Trust & transparency
How UpgradeOrSave Decides
Last updated October 13, 2026
UpgradeOrSave answers one question: is upgrading from the hardware you own to the hardware you are considering worth it for the work you actually do? The answer comes from verified evidence and fixed, versioned rules. AI models may help explain a result, but they never make the call.
The decision path
- Your PC — the hardware you own; for a graphics card upgrade, also power supply, connectors, and case clearance if you provide them.
- Compatibility — whether the candidate fits what you told us. Published requirements alone never prove fit.
- Performance — whether the candidate is faster in the workload you selected.
- Eligible current price — only prices whose source permits public purchase advice.
- Resale and net cost — what your current card may be worth and the resulting net cost.
- Timing — whether the market supports waiting.
Four possible calls
- UPGRADE — The measured workload result, compatibility, and price evidence meet the current decision checks. This does not predict FPS or guarantee purchase value.
- WAIT — The current checks do not support buying yet. Review the budget, price, and evidence details below to see what is holding this decision.
- KEEP CURRENT — Keep your current setup for this workload: the candidate does not show a measured improvement. This says nothing about other workloads.
- NEED MORE DATA — Performance, compatibility, or current-price evidence is missing. Unknown is not a zero or a confirmed fit, so no upgrade call is made.
A faster card is not automatically a purchase recommendation. Performance direction and the purchase call are always shown separately. SAVE can be decided on performance alone, when every decisive test shows the candidate is not faster. UPGRADE and WAIT require price evidence we are allowed to use publicly.
No combined score
Gaming at each resolution and ray tracing mode, rendering, local AI, compatibility, price, and resale are shown as separate dimensions. We do not blend them into a single number, and results from different sources are never averaged together.
Where performance scores come from
Every performance score on rankings, hardware pages, comparisons, and upgrade checks is a UOS workload index read from one place. Indices are derived from measured results published by independent sources, connected through comparable test cohorts to a frozen reference for each workload domain. A product gets a score only when its result is ready at the current reference for that domain and is cleared for publication: every source behind it must have confirmed rights for public display and derived scores, and none of its evidence may still be under review. A result that is not cleared is shown as "Evidence under review" or "Not yet approved for publication" — never as a score, and never as zero. Older source composites are never used as a fallback, and specifications or estimates never enter an index.
Research that has not passed admission — candidate documents, extractions awaiting review, and user-run diagnostics — is not an index result. The UOS lab benchmark suites described on the UOS Benchmark methodology page have no measured reference sets yet and do not produce public scores.
Evidence strength
Upgrade checks compare two products only when both have a ready UOS index in the same workload domain. The evidence explorer also lists older source records, labeled by the strength of the evidence behind them:
- UOS index — Both products have a ready UOS workload index on the same frozen reference, methodology and graph semantics.
- Measured — Measured on the exact workload (model, settings, and runtime) being compared.
- Standardized test — The same benchmark suite and version run on each GPU.
- Public report — An attributed publication that tested the GPUs in one test setup.
- Limited evidence — Sparse, mixed, or conflicting results — treat the direction with caution.
- Not measured — No valid comparable evidence. Nothing is estimated in its place.
UOS index results record how many measurements and independent source cohorts contribute. We do not publish confidence percentages; published results often omit sample spread, and we say so.
Missing evidence is not zero
When evidence is missing we show that directly — for example "Not measured", "Needs your PC details", or "Current pricing unavailable" — and the purchase call becomes NEED MORE DATA. We never fill a gap with an estimate, a typical value, or a result from a different GPU or workload.
Freshness and provenance
Publication, retrieval, and benchmark measurement dates are different facts. UOS preserves source attribution and review status in its evidence layer, but public ranking tables do not yet display per-measurement timestamps. A ranking page loading today does not mean its measurements were collected today. Where an exact source observation date is not exposed, freshness is unknown — not current by default.
Dataset coverage is scoped to a specific hardware population and comparable workload. A configuration can be indexed in rendering and unindexed in gaming. Unindexed never means a zero score. Reviewed checkpoint-file bytes are not evidence that a Local AI model runs on an exact GPU or backend.
Prices and rights
A current retailer price appears only when the exact source permits public display and the observation meets its freshness and attribution requirements. Using price in a purchase recommendation is a separate permission check. Asking prices are what sellers list, not completed sales. Price never changes a performance result.
Commissions do not influence results
Whether a retailer link earns a commission has no effect on evidence, rules, or calls. See our Affiliate Disclosure.
Known limitations and corrections
Coverage is intentionally narrow: rankings list only reviewed hardware identities, and upgrade checks are offered only for curated pairs that share a measured workload domain. During the beta, individual upgrade check pages are not included in our sitemap. If you believe a result or source is wrong, contact us. Every page is computed from the same underlying evidence, so a correction there updates every page that uses it.