METHODOLOGY
How we score, grade, and compare tech reviews.
The BPR Formula
BPR — Battery Performance Ratio — is a single 0–100 score that summarizes how much of a laptop's plugged-in performance it keeps on battery. We compute it as the geometric mean of per-discipline battery/AC ratios across the seven BPR-eligible disciplines.
BPR = exp( (1/n) × Σ ln(battery_i / ac_i) ) × 100Why geomean, not arithmetic mean? A geometric mean treats a 50% drop in one discipline the same weight as a 200% gain in another — a single catastrophic regression can't be hidden by a strong average. It's the right tool when you're combining ratios.
The 7 Eligible Disciplines
Of the 13 disciplines in Rubric v1.1, seven are BPR-eligible. These are the workloads where battery mode materially changes performance — the ones where "unplugged penalty" matters to a real buyer. Disciplines like Storage and Wireless don't vary meaningfully on battery and are excluded to keep the score signal-heavy.
| Discipline | Description |
|---|---|
| CPU | Core compute performance across Geekbench 6, Cinebench 2024, and standard multi-thread tests. |
| GPU | Graphics compute across 3DMark, Blender GPU, and real-world render workloads. |
| LLM | Local large-language-model inference throughput (tokens/sec) on standardized models. |
| Video | Video encoding and decoding throughput — H.264, H.265, AV1. |
| Dev | Developer workload benchmarks — compile, link, test suite runtime. |
| Python | Python-specific numeric and scripting workloads. |
| Games | Real-world game performance — frame rates at standardized settings across a fixed title list. |
Medal Thresholds
| Medal | Score | What it means |
|---|---|---|
| PLATINUM | 90% and up | Near-zero battery penalty. |
| GOLD | 80%–89% | Minor battery impact — holds most performance unplugged. |
| SILVER | 70%–79% | Noticeable dropoff on battery — workable but visible. |
| BRONZE | 60%–69% | Significant battery drop — plan on plugging in for heavy work. |
| No medal | Below 60% or < 5 of 7 | See exclusion policy → |
GlitchMark
GlitchMark is a single number that summarizes a device's performance across every benchmark we record. One score per device. Higher is better; the reference device sits at 100.
Where BPR grades the qualitative value of a review (Platinum / Gold / Silver / Bronze on a 7-of-13 rubric), GlitchMark is the raw aggregate sortable number across the full benchmark set. Both surface side-by-side; they answer different questions.
Formula
For each benchmark a device runs, we compute a normalized ratio against a fixed reference value. Higher-is-better tests use raw / reference; lower-is-better tests use reference / raw. Reference device → ratio 1.0 → score 100.
GlitchMark is the geometric mean of all those ratios, multiplied by 100:
GlitchMark = (r₁ · r₂ · ... · rₙ)^(1/n) × 100
Worked example: a device with 8 measured tests where two tests are 2× and 8× the reference and the rest equal the reference (ratios = [2, 8, 1, 1, 1, 1, 1, 1]) yields a geometric mean of 16^(1/8) ≈ 1.41 — GlitchMark ≈ 141.
Test count policy
- Below 8 measured tests: no GlitchMark — the signal is too thin to publish.
- 8–11 measured tests: score published with a partial flag (e.g. GlitchMark 142 · partial (10/18 tests)).
- 12 or more measured tests: score published without a partial flag.
Reference baselines
Reference baselines are populated as benchmark tests are calibrated. None published yet.
Version history
v1 — initial release. Geometric mean × 100, per-test reference baselines as listed above, ≥8-test floor, partial flag for 8–11.
Exclusion Policy
A discipline may be excluded from a specific review when the hardware lacks the relevant component (no discrete GPU, no neural engine), the test harness isn't supported on that platform, or the reviewer documents a specific opt-out reason. Every exclusion is recorded per-review and shown on the review page.
When fewer than five of the seven BPR-eligible disciplines have complete AC + battery data after exclusions, we render no medal at all. A "Not enough data" placeholder appears in its place. Showing a misleading low-data score would do more harm than showing none.
Rubric Changelog
Get Tech Reviews in Your Inbox
Weekly reviews, benchmarks, and deep dives.