Methodology · v0.4-unified

How the Chalais Score works.

The Chalais Score is a 0–100 grade of how much value is hiding in a residential property — and how much of it the right renovation can unlock. It blends eight dimensions, real comparable sales, and a property-specific ceiling estimate into one number that maps to a concrete next move. This page describes how that number is constructed.

The eight dimensions.

Every Chalais Score is a weighted sum of eight 0–10 dimensions. Each dimension answers a separate question about your home. The questions that best predict what renovated homes actually sell for count the most. Weights sum to 1.0.

• Finish quality

  What quality are the finishes today? Stone vs laminate, built-in vs bolt-on, paneling vs paint. Better bones mean the renovation can reach higher.

  20%

• Resale headroom

  How much of the sub-market's price ceiling does this property currently capture? Lower = more headroom hiding inside the address.

  20%

• Smart use of space

  Are the existing walls in the right places? Bonus rooms, wasted halls, awkward kitchens, dead corners — features the audit flags for removal or repurposing.

  10%

• Style fit

  Does the property read as a single confident style, or is it three remodels in a trench coat? Coherent properties realize closer to ceiling per dollar spent.

  10%

• Light & views

  Natural light, sightlines, and what the windows actually frame. Hard to fix, expensive to retrofit, easy to undervalue.

  10%

• Bones & systems

  Roof age, HVAC, plumbing, electrical. The unsexy stuff that drives 30% of post-sale buyer regret if it's old.

  10%

• Right-sized

  Right-sized for the sub-market. A 2BR in a 4BR neighborhood under-captures; a 6BR in a 3BR neighborhood over-builds. The Score penalizes both.

  10%

• Overbuild margin

  Headroom for the renovation itself to go over budget or under-deliver and still beat the buy price. The Score rewards properties with margin to absorb execution risk.

  10%

Two anchor dimensions — finish quality and resale headroom — each carry 20% of the composite weight. The remaining six split the rest evenly at 10% each. Anchors carry more weight because they explain the largest share of cross-property variance in our reference dataset.

The ceiling.

The Score doesn't exist in a vacuum. Every property gets a ceiling estimate — the top of what this property could realistically sell for, post-renovation, given the sub-market. The ceiling is what makes the headroom number ($X of upside left on the table) defensible.

We build the ceiling three ways and pick the one most appropriate for the data we have:

1. Top-quartile comp. When we have enough recent, similar sales within 1mi, we anchor the ceiling to the 75th-percentile $/sqft. Highest confidence.
2. Thin-cluster median. When comps are scarce — rural, unique, or freshly post-COVID markets — we fall back to a tighter cluster median, with a wider confidence band.
3. LLM counterfactual. When neither comp method fits, we ask a frontier vision model to estimate the ceiling given the property and the sub-market. Lowest confidence — flagged in the UI when used.

Data sources.

Every audit pulls from named, third-party data providers. Sources are credited inline wherever their data is shown.

RentCast· Comparable recent sales

Used for the Verified Comps section and the top-quartile ceiling. Pulled live per audit. Returns standardized fields (price, sqft, beds, baths, sold-date, distance) for sales within 1mi of the subject property.

Google Maps Platform· Geocoding + Street View imagery

When a read is started by address (no upload, no listing URL), we geocode the address and pull a Street View Static photo as the property hero. The image stays attributed to Google.

Anthropic Claude (Sonnet)· Visual audit reasoning

The audit body — what's working, what's not, the three highest-ROI moves — is generated by Claude with vision. We pass the property photo, listing facts (sqft / beds / baths / year built), and the RentCast comp set; Claude returns a structured audit report which we validate against a Zod schema before storing.

Google Gemini (with grounding)· Listing-URL hunt

When you submit an address, we use Gemini's web grounding to find the listing on Zillow / Redfin / Realtor and pull the canonical photo and listing facts. Reduces the human-in-the-loop step for address-only entries.

OpenAI GPT (with vision)· Render brief refinement + budget estimation

Used inside the per-render Build Sheet to extract ZIP-grounded line items from the rendered before/after pair. Plus plan and above.

What we don't do.

• We're not an appraiser. Chalais output is for decision-making, not lending or tax basis. A licensed appraiser is still the right call for those uses.
• The render is not a construction document. It's an AI visualization of one possible execution of the recommended scope. A licensed contractor — and, for structural changes, a licensed architect — is what bridges from render to building permit.
• We don't guarantee a sale price. Every projected ARV is labeled as a model output, with confidence. Market conditions, execution quality, and timing all move the actual realized value.

More detail in our disclaimers.