決定DECISIONS

Every choice that shapes the ranking, and why. Written so a collaborator (or you in six months) can tell which numbers are facts, which are judgment calls, and which were already argued about and settled.

Last updated 2026-09-01.

1. What this is

A proprietary algorithmic ranking of every Godzilla film, published at godzillarank.com, plus the ability for visitors to build their own rankings from the same data.

The ranking is deterministic and explainable. No machine learning, no black box. Every film's position can be justified in a sentence, which is what makes it defensible as the definitive ranking rather than an arbitrary one.

2. Scope

Score everything, filter on the site. The dataset covers 43 rankable films:

Films carry era, canon_status and continuity tags so the site can offer "canonical 38 only", "live-action only", "Toho only" and so on. Decided: don't restrict the dataset; give users the filter.

3. The evidence model

Five classes, kept in separate files because they are different kinds of claim and must not be averaged together carelessly.

Never impute a missing value as zero. Absence is nearly always ineligibility: Screen Rant's list predates four films, Paste and Variety exclude the anime trilogy, RT publishes no Tomatometer below five reviews. A missing score is an absence of evidence, not a bad verdict.

Prefer average ratings over thresholded percentages. RT's Tomatometer is binary — the share of critics who were positive. Its average rating is not. Destoroyah's 100% Tomatometer averages 6.7/10; Minus One's 99% averages 8.5/10. The composite uses the average.

Sample sizes must be the real ones. RT displays banded audience counts ("5,000+") that are wildly wrong — Son of Godzilla displays "5,000+" against a true denominator of 676, and The Return of Godzilla's Popcornmeter rests on 28 ratings. sample_size carries likedCount + notLikedCount from RT's scorecard JSON, verified to recompute every published score to within 0.5pp.

4. The regions

Audience scores carry a region. One region, one vote.

A source must cover ≥30 of 38 films to be a pillar input. Below that it is display-only, because films would otherwise be scored on different recipes. This excludes Metacritic (10), eiga.com (10) and AlloCiné's press score (6).

Not weighted by sample size. That would hand the ranking back to IMDb and Letterboxd's huge English-language samples and undo the whole point. Per-film confidence is handled by shrinkage instead (§5), where the variation actually is.

Not weighted by population. Douban is not a sample of 1.4 billion people; it is a self-selected platform skewing urban and cinephile. Population weighting prices in people the data never touched, and would have to apply to India, Indonesia, Brazil and Nigeria — none of which have Godzilla rating data at all. One region one vote is a claim about independent measurements, not about people.

5. The algorithm

rank.py, ~120 lines, no dependencies.

1. Normalise every score to 0–1 (value / scale_max).
2. Shrink toward the source's franchise mean by sample size: (n·x + m·prior) / (n + m), where m is that source's median n. A film rated by 200 people cannot swing as hard as one rated by 200,000. This is what handles Godzilla 1985's 194 Filmarks ratings and Destoroyah's 6 critic reviews.
3. Z-score within each source, so the source's own scale and generosity drop out and what remains is "where this film sits among Godzilla films, on this source".
4. Average sources within a region → one score per region.
5. Average regions → the audience pillar.
6. Combine three pillars by the chosen weights, renormalised over whichever pillars that film actually has.

Output carries composite, all three pillars, all six region scores, n_regions, n_pillars, and contentiousness (the spread across pillars) so the site can show why a film ranks where it does.

6. The house ranking

Godzilla Rank = the audience-led preset: 70% audience, 15% critics, 15% critic-lists. Marked OFFICIAL in rank.py.

Naming matters here, because two earlier names were misleading. Both remaining critic pillars are critic-derived, so critics are counted twice. The honest axis is audience voice vs critic voice:

A 1/3–1/3–1/3 split is not neutral; it is 33/67 in favour of critics.

Why audience-led — representativeness, not output. Count the people behind each half of the composite across the canonical 38:

A 50/50 split hands 2,176 critics and 8 writers the same weight as 8.75 million ratings. Even at 70/30 the critics remain enormously over-represented per head — audience-led is still generous to them.

This is deliberately not the argument first made for audience-led, which was that it "says something the English-language web doesn't already say". That was an argument about wanting a particular result, and it was correctly rejected as a basis for a default. The representativeness argument holds regardless of what the ranking produces, which is the test a default has to pass.

What is given up, knowingly: the critic pillars carry real information that the audience pillars do not, particularly about the older films. Expect to defend the Heisei run ranking above where critics place it, and Godzilla, King of the Monsters! (1956) ranking below.

balanced and critic-led remain first-class slices on the site.

7. Deliberately excluded from the score

8. Website requirement: user-defined rankings

Locked in before any design work, so it cannot be lost.

The atomic unit is the per-film, per-source z-score — already shrunk and normalised. Every slice is a different average over the same numbers. There is one algorithm, not one per view.

build(weights, sources=None, film_subset=None, pillars=None) in rank.py is the reference implementation. data/engine.json (24KB) ships the whole substrate so the browser can do it with no server: z-scores, critic pillars, source→region map, film metadata for filtering, and a starter menu of 11 slices as data, not code.

Documented gotcha: filtering the film set does not renormalise the z-scores. A Showa-only view answers "how good is this among all Godzilla films", not "among Showa films". Renormalise client-side if you want the latter.

Slices produce genuinely different answers, which is the point:

9. How caveats are tracked

Two lists, and the distinction is load-bearing. In gzconfig.py, rendered into both SOURCES.md and the almanac artifact by build_caveats.py so they cannot drift.

• CAVEATS_DEFECT — something is wrong or unknown and can be fixed. A worklist. Currently empty.
• CAVEATS_CONSTRAINT — permanent properties of the sources. Nothing closes these. They are instructions to the algorithm, not unfinished work. Currently 5. A non-empty constraint list is not a sign of incompleteness.

The test for which list an item belongs in: name the action that would close it. If the answer is "acquire data nobody ever created" or "make a moving number stop moving", it is a constraint.

The five constraints: two films have no Tomatometer at all; RT publishes only one audience score type per film (VERIFIED vs ALL, not comparable); Metacritic drifts; Showa yen figures are sparse and one is self-contradictory; Filmarks n scales with recency.

10. Positions reversed along the way

Recorded so they are not re-litigated.

11. Not yet decided

• Website design. Not started. The data layer imposes no constraints on it.
• Whether to publish a "biggest" ranking alongside "best", using box office and attendance.
• Whether to add the two remaining Anglophone lists (AV Club, Forbes/Mark Hughes). Both exist and were deliberately not added — they would deepen the Anglophone skew.
• Kinema Junpo readers' poll (読者選出) — runs separately from the critics' poll and is usually kinder to genre film. The next real Japanese critic axis if one is wanted.

12. Current state