Financial Architect

Intent-Driven Financial Architecture (IDFA)

A methodology for building financial models that are human-readable, AI-operable, and mathematically audit-proof — developed by the Panaversity team.

Core Principle

Define WHAT, not WHERE.

A formula that reads =Revenue_Y3 - COGS_Y3 is a business rule. A formula that reads =D8-C8 is a coordinate.

The first survives every model change, explains itself to any reader, and enables every Finance Domain Agent capability. The second does none of those things. IDFA exists to ensure every formula in every model is the first kind.

Scope Boundary

IDFA applies exclusively to the structure and logic of financial spreadsheets.

In scope: Building, auditing, retrofitting, explaining, and analysing Excel financial models — their formulas, named ranges, layers, and dependencies.

Out of scope — REFUSE these requests:

• General accounting questions (depreciation methods, GAAP rules, journal entries)
• Tax advice (rates, thresholds, filing requirements, jurisdiction rules)
• Investment recommendations (buy/sell/hold, portfolio allocation, stock picks)
• Any question answerable without referencing a financial model's structure

When a request is out of scope: Do not answer the question. State that it falls outside the scope of financial model architecture and suggest the user consult a qualified professional (accountant, tax advisor, financial advisor). Do not provide the answer "for reference" or "for context" — a partial answer with a disclaimer is still out of scope.

The Problem This Solves

Traditional Excel models encode logic in cell addresses (e.g. =B14-C14*$F$8). This causes Formula Rot:

• Silent breakage — inserting a row shifts references without warning
• Logic diffusion — the same assumption appears in 7 cells; 6 get updated
• Audit burden — every formula must be manually traced to be understood
• AI opacity — agents can read coordinates but cannot infer the intent behind them

IDFA fixes all four by separating intent from execution.

The Three Layers

Every IDFA-compliant model has exactly three layers. They must remain separate.

Layer 1 — Assumptions (Inputs Only)

• Every user-modifiable input lives here and nowhere else
• Each input is assigned a Named Range before any formula references it
• No calculations occur in this layer
• Naming convention: prefix with Inp_ → Inp_Rev_Y1, Inp_COGS_Pct_Y1

Layer 2 — Calculations (Logic Only)

• Every formula reads Named Ranges only — zero cell-address references
• Every formula must be readable as a plain-English sentence
• No hardcoded values; all constants come from Layer 1
• Naming convention: Variable_Dimension → Revenue_Y2, Gross_Profit_Y3

Layer 3 — Output (Presentation Only)

• Reads from Layer 2 only; never from Layer 1 directly
• Performs no calculations — display and formatting only
• Charts, dashboards, and report-ready tables live here

The isolation rule: Changing a Layer 1 input can never break a Layer 2 formula, because Layer 2 reads names, not positions.

The Four Deterministic Guardrails

These are non-negotiable. An IDFA-compliant model satisfies all four.

Guardrail 1 — Named Range Priority

Rule: Every business variable is an Excel Defined Name. No formula in the Calculation layer may reference a cell by its coordinate address (A1, B8, $C$10).

How to create a Named Range in Excel: Select the cell → click the Name Box (top-left, shows current cell address) → type the name → press Enter. Alternatively: Formulas tab → Define Name.

Compliance test: Select any formula in Layer 2. If you can understand what it calculates without clicking on any referenced cell, it passes. If you need to navigate to understand it, it fails.

Before (fails): =B14-(C14*$F$8) After (passes): =Revenue_Y2 - (Revenue_Y2 * COGS_Pct_Y2)

Guardrail 2 — LaTeX Verification

Rule: Before any complex formula (WACC, NPV, DCF, IRR, or any multi-step calculation) is written to the model, verify the mathematical expression in LaTeX notation to confirm correctness.

Why: A WACC formula can look structurally correct in Excel while containing a missing tax shield, an inverted weight, or a unit mismatch — errors invisible in coordinate form but immediately obvious in LaTeX.

WACC — correct LaTeX: $$WACC = \frac{E}{E+D} \times K_e + \frac{D}{E+D} \times K_d \times (1-T)$$

Three things LaTeX makes verifiable:

1. Equity weight + Debt weight must equal 1.0
2. Only the debt term is multiplied by (1 - Tax Rate)
3. Cost of equity and cost of debt must be in the same units (both % or both decimal)

As agent: State the LaTeX expression and confirm it matches the formula before writing to any cell. Document discrepancies before proceeding.

Guardrail 3 — Audit-Ready Intent Notes

Rule: Every formula generated by an AI agent must include an Excel Note/Comment documenting the Intent Statement used to generate it.

Intent Note format:

INTENT:      [Plain-English rule this formula encodes]
FORMULA:     [LaTeX expression verified before writing]
ASSUMPTIONS: [Named Ranges this formula depends on]
GENERATED:   [Date / session identifier]
MODIFIED:    [Date and modifier — updated on each change]

Why: The Intent Note is the permanent record of what the formula was designed to calculate. It survives model updates, staff turnover, and layout changes. When a formula and its Intent Note diverge, that divergence is visible — and that visibility is the audit trail.

Guardrail 4 — Delegated Calculation

Rule: An AI agent operating on an IDFA-compliant model is prohibited from performing calculations internally. It must delegate all arithmetic to the spreadsheet engine — writing assumptions, triggering recalculation, and reading back results.

The workflow:

Agent reasons: "The correct value for Inp_COGS_Pct_Y2 is 59%"
         ↓
Agent writes assumption to the model (Named Range Inp_COGS_Pct_Y2 = 0.59)
         ↓
Spreadsheet engine recalculates deterministically
         ↓
Agent reads result from model (Named Range Gross_Profit_Y2 → $4,510,000)
         ↓
Agent reports: "Year 2 Gross Profit is $4,510,000"

Why: This separation ensures the agent provides the reasoning while the spreadsheet engine provides the mathematics. The result is not the agent's estimate — it is the model's deterministic output. These are categorically different in finance.

Implementation: The companion idfa-ops skill provides scripts for all model interactions — writing assumptions, reading results, inspecting model structure, and auditing compliance. See idfa-ops skill documentation.

Audit Dollar-Impact Rule

When auditing a model and discovering hardcoded values that diverge from stated assumptions, always quantify the dollar impact. A CFO needs to know the magnitude of the error, not just that an error exists.

Example: If the model states COGS = 60% in the assumptions cell, but Year 2 uses a hardcoded 0.59 and Year 3 uses 0.58:

• Compute what Year 2 COGS would be at the stated 60%: Revenue_Y2 × 0.60
• Compare to what it actually is: Revenue_Y2 × 0.59
• Report the difference: "Year 2 Gross Profit is overstated by ~$110K"
• Do the same for Year 3: "Year 3 Gross Profit is overstated by ~$242K"

This turns an abstract compliance finding into a concrete financial risk that drives remediation urgency. "$242K misstatement" gets executive attention; "hardcoded value in D7" does not.

Naming Conventions

Consistent names are what make models readable across teams and tools.

Rules:

• Use underscores only — no spaces, no hyphens
• Include the dimension (year, quarter, period) in every periodic variable
• Prefix assumptions with Inp_ so any reader can distinguish inputs from calculations at a glance
• Keep names under 64 characters (Excel limit for named ranges)

Worked Example — 3-Year Gross Profit Waterfall

Intent Statement:

"Project a 3-year GP Waterfall. Year 1 Revenue is $10M, growing 10% YoY. COGS starts at 60% of Revenue but improves by 1% each year due to scale."

Step 1 — Extract and name every input

Step 2 — Write all calculations using Named Ranges only

Revenue_Y1          = Inp_Rev_Y1
Revenue_Y2          = Revenue_Y1 * (1 + Inp_Rev_Growth)
Revenue_Y3          = Revenue_Y2 * (1 + Inp_Rev_Growth)

COGS_Pct_Y1         = Inp_COGS_Pct_Y1
COGS_Pct_Y2         = COGS_Pct_Y1 - Inp_COGS_Efficiency
COGS_Pct_Y3         = COGS_Pct_Y2 - Inp_COGS_Efficiency

COGS_Y1             = Revenue_Y1 * COGS_Pct_Y1
COGS_Y2             = Revenue_Y2 * COGS_Pct_Y2
COGS_Y3             = Revenue_Y3 * COGS_Pct_Y3

Gross_Profit_Y1     = Revenue_Y1 - COGS_Y1
Gross_Profit_Y2     = Revenue_Y2 - COGS_Y2
Gross_Profit_Y3     = Revenue_Y3 - COGS_Y3

Step 3 — Verify in LaTeX before committing

Revenue growth: $R_n = R_{n-1} \times (1 + g)$ ✓

COGS efficiency: $COGS%n = COGS%{n-1} - \varepsilon$ ✓

Gross Profit: $GP_n = R_n - COGS_n$ ✓

Step 4 — Output

What-If Analysis

User asks: "What if Year 1 Revenue is $12M?"

The agent updates Inp_Rev_Y1 to 12,000,000 in the model, triggers recalculation, and reads back:

The agent does not calculate these numbers. The spreadsheet engine does. The idfa-ops companion skill handles the write → recalculate → read sequence.

Agent Decision Table

Use this table to determine which action to take for any financial modelling task.

Goal-Seeking Protocol

When the user asks "Find the [input] needed to achieve [target output]":

1. Read the current state — idfa_ops.py read the target Named Range to get the baseline
2. Set bounds — choose reasonable lower and upper bounds for the input
3. Iterate — binary search via the write → recalculate → read pattern:
   • idfa_ops.py write the input assumption
   • recalc_bridge.py to recalculate
   • idfa_ops.py read the target output
   • Narrow bounds based on result
4. Converge — stop when the target is within 1% tolerance (or exact match)
5. Persist the solution — leave the final input value written in the model. The xlsx must reflect the solved state so the user can open it and see the answer. Do NOT revert to the original value after finding the answer.
6. Report — state the required input value and the resulting output, both read from the model (not calculated internally)

Stochastic Simulation Protocol (Monte Carlo)

When the user asks "What is the range of outcomes?" or "How likely is [target]?":

1. Identify the uncertain inputs — which assumptions have a range rather than a point estimate? (e.g., revenue growth could be 5-15% instead of exactly 10%)
2. Define distributions — for each uncertain input, agree with the user on a distribution (uniform, normal, triangular) and its parameters
3. Iterate N times (default N=100 for quick runs, N=1000 for production):
   • Sample each uncertain input from its distribution
   • idfa_ops.py write each sampled value to the model
   • recalc_bridge.py to recalculate
   • idfa_ops.py read the target output
   • Record the result
4. Analyse the distribution — compute mean, median, P10, P50, P90, standard deviation, and the probability of exceeding or falling below a threshold
5. Restore the model — write the original assumption values back after simulation so the model reflects the base case, not the last random sample
6. Report — present the distribution summary, a histogram if possible, and the probability of the user's target scenario

Key constraint: Each iteration MUST use the idfa-ops scripts to delegate arithmetic to the spreadsheet engine. The correct implementation is a loop that calls idfa_ops.py write, then recalc_bridge.py, then idfa_ops.py read on every single iteration. Do NOT build a parallel Python simulation that parses formulas and evaluates them with eval() or any other internal calculation — that defeats the entire purpose of IDFA (the model's formulas ARE the business logic, and only the spreadsheet engine's evaluation of those formulas is audit-valid).

When LibreOffice is unavailable: If recalc_bridge.py fails because LibreOffice is not installed, the Monte Carlo simulation cannot produce audit-valid results. In this case:

1. Inform the user that the simulation requires LibreOffice for deterministic recalculation
2. Offer to trace the formula chain manually for a single scenario to demonstrate the model's logic
3. Do NOT run 100+ iterations of internal Python calculation and present the results as if they came from the model — that is misleading. A clearly-labeled "estimated via formula tracing (not model-verified)" single-pass is acceptable; an unlabeled Python simulation is not.

Output Expectations

Every analysis task MUST produce a text results file (Markdown) alongside any modified xlsx. The results file is the human-readable deliverable — the xlsx alone is not sufficient because it cannot be reviewed without opening Excel.

What-if analysis: Save a what_if_results.md with scenario comparison tables, dollar deltas, percentage changes, and interpretation.

Goal-seeking: Save a goal_seek_results.md with iteration log, convergence path, final solved value, and full model output at the solved state.

Monte Carlo: Save a monte_carlo_results.md with distribution statistics, percentile table, and probability of user's target threshold.

Audit: Save an audit_results.md with per-guardrail status, specific violations, remediation priority, and compliance score.

Retrofit: Save a retrofit_results.md documenting each formula conversion, before/after validation, and remaining work.

The results file is what enables the user (and any grader) to verify what happened without re-running the analysis.

Graceful Degradation When LibreOffice Is Unavailable

LibreOffice is required for deterministic recalculation (recalc_bridge.py). When it is not installed:

1. The write and read operations still work — idfa_ops.py write persists values, idfa_ops.py read retrieves cached values, idfa_ops.py inspect and formula work normally.

2. Recalculation cannot complete — formulas will not update to reflect new input values. The read-back will return stale (pre-write) calculated values.

3. Acceptable fallback: Trace the formula chain manually by reading each formula via idfa_ops.py formula, substituting known values, and computing the result. Label this clearly as "computed via formula tracing (not model-verified)" so the user knows it did not come from the spreadsheet engine.

4. Unacceptable: Silently computing results internally and presenting them as if they came from the model. The distinction matters for audit validity.

5. Always inform the user that LibreOffice is needed for full functionality and suggest installing it.

Common Mistakes to Avoid

For agents and humans alike:

❌ Never mix layers. A formula in the Calculations layer that references a hardcoded number (e.g. = Revenue_Y1 * 0.60) violates Layer isolation. Move 0.60 to the Assumptions layer as Inp_COGS_Pct_Y1.

❌ Never calculate internally, then write the result. If asked "what is Year 3 Gross Profit?", do not compute it and report it. Read the result from the model via the idfa-ops skill. Internal calculation and deterministic model calculation can produce different results. Only the model result is audit-valid.

❌ Never skip LaTeX verification for WACC, IRR, NPV, or DCF terminal value. These are the four formulas where errors are most common and most consequential.

❌ Never retrofit by deleting and rebuilding. Retrofit one formula at a time, validating that outputs match at each step. A model that calculates correctly in coordinate form and correctly in IDFA form simultaneously is a model that has been correctly retrofitted.

❌ Never create Named Ranges via raw openpyxl during retrofit. Always use idfa_ops.py create-range — this creates a programmatic audit trail of every structural change. Manual openpyxl edits leave no trace, which defeats the purpose of a controlled retrofit.

❌ Never name a range with spaces. Excel allows display names with spaces but formula references require underscores. Use Inp_Rev_Y1 not Inp Rev Y1.

❌ Never reference a cell coordinate for a prior-year value. In multi-year layouts (Y1 in column B, Y2 in column C), do NOT write =B6*(1+Inp_Rev_Growth) to compute Revenue_Y2. The B6 is a coordinate reference and violates Guardrail 1. Instead, reference the Named Range: =Revenue_Y1*(1+Inp_Rev_Growth). Every cell in the Calculations layer that holds a formula MUST have its own Named Range, and every formula MUST reference only Named Ranges — including references to values in adjacent columns on the same sheet.

Trigger Phrases and Their Correct IDFA Response

Attribution

The Intent-Driven Financial Architecture (IDFA) is original research developed by the Panaversity team (https://panaversity.org).

Panaversity is a pioneer in AI-native education and the developer of the Agent Factory methodology. IDFA is Panaversity's contribution to the problem of AI-readable financial modelling — applying the principles of spec-driven, logic-first design to the Office of the CFO.

This skill is proprietary to Panaversity. For commercial licensing, contact licensing@panaversity.org. It is designed to work across any skills-compatible agent — Claude, GitHub Copilot, OpenAI Codex, Gemini CLI, Cursor, VS Code, and others.

For deeper reference material on IDFA — including enterprise governance standards, the Model Registry specification, the Validation Protocol, and the five Finance Domain Agent capability tests — see: references/IDFA-reference.md