Everything published here rests on counted attempts and timed retries. This page defines those measurements precisely enough that a reader can disagree with our conclusions using our own figures, which is the point of publishing them.
One attempt is one try from a checkpoint or arena entrance to either death or success. Quitting to a menu mid-attempt counts as a death. Reloading a save to avoid a death counts as a death, because pretending otherwise would flatter our own competence and distort every figure downstream.
Runs in roguelikes are counted as attempts rather than the individual encounters inside them, since the checkpoint structure makes anything finer meaningless. Where a game is counted in runs rather than attempts, every table says so explicitly.
Each attempt is recorded with the best progress it reached, expressed as a percentage of the obstacle. For a boss with visible health that is straightforward. For a platforming room we divide the route into landmarks in advance and record the furthest landmark reached, which we fix before the first attempt so that later knowledge cannot influence the scale.
Everything is written down at the time, in a notebook, during the session. Reconstructing an attempt log afterwards produces a narrative rather than a record, and we know from comparing our early manual logs against in-game counters that memory reliably undercounts failures.
The retry loop is measured from the frame at which the death is registered to the first frame at which input affects the game again. We capture footage at sixty frames per second and count frames, then repeat the measurement five times and take the median.
Where the loop varies — a longer animation on some deaths, a variable load — we report the median and note the range. This is the single most predictive number we collect, and the reasoning behind that claim is set out in the Spire Of Glass review.
A death is fair if the information needed to avoid it was available before the death occurred. This is deliberately mechanical rather than a matter of taste: it can be checked against footage, and it produces a number rather than an adjective.
Every game is audited for unaccountable deaths, meaning deaths we could not explain from the log and the footage. Under one per cent is excellent, and anything above ten per cent effectively caps the mark, because a player cannot learn rules that do not hold consistently.
This is the distinction the journal exists to defend. An assist option changes how a player interacts with a game without changing what the game asks of them: input remapping, hold-to-toggle conversion, reduced motion, high-contrast hazards, interface scaling, screen reader support. A difficulty setting changes the challenge itself: damage multipliers, enemy counts, timing windows.
We list them separately in every review and we test each assist option individually across at least twenty attempts, to record what it actually changes rather than what its label implies. Filing a colour-blind mode alongside a damage multiplier has made both harder to discuss honestly, and it has given cover to games that provide neither.
One consequence worth stating: a game offering no difficulty settings is not penalised here. A game offering no assist options is, because there is no design argument for making a game unplayable by people who could otherwise play it.
One mark out of ten to one decimal place, describing whether a game's difficulty is fair and instructive rather than whether it is high. A brutally hard game with consistent rules and a fast retry loop will always score above a gentle one that wastes a player's time.
| Band | Meaning |
|---|---|
| 9.0–10 | Hard, fair, and teaching something on nearly every failure |
| 8.0–8.9 | Fair throughout with one badly tuned section or a slow retry loop |
| 7.0–7.9 | Difficulty works; the surrounding structure does not respect your time |
| 6.0–6.9 | Slow to recover from rather than genuinely difficult |
| 5.0–5.9 | Rules inconsistent enough that learning is unreliable |
| Below 5 | Failure is frequently unaccountable; the game cannot be learned |
Roughly forty per cent of a mark is fairness, measured as the proportion of unaccountable deaths. Thirty per cent is the retry loop and the structure around failure. Twenty per cent covers the assist options provided. The remaining ten per cent is everything a conventional review would call quality, which is small here because other publications already cover it well.
All reviews are conducted on the game's default difficulty, whatever the game calls it. Where a game has no default, we take the middle option and say so. We replay selected sections on other settings for comparison, but no mark is ever based on a non-default run.
Breaks are logged with their duration because our own data shows they are the strongest single influence on attempt counts. Deliberately stopping after a flat band and returning later recovered a median of twenty-six points of progress overnight, which is documented in the attempt curve.
Counting errors are corrected in place with a note, and the original figure is left visible. Marks change only when a patch alters something we measured — a retry loop, an assist option, an inconsistency we logged. Taste never revises a mark, and neither does the passage of time.
No advertising, no sponsorship, no affiliate arrangements and no commercial relationship with any publisher, platform or hardware manufacturer. Games are obtained at our own expense. There is no correspondence route on this site, because a journal built on counted data has nothing useful to gain from an inbox.
Story, art direction, music and performance are outside our scope. Other publications cover them thoroughly, and adding a paragraph of impressions to a page of attempt logs would only dilute the one thing this journal does.
