Methodology

What the language model does and does not do

The only place natural language enters the decision path is turning your description into a structured requirement profile. In this MVP that parsing is done by a deterministic local interpreter with no API key and no network call. An optional model adapter may be added later, but it would have to return the same validated schema, and it would still never select parts, read prices, or judge compatibility.

How three builds are produced

1. Candidate retrieval: every published product with at least one in-stock offer that satisfies your hard requirements (brand exclusions, locked parts, style preference).
2. Top-k reduction per category, so the search stays bounded and reproducible.
3. Bounded backtracking over the eight categories, pruning any branch whose cheapest possible completion already exceeds the budget, and any partial selection that already contains a compatibility error.
4. Price and estimated power computed for every complete candidate.
5. Builds over budget are discarded, not down-ranked.
6. Every survivor is scored once per profile using the weights below.
7. The three published builds are the best-scoring per profile, preferring builds that differ from the others in at least two components — but never sacrificing more than 4.5 score points for that difference.

The same requirements against the same catalog version always produce the same three builds. Ties are broken by total price and then by a stable identifier.

How Upgrade Advisor works

Version 2026.08.3 analyzes a complete current build and an explicit maximum upgrade spend. It substitutes every purchasable same-category product one at a time and separately enumerates genuine CPU, motherboard and memory migrations. Every path recomputes the complete build rather than scoring products in isolation.

1. Discard a replacement if any compatibility rule returns an error.
2. Discard it if the full delivered checkout price exceeds the cash budget.
3. Compare category-relevant specifications and, where applicable, selected-game frame rate, whole-build score and power draw before and after the swap.
4. Apply diminishing returns: capacity already adequate for the target is worth less than fixing a real shortage.
5. Rank by measurable impact, current resolution and stated priority, then show at most one recommendation per category.
6. For a platform migration, replace CPU, motherboard and memory atomically, require a socket or memory-generation change, combine all three delivered prices and apply an explicit installation-complexity penalty.
7. Estimate recovery only from fresh completed sales for the exact product and selected condition. Require at least four observations across two marketplaces after fees and seller-paid shipping.
8. Keep the gross checkout budget gate. When every removed part has verified coverage, rank against the conservative end of the net-cost range; otherwise use gross cost.

Affiliate status and commission are not read by this calculation. Asking prices are excluded from resale evidence, partial bundle coverage is never deducted, and an expected recovery is shown separately from the checkout price rather than presented as guaranteed cash.

Scoring components

Power model

Estimated system load is the sum of the following terms. Every coefficient is fixed and published so you can reproduce the number yourself (model version 2026.08.1):

• Processor peak package power × 1.25 to cover short boost windows
• Graphics-card board power × 1.1 to cover transient spikes
• Motherboard and chipset: 35 W
• Memory: 3 W per module
• Storage: the drive’s published typical draw
• Cooler: 5 W for an air tower, 12 W for a pump and radiator fans
• Chassis fans: 2.2 W each
• Headers, USB and unmodelled loads: 10 W

Recommended supply size is the estimated load plus the profile headroom, rounded up to the next 50 W.

Compatibility rule set

Version 2026.08.1, 13 rules. Each rule is a pure function of the selected parts, returns a severity, and records the canonical values it read.

Severities

• Error — the combination cannot work. The builder never returns one.
• Action required — it works after a specific step, such as a BIOS update, an adapter or a mounting bracket.
• Limitation — it works, but below the rated capability of a part.
• Note — information worth reading before ordering.
• Verified — the rule was evaluated against real values and passed.

Benchmarks: measured, normalized and estimated

The catalog currently holds 110 benchmark observations across 5 demo titles.

• Direct — measured on the reference bench under a published methodology.
• Normalized — measured elsewhere and rescaled onto that methodology, carrying a confidence value.
• Estimated — not measured at all. For a given game, resolution and preset we fit a line through the origin over the measured cards, relating the performance index of a card to its frame rate, then read the unmeasured card off that line. Confidence falls when the card sits outside the measured range, and the figure is capped by the processor frame-rate ceiling.

The processor ceiling model treats gaming index 100 as roughly 240 fps and scales linearly below that. An estimate is never displayed with the same badge as a measurement.

Offer ranking

Offers are sorted by, in order:

1. availability — in stock, then low stock, then out of stock;
2. freshness — offers observed within the last 72 hours rank ahead of older ones;
3. total price including shipping;
4. retailer slug and offer id, purely to make the order deterministic.

Affiliate status and commission are not inputs. A retailer with no affiliate agreement can and does win the ranking in this demo dataset. See the affiliate disclosure.

Retailer rows reach this ranking only after deterministic identity matching, price anomaly guards and a feed-level rollback gate. Read the complete Merchant & Offer Autopilot methodology.

Data quality and publication

Product data is assembled from source claims, not written by hand. Each source has a trust score, and when two sources disagree on a field, the higher-trust source wins and the disagreement is recorded and shown on the product page. A product is published only when every critical field for its category is present and the whole specification passes its schema contract. Anything else is quarantined rather than published with risky data — coverage is sacrificed before trust is.

Demo dataset

Every brand, product, game, retailer and price in this MVP is fictional and created for demonstration. Specifications are internally consistent so the engines above can be exercised end to end, but they do not describe real hardware, and no price shown here is a live quote.