The Arithmetic of a Comment Score, Published While the Contest Was Still Running
Japan's International AI Creator Award published the exact weights behind its best-commenter prize with two days of voting left: 10 points per comment, another 10 for each category you touch, and hearts on your comment worth 10 × log₂(1 + hearts). The curve is the interesting decision, and the referral term is the hole.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

Japan's International AI Creator Award 2026 had 360 published entries by 12 November, and by the organisers' own count those entries had collected 350 written comments and more than 20,000 emoji reactions. The award runs ten categories. Public voting closes on 14 November at 21:00 Japan time, and the commenter prize, called CCCP, closes during a watch party at a minute the organisers deliberately do not announce.
Two days before the end, the organisers did something contests usually avoid. They published the scoring formula, the actual numbers, in a post that also reported interim standings and admitted the weights might change again.
That post is a rare document: a live community score, disclosed while it still matters, with a stated reason for each term. Reading it as a piece of design rather than an announcement is worth an hour.
The six terms
For the period up to 12 November, the CCCP score added up like this:
- A comment you write on someone else's work: 10 points.
- Hearts (❤️) collected by that comment: 10 × log₂(1 + number of hearts).
- Each category in which you have written at least one comment: 10 points.
- Each work you gave an emoji reaction to: 1 point, capped at 50 works.
- Each time you shared a work: 1 point, capped at 20 shares.
- Each person you referred who signs up: 5 points, plus another 20 if that person goes on to write a comment.
Comments on your own work do not count. Neither do hearts you gave yourself.
The organisers also reported who was winning under the corrected formula. First place had written more than 40 comments spread across eight of the ten categories. Second place, a character designer, had written more than 20 comments spanning all ten. Most of the leaders are also entrants, which the organisers describe as the point: makers encouraging makers rather than an audience ranking them.
The heart curve is the design decision
Everything above is bookkeeping except one line. A comment is worth a flat 10 points no matter how good it is, and the applause on that comment is compressed logarithmically.
The published mapping makes the compression concrete. One heart on a comment adds 10 points, three hearts add 20, seven add 30, fifteen add 40. Each additional 10 points of heart bonus costs roughly twice the applause of the one before, so the thresholds run 1, 3, 7, 15, 31. Counted as totals, a comment with 15 hearts scores 50 points against a bare comment's 10, which is five times as much, and 30 hearts reaches about 59.5 points, just under six times. Fifteen hearts at the top of the range buy the same 10 points the first heart bought.
I think this is the right curve for the problem they have. A linear heart count rewards the comment most likely to be clicked, which in a gallery of 360 works is the one attached to the most-viewed entry, which is usually the most polished entry, which is the one that needed the least feedback. Compression flattens that: the bonus keeps growing past a dozen hearts, but every ten-point step costs twice the applause of the previous one, so the favourite stops pulling away at the rate a linear count would hand them, and the cheapest next ten points are another comment on another work. Reactions to works, shares and referrals also pay, but those are capped or cheap for anyone with a script. Volume of attempt is rewarded; volume of applause is not.
There is a second effect the curve produces whether or not it was intended. Going from zero hearts to one heart is worth a full 10 points, the same as writing an entire new comment. So the marginal value of the first reaction is enormous, which makes people reply to each other's comments early, which is exactly the reciprocal behaviour that turns a gallery into a conversation. It also means a small circle of four or five commenters can lift each other efficiently. The organisers can see that in the data; from outside, it is the predictable cost of the first step being the steepest.
The caps are where the bots live
The flat terms are the ones a script can farm. Fifty emoji reactions across fifty works is a 50-point walk through the gallery, and twenty shares is another 20. Neither requires reading anything. The caps are the anti-automation device, and they are set low enough that a person who actually browses will hit them without noticing.
Then there is the referral term, which the caps do not cover: 5 points per sign-up, plus 20 more if the referred account writes a comment. That is 25 points for one throwaway account that leaves one comment, and the formula does not appear to ask what else that account did. Twenty-five such accounts come to 625 points, more than sixty comments from a dedicated critic. The award's rules separately ban multiple accounts and reserve the right to exclude inflation, so this is a score available before detection rather than after it.
The organisers name this problem directly. Their line is that creators in the AI era also arrive with bots, so they have added security, the chair reviews suspicious activity case by case, the formula may be revised again after 12 November, and standings can be corrected retroactively. Last-hour gaming is handled by closing the counting window at an unannounced moment inside the watch party, and by stating that a mid-event position is not an award.
That last pair of measures is honest but not free. An unannounced close is fair against snipers and unfair against anyone with a job in another timezone, because the people who can sit in a livestream at 20:00 Japan time are the people whose evening that is. The award is Japanese-run and the audience is international, so the rule quietly favours one region.
What the score cannot see
Six terms, and none of them measure whether the comment was useful.
A comment that tells a maker their fog layer only exists behind the character, and a comment that says "beautiful work, love the colours" score the same 10 points. The artist's reply does not add anything. Whether the maker changed the render does not add anything. Whether the work had nine views or nine hundred does not change the value of a comment on it, which is arguably correct but means the score does not distinguish the person who watched the unwatched entries.
This is not a flaw unique to a contest. It is what every engagement score does, and the reason to read the formula rather than the leaderboard is that the formula tells you which behaviour will be produced. Here the produced behaviour is: many short comments, spread wide, written early enough to collect a heart or three, from people who also bring a friend. That is a decent description of a healthy fan community, and a poor description of a critique practice.
If you run a review loop, steal the shape and patch the holes
Anyone running a small creative group, a prompt club, or an internal design review has the same problem at a smaller scale: you want feedback that is specific, and the cheapest way to get engagement is to count reactions.
Four things from this formula transfer, and one I would change.
Score the writing, not the applause. A flat value per comment, with reactions compressed, is sturdier than it looks. If reactions dominate, you are scoring the maker's existing audience rather than the critic's attention.
Make revisions replace. One comment per work, editable, with the new text overwriting the old. This is the rule that does the teaching, and it costs nothing to implement. I wrote about the mechanics of that separately in why an award started scoring the comments.
Bonus breadth, do not require it. Awarding 10 points per category touched caps at 100 points here, roughly the value of ten comments, so it nudges people out of their favourite category without letting category-hopping become a strategy on its own.
Cap every zero-effort action and price the invite term at zero. Referral bonuses are the one term in this design with no natural ceiling and no connection to the thing being measured. If you need invites, count them separately from the critique score.
Publish the weights, and date them. The organisers published the formula mid-contest and said it might change, which is braver than a fixed secret. Participants who know the rules play the game you designed; participants who guess play a game you will have to clean up afterwards. The caveat is that a published formula invites optimisation, so publish it together with the revision date and the rule for corrections, which is what this post did.
The numbers themselves expire with the contest. The useful part is the shape I would copy for an internal review loop, which is not quite the award's shape: a flat score for writing one note, a logarithmically compressed score for how many people reacted to it, and no referral term at all. The award pays 10 points a comment, 5 for each recruited account, and another 20 when that account writes a comment; a small team does not need invites, so drop that term and the score gets harder to farm. If your own review system weights those pieces in a different order, you are measuring something else, and the something else is usually popularity. The same award's earlier numbers are worth a look if you want to see how AI entries actually distribute across categories, which I went through when 303 entries had come in.