Method
Everything the engine does, in the open
A verdict you cannot check is just an opinion. This page documents every rule the engine applies. The same inputs always produce the same output, and no language model is involved in reaching the conclusion — text extraction, where used, only pulls a published number across and always shows it next to its source.
1. Global evidence collection
Every factual finding is a published figure attributed to the publisher that produced it: expert reviews, laboratory measurements, benchmark database entries, professional evaluations and manufacturer commitments. The publisher registry currently holds 40 independent publishers across 16 countries and is designed to scale to hundreds, with a per-publisher collector that discovers documents and extracts figures — never estimates them. The corpus behind the live smartphone category is 1746 documents and 6969 findings across 202 models.
Review scores are never averaged into a rating. For each metric the engine takes the median of the independent published figures and reports the spread between them: agreement is strong under 5%, moderate under 15%, and anything wider is flagged as contested so you can see that the testers do not agree.
Why the median, not an average
Five illustrative publisher figures, one of them a high outlier. The average drifts toward it; the median doesn't.
Lab measurement or public benchmark database
Independent editorial testing with a documented procedure
Manufacturer specification, used only for stated commitments
Live collection is being rolled out publisher by publisher. Until a publisher is live, its findings are marked "seeded corpus" in the analysis so a demonstration figure is never presented as a verified one.
2. Metrics and base weights
| Metric | Group | Better | Weight | Threshold |
|---|---|---|---|---|
| Active use runtime | Battery & charging | higher | 1.00 | 8% |
| 0–100% charge time | Battery & charging | lower | 0.55 | 12% |
| CPU multi-core | Performance | higher | 0.85 | 12% |
| Sustained GPU | Performance | higher | 1.00 | 12% |
| Stills quality | Camera | higher | 1.00 | 4% |
| Zoom quality | Camera | higher | 0.70 | 6% |
| Video quality | Camera | higher | 0.80 | 4% |
| Peak brightness | Display | higher | 1.00 | 15% |
| Refresh rate | Display | higher | 0.60 | 10% |
| Remaining support | Longevity | higher | 1.00 | 12% |
| Weight | Handling | lower | 1.00 | 6% |
A difference smaller than the threshold is reported as "no real difference" and contributes exactly zero points, however good it looks on a spec sheet.
3. Your priorities become weights
Final weight = base metric weight × priority multiplier. All weights are then normalised so the score always lands on the same −100…+100 scale.
4. Your phone is scored as it is today
- Battery: measured runtime × your battery health percentage.
- Performance: benchmark figures reduced by a condition allowance — flawless 0%, good 2%, worn 5%, damaged 12% — covering thermal and throttling behaviour of a used device.
- Longevity: only the support years still remaining count, for both devices.
- Resale: launch price depreciated 24% per year, then adjusted for condition, battery health and storage.
Battery health
82%Example: 82% health counts as 82% of the measured runtime figure.
Condition allowance (performance metrics)
−0%
Flawless
−2%
Good
−5%
Worn
−12%
Damaged
5. The Upgrade Value Score and the verdict
Each metric's relative change is capped at ±60% — past that point more is not meaningfully better — then multiplied by its normalised weight and scaled to points. The Upgrade Value Score is the sum.
The break-even score is net cost ÷ 14, so a more expensive upgrade has to prove more. Verdict bands:
- score ≤ 8, or below 70% of break-even → Skip
- 70% of break-even up to break-even → Not worth this price
- at or above break-even → Worth it
- ≥ 1.6× break-even and ≥ 35 points → Clear upgrade
≤ 8 pts, or < 70% of break-even
70% of break-even up to break-even
at or above break-even
≥ 1.6× break-even, ≥ 35 pts
Illustrative — the exact break-even point depends on your own net cost, not a fixed score.
6. Limits we state openly
The engine only judges what is measurable and published. Preferences like design, ecosystem lock-in or brand loyalty are not scored. Where a model lacks a published figure for a metric, that metric is marked unknown and contributes nothing rather than being guessed.
7. Product photography
Where a device photo appears next to a comparison, it is a real, freely-licensed photograph, not an AI-generated or stock image. Every photo we use is credited here in full, as its licence requires.
- Apple iPhone 13 — Kskhh, CC BY-SA 4.0, via Wikimedia Commons (CC BY-SA 4.0)
- Samsung Galaxy S25 Ultra — Jakub CA, CC BY 4.0, via Wikimedia Commons (CC BY 4.0)
- Xiaomi 15 Ultra — 毕明明, CC BY-SA 4.0, via Wikimedia Commons (CC BY-SA 4.0)
- Apple iPhone 17 Pro Max — 茅野ふたば, CC BY-SA 4.0, via Wikimedia Commons (CC BY-SA 4.0)