artifacts/incoming

FSGF-0.2 — Finance Semantic Grounding Framework

artifacts/incoming/fsgf_0.md

Rendered from markdown source. Open raw source on GitHub.

FSGF-0.2 — Finance Semantic Grounding Framework

Status

Second-draft anchor.

Purpose

The Finance Semantic Grounding Framework (FSGF) defines how financially material language becomes machine-reliable, human-auditable, context-stable, and runtime-maintainable.

Its purpose is to ensure that every materially relevant finance term used by humans, systems, and AI agents is bound to explicit meaning rather than assumed familiarity.

A grounded term must be bound to:

a semantic identity
a context envelope
an authority source
a transformation lineage
an ambiguity state
a governance owner
a maintenance and revalidation posture

The framework exists to prevent silent semantic drift in high-stakes financial operations.

---

Core principle

A finance term is not grounded by familiarity. It is grounded by binding.

Compactly:

Grounded finance language = meaning + context + authority + lineage + ambiguity + reviewability

---

Why this framework exists

Finance language often appears stable while actually remaining context-dependent.

Examples:

Revenue
EBITDA
Operating Expense
Free Cash Flow
Recognized ARR
Adjusted EBITDA
Headcount Cost

These terms may vary across:

accounting basis
internal policy
entity scope
reporting context
time validity
source system
intended audience

Humans often reconcile this implicitly. Machines usually do not.

FSGF exists to prevent this failure mode:

surface linguistic sameness masking materially different financial meaning

---

Foundational assumptions

1. Meaning must be scoped

No financially material term is globally self-evident.

2. Translation is first-class

Local vocabularies are not errors. They are real operating conditions and must map explicitly to canonical meaning.

3. Lineage outranks wording

A term is not grounded because its label feels familiar. It is grounded because its derivation and use constraints are inspectable.

4. Ambiguity must be representable

The system must be able to preserve unresolved, overloaded, and conditional meaning rather than silently collapsing it.

5. Definitions are versioned artifacts

Meanings change over time and must remain historically legible.

6. Use-context matters

Statutory, management, board, investor, planning, and ad hoc analytical contexts may validly differ.

7. Grounded meaning requires maintenance

A meaning object is not self-sustaining simply because it was once defined. It must be periodically reviewed, revalidated, and checked against drift.

---

Relationship to adjacent layers

FSGF

Defines what a grounded financial meaning object is.

FSP

Defines how that meaning is symbolically referenced in text.

SRE

Maintains live binding, drift detection, and runtime safety for grounded meaning.

Compactly:

FSGF = what meaning is
FSP = how meaning is pointed to
SRE = how meaning is kept alive

---

Framework stack

Layer 0: Surface term

The human-visible label.

Examples:

EBITDA
Revenue
Opex
Free Cash Flow

This layer is unstable by default.

Layer 1: Semantic identity

A canonical meaning object independent of local phrasing.

Example:

CANON.METRIC.EBITDA.ADJUSTED.INTERNAL.V3
CANON.REVENUE.RECOGNIZED.EXTERNAL.GAAP.V2

This is the anchor layer.

Layer 2: Context envelope

The conditions under which the semantic identity is valid.

Examples:

accounting basis
organizational scope
jurisdiction
intended use
time validity
currency basis
inclusion / exclusion rules

Layer 3: Authority

What legitimizes the definition.

Examples:

GAAP / IFRS standard
tax rule
internal accounting policy
CFO-approved metric definition
board reporting policy
approved transformation logic

Layer 4: Lineage

How raw inputs become this semantic object in use.

Examples:

source systems
transformation logic
adjustments
eliminations
rollups
overrides

Layer 5: Ambiguity state

How semantically clear, contested, or overloaded the term is.

Layer 6: Governance and maintenance

Who owns the term, how often it must be reviewed, and how drift is detected.

---

Grounded term requirements

A financially material term is grounded only when all of the following are explicit:

canonical semantic identity
valid context envelope
authority source
lineage path or lineage reference
ambiguity state
governance owner
review posture

Any missing element downgrades semantic trust.

---

Canonical semantic object model

Each grounded finance term should be represented as a semantic object.

1. Identifier

canonical ID
semantic version
status

2. Surface forms

preferred label
aliases
abbreviations
deprecated synonyms
prohibited substitutions

3. Definition

canonical definition
plain-language definition
examples
counterexamples

4. Scope

reporting basis
entity or business scope
geography or jurisdiction
time validity window
currency basis
intended use

5. Authority

authority type
policy or standards reference
approval state
approving function or person

6. Lineage

source systems
transformation logic references
rollup relationships
predecessor and successor objects
manual override conditions

7. Ambiguity metadata

ambiguity state
common confusions
nearby terms
mapping caveats
confidence notes

8. Governance

semantic steward
lineage owner
review cadence
last reviewed date
deprecation policy

9. Runtime posture

materiality tier
revalidation horizon
expected decay rate
resolver mode defaults

---

Context envelope schema

A term is not grounded unless its context envelope is explicit.

Required context fields

reporting basis
organizational scope
time basis
currency basis
inclusion / exclusion policy
intended use

Optional but often important context fields

jurisdiction
consolidation state
audience
materiality threshold
scenario type (actual, forecast, plan, pro forma)
system of record

Example context envelope

{
  "reporting_basis": "management",
  "organizational_scope": "consolidated_company",
  "time_basis": "fiscal_quarter",
  "currency_basis": "USD_reporting",
  "intended_use": "board_review",
  "inclusion_exclusion_policy": "FIN-MET-014.V3"
}

---

Semantic classes

Different classes require different grounding rigor.

Primitive ledger classes

Examples:

account
journal entry
legal entity
cost center
customer
vendor
currency

These require strong structural grounding.

Policy-defined classes

Examples:

recognized revenue category
capitalizable expense
bad debt reserve
material adjustment

These require date-aware policy grounding.

Management-defined metric classes

Examples:

Adjusted EBITDA
contribution margin
productivity spend
normalized headcount cost

These require explicit inclusion and exclusion logic.

Derived KPI classes

Examples:

ARR
CAC
LTV
Free Cash Flow
Rule of 40

These require formula lineage and presentation context.

Narrative classes

Examples:

one-time
exceptional
normalized
healthy
efficient

These must never quietly masquerade as grounded objects.

---

Grounding tiers

Not all terms require the same rigor.

Tier 1: Financially determinative

Used in books, filings, tax, or audit-critical outputs.

Required:

canonical identity
authority source
lineage
full context envelope
prohibited substitutions
named steward
review cadence
runtime revalidation posture

Tier 2: Management-critical

Used in board, forecasting, operating review, or executive decisions.

Required:

canonical identity
explicit formula or composition logic
use context
exclusions / adjustments
steward
review cadence

Tier 3: Analytical

Used in internal exploration and ad hoc analysis.

Required:

provisional or canonical definition
owner
intended scope
ambiguity disclosure where relevant

Tier 4: Narrative

Used rhetorically or descriptively.

Required:

mapping to grounded object where possible
or explicit non-authoritative designation

---

Mapping model

Local terms must not be assumed equivalent to canonical identities.

The framework must support explicit mapping records.

Mapping types

exact
narrower-than
broader-than
conditional
approximate
deprecated alias
prohibited

Mapping conditions

Mappings may depend on:

report type
source system
legal entity
business unit
policy version
date range
intended use

Example mapping logic

A local label such as Revenue may be:

exact in statutory P&L context
approximate in board-pack commentary
prohibited as a substitute for Bookings

---

Ambiguity protocol

The framework must preserve ambiguity instead of erasing it.

Ambiguity states

unambiguous
ambiguous
overloaded
conditional
deprecated
contested
unresolved

Ambiguity handling rules

ambiguity must be surfaced, not hidden
unresolved meaning must not be treated as authoritative
ambiguous mappings must carry candidate bindings where available
high-materiality ambiguity must escalate to review

---

Governance model

A semantic framework fails if nobody owns it.

Semantic steward

Owns term quality and definition integrity.

Policy authority

Approves financially binding definitions.

Lineage owner

Owns source mapping and transformation references.

Runtime governor

Defines allowed runtime and AI usage for the term.

Audit witness

Ensures inspectability across review, replay, and accountability workflows.

---

Maintenance and decay posture

FSGF should explicitly acknowledge that semantic objects decay if not maintained.

Each grounded object should include:

last review date
review cadence
revalidation horizon
decay risk
deprecation path

Example decay signals

policy superseded
lineage changed
repeated local overrides
unresolved alias collisions
organizational restructuring
change in intended use

This allows SRE to operate against a framework that was designed for maintenance rather than assumed permanence.

---

AI operating modes against FSGF

Different AI actions require different grounding thresholds.

Retrieval

AI may explain grounded terms.

Translation

AI may map local terms to canonical terms if mapping confidence is sufficient.

Composition

AI may generate summaries or analyses using grounded terms when all materially relevant terms are sufficiently bound.

Recommendation

AI may suggest classification or treatment only when policy-backed reasoning exists and review posture is explicit.

Autonomous action

AI may act only with deterministic execution controls, explicit policy authorization, lineage capture, and rollback support.

---

Artifact binding requirement

A grounded framework must attach to real artifacts, not only definitions in isolation.

Artifacts may include:

spreadsheets
reports
dashboards
models
board materials
filings
reconciliation workpapers

A finance term reaches practical grounding only when it can be bound to real artifacts and referenced consistently across them.

---

Minimum viable implementation

Phase 1: Critical term register

Define 50–100 high-value finance terms.

For each term:

canonical identity
preferred definition
aliases
context envelope
authority source
steward
ambiguity notes
review cadence

Phase 2: Report grounding

Apply the framework to:

P&L
balance sheet
cash flow
key board metrics
forecast glossary

Phase 3: System mapping

Map ERP labels, warehouse metrics, dashboard labels, and planning models to canonical semantic objects.

Phase 4: Runtime integration

Enable SRE and AI workflows to use the semantic layer as an operational dependency rather than a passive glossary.

---

Evaluation criteria

The framework is working when:

materially important terms have explicit, inspectable meaning per context
conflicting usages are visible instead of latent
reports can declare semantic versions
AI outputs can ground themselves in semantic objects rather than linguistic guesses
audits can trace not only numeric origin, but semantic validity
reorganizations and system migrations do not erase financial meaning

---

Failure modes

over-normalization that flattens meaningful distinctions
ontology theater disconnected from actual workflows
static glossary syndrome with no binding to systems or artifacts
hidden local drift outside steward review
false AI confidence built on approximate mappings
failure to model semantic maintenance and decay

---

Compact definition

The Finance Semantic Grounding Framework is the structural layer that binds financially material language to explicit meaning, context, authority, lineage, ambiguity, and governance so that financial reasoning can be safe, inspectable, and machine-usable.

---

Protocol statement

No financially material term should be treated as authoritative solely by linguistic familiarity. Authoritative use requires explicit semantic identity, scoped context, governing source, lineage relationship, ambiguity handling, and review ownership.

---

Anchor note

This document is intended as the next stable anchor for FSGF. It has been brought forward to align more tightly with FSP and SRE, especially around governance, artifact binding, runtime posture, and semantic maintenance over time.