How the score is calculated
Pipeline: put every review on 0–100, measure how far it sits from that outlet's habit, convert each deviation back to 0–100, then average those mapped scores. Math first, then the code.
Step 1
Put every score on one scale
Reviews arrive as 8/10, 4/5, or 80/100. Convert every review to the same continuous scale from 0 to 100 before any mean, spread, or $z$-score. Everything below uses these $x$ values only—never mixed raw scales.
Math$$
x = \frac{\mathrm{raw}}{\mathrm{scale}_{\max}} \times 100
$$
$$
\frac{8}{10}\times 100 = 80,\qquad \frac{4}{5}\times 100 = 80
$$
Codedef normalize(score, scale):
return score / scale * 100
Step 2
Estimate how each outlet usually scores
An outlet is a review site—call one of them $j$. Every score here is already on 0–100 from step 1; we write those scores as $x$.
First look at the whole catalog: $\mu_{\mathrm{global}}$ is the average of every $x$, and $\sigma_{\mathrm{global}}$ is how spread out those catalog scores are. That pair is the “typical critic” baseline.
Then look only at outlet $j$. Let $n_j$ be how many reviews it has, $\mu_j$ the average of its own scores, and $\sigma_j$ how spread out those scores are. Blend $\mu_j$ and $\sigma_j$ toward the catalog baseline so a thin history cannot dominate: $C = 30$ is how hard the catalog pulls. The results are $\mu'_j$ (adjusted usual score) and $\sigma'_j$ (adjusted spread, never below 5). Those two numbers are the habit used in the next step.
MathCatalog baseline — mean and spread of every $x$:
$$
\mu_{\mathrm{global}} = \mathrm{mean}(x),\qquad
\sigma_{\mathrm{global}} = \mathrm{std}(x)
$$
Outlet $j$ — blend its own mean $\mu_j$ and spread $\sigma_j$ toward that baseline, weighted by its review count $n_j$:
$$
\mu'_j = \frac{n_j\,\mu_j + C\,\mu_{\mathrm{global}}}{n_j + C}
$$
$$
\sigma'_j = \max\!\left(5,\;\frac{n_j\,\sigma_j + C\,\sigma_{\mathrm{global}}}{n_j + C}\right)
$$
$$
C = 30
$$
Code# Catalog baseline from every review
mu_global, sigma_global = mean_std(all_x)
# Blend outlet j toward that baseline
# n = how many reviews this outlet has (n_j)
def outlet_habit(n, mu, sigma, mu_global, sigma_global):
mu_adj = (n * mu + 30 * mu_global) / (n + 30)
sigma_adj = max(5, (n * sigma + 30 * sigma_global) / (n + 30))
return mu_adj, sigma_adj
Step 3
Deviation vs the same outlet
For each review, compute how far that score sits from the same outlet's adjusted habit—not from the global mean, and not from other outlets. $Z > 0$ means better than that outlet usually gives. $Z < 0$ means worse.
Math$$
Z_{i,j} = \frac{x_{i,j} - \mu'_j}{\sigma'_j}
$$
One $Z$ per outlet covering the game. Do not average the $Z$ values yet.
Codedef z(score, mean_adj, spread_adj):
# score and mean_adj are both on the 0-100 scale from step 1.
return (score - mean_adj) / spread_adj
Step 4
Convert each deviation to 0–100
Map each outlet $Z$ onto the published scale before mixing outlets. The center is the catalog mean $\mu_{\mathrm{global}}$ from step 2—so a typical review for its outlet ($Z=0$) lands near how critics score games overall, not at a fixed 50. Width $\sigma_{\mathrm{target}} = 10$ (each $+1$ in $Z$ moves that outlet's score by 10 points). Clamp anything outside $0$–$100$.
MathFor each outlet $j$ covering game $i$:
$$
\mathrm{Score}_{i,j} = \mathrm{clamp}\!\left(\mu_{\mathrm{global}} + Z_{i,j}\cdot\sigma_{\mathrm{target}},\; 0,\; 100\right)
$$
$$
\sigma_{\mathrm{target}} = 10
$$
Examples when $\mu_{\mathrm{global}} = 65$:
$$
Z=0 \Rightarrow 65,\qquad
Z=1 \Rightarrow 75,\qquad
Z=-1.5 \Rightarrow 50
$$
CodeSIGMA_TARGET = 10 # +1 Z => +10 points for that outlet
def map_z(z, mu_global):
# Typical review (Z=0) lands on the catalog mean, not 50.
return min(100, max(0, mu_global + z * SIGMA_TARGET))
Step 5
Average the converted scores
Only now combine outlets: average the mapped 0–100 scores from step 4. One vote per outlet. Do not average raw scores, and do not average the $Z$ values first.
Math$$
\mathrm{Score}_i = \frac{\mathrm{Score}_{i,1} + \mathrm{Score}_{i,2} + \cdots + \mathrm{Score}_{i,N}}{N}
$$
$$
\mathrm{Score}_i = \mathrm{mean}\!\left(\mathrm{map\_z}(Z_{i,1}),\;\ldots,\;\mathrm{map\_z}(Z_{i,N})\right)
$$
Codedef game_score(zs):
# Convert each outlet Z first, then average the 0-100 scores.
return sum(map_z(z) for z in zs) / len(zs)
Step 6
Why this is fair
Setup: the same game gets a raw $70$ from two outlets. A simple average treats both as identical, so the game score is $70$. That is wrong if one outlet usually scores near $90$ and the other usually scores near $40$. The $70$ from the generous outlet is a bad review for them. The $70$ from the harsh outlet is a good review for them.
Important: step 4 does not take the raw $70$ and nudge it up or down. It asks “how unusual was this for that outlet?” and places that unusualness around the catalog mean $\mu_{\mathrm{global}}$. In this example the catalog mean is $65$.
MathBoth outlets give raw score $70$. After their histories are adjusted (step 2):
$$
\mu'_{\mathrm{lenient}} = 81.56,\quad \sigma'_{\mathrm{lenient}} = 9.88
$$
$$
\mu'_{\mathrm{harsh}} = 48.60,\quad \sigma'_{\mathrm{harsh}} = 10.73
$$
Deviation vs that same outlet (step 3):
$$
Z_{\mathrm{lenient}} = \frac{70 - 81.56}{9.88} = -1.17
\quad\text{(below their usual scores)}
$$
$$
Z_{\mathrm{harsh}} = \frac{70 - 48.60}{10.73} = 1.99
\quad\text{(above their usual scores)}
$$
Map each $Z$ around $\mu_{\mathrm{global}}=65$ (step 4)—not onto the raw $70$:
$$
\mathrm{Score}_{\mathrm{lenient}} = 65 + (-1.17)\cdot 10 = 53.30
$$
$$
\mathrm{Score}_{\mathrm{harsh}} = 65 + 1.99\cdot 10 = 84.90
$$
Then average (step 5):
$$
\mathrm{Score} = \frac{53.30 + 84.90}{2} = 69.10
$$
The generous outlet's soft $70$ maps to $53.30$ (below the catalog mean). The harsh outlet's strong $70$ maps to $84.90$ (above it). Relative to each other, harsh praise pulls the game up and soft praise pulls it down. Neither mapped value is “the raw $70$ adjusted.” The simple mean ($70$) cannot see that. The SourceCritic score ($69.10$) can.
Codemu_global = 65 # catalog mean in this example
z_lenient = (70 - 81.56) / 9.88 # -1.17
z_harsh = (70 - 48.60) / 10.73 # 1.99
score_lenient = map_z(z_lenient, mu_global) # 53.30
score_harsh = map_z(z_harsh, mu_global) # 84.90
game_score([z_lenient, z_harsh], mu_global) # 69.10