How SENTINEL scores Washington properties

The property-priority score is a 0-100 number that combines four weighted signals: equity ratio, ownership tenure, absentee status, and appreciation gap. Each factor is worth 25 points. We compute the score from public county-assessor records — no machine-learning model, no neural network, no proprietary black box.

The score is not a prediction. It is a transparent ordering of the four inputs above. A higher number means more of those rules fired; it does not mean Sentinel has measured a sale probability or future-listing window. We have not completed a prospective validation cohort large enough to make that claim.

Everything is open. The exact code path is at src/scoring/engine.py. The county data is public. The Redfin market data is licensed for republication with attribution. The methodology is versioned — when the algorithm changes, the prior version stays live forever at /methodology/v0-0-1 so any agent's screenshot URL still resolves with the math the score was actually computed under.

SENTINEL ingests from a small set of public and licensed sources. Every datapoint surfaced in the product traces to one of these.

The score combines four equally weighted factors, each worth a maximum of 25 points. Total range: 0-100. The algorithm is ~150 lines of Python with no dependencies beyond the standard library.

Factor 1 — Equity ratio (0-25 pts)

We estimate the share of the property's assessed value that is owner equity rather than mortgage debt. Higher equity = more options to sell.

equity_ratio = (assessed_value − remaining_mortgage) / assessed_value

Remaining mortgage is estimated from the last recorded sale price under three assumptions: the buyer financed 80% of the purchase (LTV = 0.80), the loan amortizes linearly over 30 years, and the property has not been refinanced. These assumptions are wrong on individual properties — they are calibrated to be approximately right at the population scale. See Section 8 for what this misses.

The 0-25 score follows a piecewise scale: ≤20% equity ranges 0-5 pts (low-equity owners rarely sell discretionarily), 20-50% ranges 5-15, 50-80% ranges 15-22, and ≥80% caps at 25.

Factor 2 — Ownership tenure (0-25 pts)

Years since the last recorded sale. The model uses a published, non-linear tenure curve as a workflow-prioritization heuristic. Tenure does not reveal why an owner may or may not move, and the score does not infer household circumstances or life events.

• 0-3 years: 0-5 pts (recent buyers rarely sell)
• 3-7 years: 5-15 pts (early discretionary window)
• 7-15 years: 15-25 pts (peak)
• 15-25 years: 20-25 pts (still active)
• 25+ years: declines toward 15 pts under the published tenure curve

Factor 3 — Absentee status (0 or 25 pts)

We compare the property address to the owner's mailing address on the assessor record. If the first token differs (street number, ZIP, or city), we flag the property as absentee-owned. Absentee owners convert to sellers at materially higher rates than owner-occupiers.

This is binary: 25 points if absentee, 0 if not. We do not attempt to infer LLC ownership or trust ownership from the mailing record — too noisy.

Factor 4 — Appreciation gap (0-25 pts)

The ratio of current assessed value to the last sale price. A property that sold for $300K and is now assessed at $800K (ratio = 2.67) is likely owned by someone sitting on a large tax-deferred capital gain.

appreciation_gap = assessed_value / last_sale_price

Scoring: <1.0 ranges 0-3 pts (the property has lost paper value), 1.0-1.5 ranges 3-10, 1.5-2.5 ranges 10-20, and ≥2.5 caps at 25.

Combining the factors

The four factors sum to a number between 0 and 100. When the assessor record is missing data (no last-sale-price, missing mailing address), the score is normalized over computable factors only and tagged as limited data so an agent knows the signal-to-noise is lower for that parcel.

A real (anonymized) parcel from King County, scored on 2026-04-15:

What an agent does with a 97 is their judgment call. The score says only that this example triggered strong equity, tenure, absentee-owner, and appreciation-gap rules. It is a reproducible research queue, not evidence that the household intends to sell and not a forecast of when a transaction will occur.

We currently do not publish a confidence interval on the property-priority score, and we are not going to fake one. A calibrated forecast requires a prospective cohort: properties scored at time t, observed through a declared window, with listing and sale outcomes measured consistently.

Sentinel does not yet have a cohort large enough to publish calibration, precision, recall, or lift honestly. When that threshold is met, we will publish the cohort definition, sample sizes, observation window, exclusions, and results here.

Until then, read the number as a rules-based priority score. A 92 triggered more or stronger published rules than a 64. The numerical gap is not calibrated to an outcome.

Status as of 2026-07-09: not yet published. Sentinel is opening to its first Washington agents and does not have a prospective validation cohort large enough to publish outcome claims honestly.

Once the cohort is large enough, this section will report calibration, lift over the base rate, precision and recall at declared thresholds, sample sizes, and observation windows by geography and property type. The methodology will be published before the results are calculated.

When a large enough cohort exists, Sentinel will publish the observation window, geography, sample size, thresholds, and underlying validation method together. Until then, the score is a transparent lead-priority heuristic, not a probability of sale.

The methodology page is versioned alongside the algorithm. When the algorithm changes, the prior version stays live forever. Any score generated under a prior version still resolves at its original URL with its original math.

• We are not an appraisal. The score is not a valuation and cannot be used as one for lending, tax, or insurance purposes.
• We do not provide Fair Housing guidance beyond listing the protected classes. We do not surface or predict demographic information of any kind. Agents asking us about "best neighborhoods for X" are routed back to neutral market math.
• We are not a substitute for broker judgment. Every legal, contract, deadline, or compliance topic ends with "verify with your broker." This is not a hedge — it is the correct operating posture.
• We do not display NWMLS listing data publicly. SENTINEL is not a Vendor Affiliate. Aggregated market stats from Redfin Data Center (republished with attribution) are fine; individual listings are not.
• We do not buy or trade backlinks, SEO tokens, or ranking placements. Every page on this site that ranks is a page we wrote.

Following the Have-I-Been-Pwned pattern: the model's failure modes are public.

• Refinanced mortgages bias the equity estimate. We assume original-purchase financing. A 2009 buyer who refinanced in 2021 looks lower-equity than they are. Roughly ~20-30% of long-tenure WA homeowners refinanced during the 2020-2022 rate window — we do not yet have a correction.
• LLC and trust ownership obscures absentee detection. When a property is held by an LLC with a registered-agent mailing address, the "absentee" flag fires even if the human owner lives at the property.
• Assessed values trail market by ~12 months. King County uses a 2025 valuation date for 2026 tax rolls. Hot or cold markets in 2026 are not yet reflected.
• Court and life-event records are excluded. They do not contribute to the published score or Radar ordering. Agents should treat tenure and equity as property-research context, never as evidence about a household or a reason for outreach.
• No multifamily handling yet. The model is calibrated for single-family + condo. Triplexes and quadplexes get scored but the absentee logic is unreliable.

If you find another failure mode, write us. The contact path is in your dashboard. We publish errata in the model-version changelog.

How to cite SENTINEL in a CMA

Suggested citation format:

Source: SENTINEL property-priority methodology v0.0.1.
Internal score generated YYYY-MM-DD. Include the county-assessor source and fetched-at date shown in the result.
Methodology: sentineliq.net/methodology.

RCW citations referenced in product

• RCW 42.56 — Public Records Act (county assessor records)
• RCW 64.06 — Real estate disclosure (Form 17)
• RCW 82.45 — Real Estate Excise Tax (REET, graduated brackets)
• RCW 18.85 — Real estate brokers and managing brokers
• RCW 64.04 — Statute of frauds (real estate contracts)

Verify any RCW citation against app.leg.wa.gov/rcw for current text. Statutes change.

License of this methodology page

This methodology document is published under CC BY 4.0. You may copy, adapt, and republish with attribution to "SENTINEL, /methodology v0.0.1". Code at src/scoring/engine.py is proprietary; the algorithm described above is published, the implementation is not.