodes are excluded at the query-engine level and cannot enter the LLM context window.
Jurisdiction, accounting framework (for example Italy / OIC or EU / IFRS) and effective date are applied as mandatory hard filters prior to GraphRAG traversal, preventing cross-regime contamination.
Yes. The KL operates on a bitemporal versioning model paired with immutable audit logging.
Every node records the exact period during which the underlying statute, accounting standard or market rule was legally or technically in force.
Every engagement pins its references to an immutable KL release tag.
For every analytical step, the system logs the exact KL release version, retrieved node IDs, applied filters, MCP tool invocations and external calculation service versions. Statutory auditors, credit committees and regulators can reconstruct the exact normative and modelling logic in force on the original valuation date.
Graph-based impact analysis automatically flags affected logic and engagements.
Primary sources are linked to Modelling Logic Nodes, Service Views and active Engagement Workspaces through typed graph relationships (SUPPORTED_BY, DEPENDS_ON, SUPERSEDES). Updating or deprecating a source paragraph triggers an automated dependency traversal.
Every impacted modelling rule is flagged for review and re-validation, and active engagements pinned to the superseded rule receive an explicit change alert. Existing engagement snapshots are never silently overwritten; migrating an active engagement to a new KL release requires explicit Partner authorisation.
AI has zero signing authority. The KL operates strictly as an indirect control plane.
No rule, accounting treatment or procedural check enters production in the KL without passing automated schema and non-contradiction validation, followed by audit-logged Partner sign-off.
AI assists in locating authoritative treatments, checking consistency and surfacing missing evidence. Professional judgement, interpretation of contractual edge cases and final sign-off on all client models and deliverables remain exclusively with the responsible COASE Partner.
Further technical questions: [email protected]
The Knowledge Layer underpins how COASE develops financial models, analysis and reporting for each assignment.