Material Change Intelligence
Don’t just see that a number changed — see why. For a key output that moved between two versions, Material Change Intelligence ranks the specific edits that drove the move, shows the precedent path from each edit to the output, rates how strong that evidence is, and reports how much of the move it couldn’t account for. It is deterministic analysis, not a guess and not a black box.
What it is
A comparison tells you what changed. Material Change Intelligence answers the next question: of everything that changed, what actually moved this output, and by how much? You select an output cell — an IRR, an NPV, an EBITDA line, a covenant ratio — and it produces a deterministic explanation of the delta between the two versions: a ranked list of the changes that drove it, each with its estimated contribution.
The engine’s internal name is Output Delta Attribution. It is not machine learning and not generative AI — it is a deterministic walk of the model’s dependencies, refined with sensitivity analysis.
Where to find it
Open a comparison (most naturally Compare to Previous Version, whose ribbon entry names Material Change Intelligence), then open the output explanation view from the results window. Outputs can be auto-suggested — the engine scans labels for known financial metrics (IRR, NPV, DSCR, EBITDA, WACC, ROI, revenue, margin, cash flow, FCF, DCF…) — or you can select any cell as the output yourself.
What an explanation contains
For the output you pick, the explanation reports:
- The delta — old value, new value, the absolute change, and the percentage change.
- Top drivers — the changes that moved the output, ranked by contribution (up to a configurable number, 8 by default). Each driver shows:
- the changed cell and what changed (value / formula / structural / named-range);
- its estimated contribution and its share of the total delta;
- a reachability path — the chain of precedents proving the changed cell actually feeds the output (not a coincidental change elsewhere);
- an evidence status (see below) and whether it is material.
- Residual — the part of the delta the ranked drivers do not account for (interactions, unmodelled effects, rounding).
- Confidence — an overall score reflecting how much of the move is path- and sensitivity-backed versus approximate or residual.
Evidence status
Each driver is labelled by how strong the evidence behind its contribution is:
| Status | Meaning |
|---|---|
| Path confirmed | A dependency path links the changed cell to the output. |
| Sensitivity backed | The contribution was refined by perturbing the input and measuring the output’s response. |
| Approximate | Estimated, with weaker direct evidence. |
| Residual | Part of the unexplained remainder rather than an attributed driver. |
| Non-material | Below the materiality bar for this output. |
How it works
The attribution is deterministic. It builds the dependency topology around the selected output, finds the changed cells that can reach it through that topology, and estimates each one’s contribution to the delta. It then refines those estimates with the Sensitivity engine — the same Sensitivity analysis used in Formula Tools — perturbing inputs to measure how the output actually responds, within a small refinement time budget. Drivers that survive this become sensitivity backed; whatever the ranked drivers don’t explain becomes the residual.
Exporting an explanation
From the explanation view you can export the report as an .xlsx workbook or a .pdf (the PDF is produced by exporting the generated report sheet to fixed format; a temporary .xlsx is written and then deleted). You can also open the explanation in Excel, and the view can highlight a driver’s cells in the grid — those highlights use a save-and-restore pattern so your original colors come back when you’re done. This .xlsx/.pdf explanation report is distinct from the main results export (.xlsx) and save (.mxcmp).
Link to Proof & Trust
Material Change Intelligence is also wired into the Proof & Trust ledger. When a comparison’s formula change falls inside a signed block, a block-broken event is appended to that workbook’s causality ledger, back-referenced to the comparison — so a cross-version change that invalidates a signature is recorded, not silently lost. This is a no-op when there are no signatures.
An attribution is an explanation of movement, not a proof of correctness
The contributions are deterministic estimates, the residual is the part they couldn’t explain, and the confidence reflects the strength of the evidence — none of it asserts that a formula is right. Use it to understand and explain a change; use Why This Number and Proof & Trust to verify the number itself.
Example: explaining an EBITDA move
Compare to last week’s file shows EBITDA up 6%. Open Material Change Intelligence on the EBITDA cell: it ranks the drivers — a revenue-assumption change contributing the bulk of the move (sensitivity backed, with a precedent path through the revenue build), a smaller cost-formula edit, and a residual of a couple of points from interactions. Export the .pdf and attach it to the model’s change note.
Caveats
- Deterministic estimates, not certainty. Contributions and confidence are computed signals; the residual is honest about what isn’t explained.
- Best on numeric outputs with a real dependency path. An output that didn’t move, or a change that can’t reach it, won’t produce drivers.
- Refinement is time-budgeted. Sensitivity refinement runs within a short budget (5 seconds by default; adjustable in settings), so very large models may lean more on path-based estimates.
- Not AI. There is no model-generated narrative or learned prediction here — only dependency analysis and sensitivity.
Related
- Sensitivity — the engine that backs driver contributions.
- Why This Number & Proof & Trust — verify the number and see the block-broken ledger events.
- Compare to Previous Version — the natural entry point.
- Comparison settings — attribution drivers, budget, and output selection mode.