Worked example: a Speedmaster Professional 3861 listing

This page is the answer to the question "what does Grail Atlas actually do when you paste in a listing?" — walked through end-to-end against a single, deliberately ordinary Speedmaster Professional 3861 listing. Every number you see comes from the engines you can read in the open source under src/engines/. Nothing here is hidden behind a model.

The listing

Picture an eBay listing — a 2023 Moonwatch on a steel bracelet, full-set with box, papers, and a service card from the seller's local Omega boutique dated nine months ago. The seller is a long-standing private seller with 187 sales over eight years, 100% positive feedback, identity verified through eBay. Asking price $6,200 USD, shipped from Switzerland to a United States buyer, with the seller absorbing duty (DDP).

Step 1 — landed cost

The first thing the calculator does is translate the asking price into a landed cost: what it actually costs to put the watch on your wrist after duty, brokerage, sales tax, and any third-party verification fee. The duty model is the engine you can read in src/engines/landed-cost/.

• Asking: $6,200 USD
• Duty: DDP — seller absorbs the import duty + brokerage on the buyer's behalf, so the listed price is the final number.
• State sales tax on the import: variable; we use the federal-only baseline of 0 in the prototype model.
• Verification fee: $0 (no third-party verification on this listing).
• Landed cost: $6,200 USD.

The model is intentionally simple in v1. Per-state sales tax, full broker-fee disaggregation, and currency forward exposure are all on the slice-1b roadmap; the prototype duty model treats DDP correctly but does not yet model the buyer-side use-tax that some US states levy post-import. The Calculator surfaces a one-line disclaimer on every result.

Step 2 — the band

Next, the valuation engine builds a fair-value band from comparable verified sales. For the Speedmaster 3861 with 24 comps in the catalog (mostly Watchfinder, Hodinkee Shop, and Wempe sold-listings from the last twelve months), the engine reports:

The band is brand-tier and era-aware. Per the sixth-panel watch-domain reset (documented in src/engines/valuation/brand-classifier.ts), Omega is in the luxury tier; the Moonwatch is modern (2021–present). Both move the condition + box/papers factor lookup; both move the momentum read.

Step 3 — deal tier

Landed cost ($6,200) sits below the band floor ($6,400). The engine reports great deal. The percentage-below-median is roughly 12.7%, which crosses the engine's great threshold.

Step 4 — trust read

Three layers contribute to the trust composite:

• Platform — eBay (public-listing tier; has authentication available but not used on this listing; has buyer protection and dispute resolution).
• Listing — full reference number on the dial visible in photos, no stock photos detected, serial format consistent with Omega's modern numbering, not on a stolen-watch registry. The listing-trust layer scores this listing at 74 on a 0–100 scale.
• Seller — private seller, 8 years on platform, 187 sales, 100% positive feedback, identity-cross-checked. The seller layer scores at 82.

The composite — weighted average of the three — comes back at 78 (trusted).

Step 5 — risk read

Risk is graded, not labeled. The engine never accuses; it publishes factual non-accusatory observations. For this listing, zero high-weight signals fire:

• No off-platform payment request.
• Not on a stolen-watch registry.
• No stock photos detected.
• Seller is established (not 12-day-old account).

Risk level: clear. (Possible levels: clear / caution / elevated / high.)

Step 6 — the verdict line

The Calculator's verdict line — added per the seventh-panel Product/Growth review — translates the engine output into a Buy / Wait / Pass recommendation plus one reasoning sentence. For this listing:

Buy — price is well below the typical band. Landed cost sits well below the excellent / full-set band and the trust + risk read is clear. This is the shape of a real deal.

What the engine did NOT do

• Read photographs of the watch. No image model is wired in v1. The Inspection Map structural data is reputation-gated — we publish what's checkable from listing text, not inferred from images.
• Authenticate the watch. We score the listing's trust posture; we do not opine on whether the specific watch is counterfeit. Third-party authentication (which the Omega boutique service card partially substitutes for) is a separate question.
• Negotiate. The verdict is "this looks like a real deal at asking price." Whether you can do better by waiting, asking the seller to drop, or finding a different listing is your decision.
• Promote the listing. Grail Atlas has no affiliate relationship with eBay or with any seller; the engine's output for a sponsored listing would be identical to its output for any other. See the disclosure page for the architectural guarantees.

How to read this page

Every step above corresponds to engine code you can read in the open source. Every number is reproducible. If the engine is wrong on a specific signal — a comp that shouldn't be in the panel, a score that doesn't match the dialog the evidence supports — that's a bug in code, not a black-box verdict from a model. The score-audit-trail engine logs the full input set for every assessment so the path from observation to recommendation stays inspectable.

This is the work we're trying to do: make a buying decision more legible without making it less honest. The catalog will grow, the comp panels will deepen, and the duty model will get more accurate. The shape of how an assessment reads — landed cost → band → tier → trust → risk → verdict — stays the same.

Open the Deal Calculator → · How the scoring works · Principles