Methodology

Not stated yet. This site will not name the organisation that publishes it until that identification is published in its own right, and it will not guess at one in the meantime. Nothing on this site is sponsored, commissioned or funded by a third party, and no subject, figure or ordering has been paid for.

If that changes, it is disclosed here first and in the changelog on the same day.

One subject is put to the reader, by name, with no description, no framing sentence and no ordering question before it. The reader presses one of three buttons. The buttons carry these labels, in this order, from left to right:

• Hate it, keyboard key 1, always on the left.
• Shrug, keyboard key 2, always in the centre.
• Celebrate it, keyboard key 3, always on the right.

The order never rotates and the labels never change per visitor, which means the instrument carries a known left-to-right bias rather than a randomised one. That is a deliberate trade: a fixed position is one of the four ways a verdict is carried for a reader who cannot use colour, and it is worth more here than the small order effect it costs.

There is no free-text box, no half-point, no abstention and no skip that is recorded. Closing the page without pressing anything records nothing at all.

Anyone who visits. There is no defined population, because there is no attempt to represent one. No age, country, sex, language or any other attribute is asked, inferred or stored against a ballot, so no subgroup exists to report and none is offered.

None. There is no list of people from which visitors are drawn, so coverage of any population is undefined rather than high or low. A reader who wants a coverage figure will not find one here, and any figure that appeared would be invented.

Non-probability and self-selected. People arrive by whatever route brought them, choose to vote or not, and choose which subjects to vote on. Nobody is recruited, nobody is invited, nobody is quota-filled and nobody is paid. The selection probability of a ballot is unknown and unknowable, which is the reason every inference on this site is confined to the ballots themselves.

Every published figure states its own n: the number of counted ballots inside the window. Below 10 counted ballots a subject has a count and nothing else. There is no percentage, no score and no rank, because there is nothing yet worth printing.

A counting interval, not a margin of error

The figure printed after a score as ± is a counting interval. It describes only the ballots on this site. It is not a margin of sampling error and it says nothing about any population, because a non-probability sample has no sampling error to report. The association whose checklist this page follows tells journalists exactly that: do not report a margin of error for an opt-in sample. So this site does not print one, and it does not print a phrase that could be mistaken for one.

What the interval does answer is narrower and still useful: if this same vote were run again under the same conditions, how far would the net figure wander purely because a count is a count. A subject at ten ballots wanders a very long way. A subject at a thousand does not.

The formula

The interval is an Agresti and Coull style adjusted-Wald interval on the multinomial contrast celebrate minus hate. One pseudo-ballot is added to each of the three cells before anything is divided, which is the standard adjustment and which closely approximates a Wilson interval. Writing h, s and c for the counted hate, shrug and celebrate ballots:

The variance term carries a covariance correction. Two cells of a multinomial are negatively correlated, so the variance of their difference is p̃c(1 − p̃c) + p̃h(1 − p̃h) + 2·p̃c·p̃h over ñ, which simplifies to the line above. Dropping the covariance term would understate the interval, which is the wrong direction for a site whose whole argument is that its numbers are softer than they look.

Why not the plain Wald interval

Because it breaks on exactly the case this site meets first. The plain normal-approximation interval takes its variance from the observed shares, and for a unanimous subject that variance is exactly zero. Twelve ballots all reading celebrate would print : a claim of perfect certainty produced by twelve people. The adjusted interval cannot collapse, because an adjusted share is never zero and never one.

None. No ballot is weighted up or down for any reason, no demographic adjustment is applied, no likely-voter style screen exists, and no post-stratification is performed. A counted ballot counts once. The only place a number is deliberately pulled toward the middle is the board's default ordering, which sorts by the score multiplied by n / (n + 100) so that a small sample cannot lead the table. That is an ordering rule, disclosed here, and it never changes a published score.

There are none. No crosstab is published, because no attribute is collected that could produce one. Where a figure covers fewer ballots than the site total, it is because the window or the subject is narrower, and both are printed beside the figure.

Self-administered on a web page, by mouse, touch or keyboard. There is no interviewer, no telephone, no panel and no incentive. A verdict is written to this site's own table the moment it is pressed, and the recalculated split is returned in the same response, which is why the bars move immediately.

Rolling. Collection began when the site began and has not stopped, so there is no field period to quote. Instead every published figure states its own window: the trailing 30 whole UTC days up to a stated date. A ballot older than the window is still in the table and still in the archive, but it is not in the current reading.

One counted ballot per session per subject per day, latest wins. A browser that votes twice on the same subject inside twenty-four hours has changed its mind, and the earlier row is removed rather than excluded, because it was never a second opinion.

How that is enforced, honestly. The check is a read followed by a write, keyed on a signed cookie named hoc_sid. There is no unique index behind it, because the shared database this site runs on accepts no schema changes. Two requests arriving in the same millisecond could therefore both pass the check. The cost of that race is one duplicate row, not a corrupted score, and this page says so rather than implying a guarantee the schema cannot give.

Ballot tokens. A verdict must carry a token minted by a page this server actually served, bound to one subject and one session, and valid for 10 minutes. The token is single-use only within that life rather than burned on use: burning it needs a unique index on a column that cannot be added. So a replayed token inside its ten minutes is bounded by the dedupe above, where it becomes a replacement rather than a second ballot.

Rate limits. Ballots are limited per session per day and per originating address per minute and per day. Limits are counted from the table itself rather than from memory, because a serverless instance is not around long enough to hold a counter.

Review heuristics. Three rules mark ballots for a human to look at. Burst: 10 consecutive verdicts with every gap under 1.5 seconds. Coordination: 3 or more distinct sessions behind one address casting the same verdict on the same subject in a day. Velocity: at least 30 ballots on one subject in an hour, at least 85% one-sided, and at least 5times that subject's own hourly norm. None of them blocks a vote. A flagged ballot is written, held out of the public numbers and reviewed, and releasing it puts it straight back into the counts.

Exclusions are published, not deleted. Nothing is ever removed from the record. An excluded ballot keeps its row, the excluded count is printed beside the counted count, and every exclusion is dated and reasoned in the changelog.

Identity. No account exists. A raw address is never written: the row holds a keyed hash of the address and a keyed hash of the session, and the address half is cleared after days. The sets out the mechanics.

The counts behind every figure are downloadable from the data page as JSON and CSV. Method changes are dated in the changelog. Anything on the site that looks wrong goes to corrections, which is a form that actually posts to this site and a public ledger of what was asked and what was done.