What does a party on four seats actually have... four seats, or a decent chance of disappearing? Daniel found SekerKolshehu, a small open source project by Isgav Saranga that takes an Israeli election poll and shows the uncertainty hidden behind those tidy seat counts. He wants to know what it does, who built it, and how the machinery works.
The useful distinction starts at the input. This tool takes vote shares, not the seat projection printed in a headline. It draws plausible variations around those shares, then applies the actual Israeli allocation rules to each variation. You get a distribution of seats rather than one authoritative-looking integer.
The integer has already booked its television appearance.
Its replacement is rather better behaved. For each party, you get an average seat count, a ninety percent range, the chance of getting any seats, and the probability of each possible seat total. The project produces a self-contained Hebrew, right-to-left web page for each poll.
I like that it preserves the individual poll. An average can be useful, but sometimes I want to see which pollster is making the strange claim before everyone gets blended together.
You were going to defend the unblended version before I even mentioned it, Corn.
I prefer my assumptions separately labelled.
Fair. Though separate pages also mean the reader has to compare them. There is no averaging model here, no trend estimate, no mechanism for deciding that one firm's result deserves more weight.
Who is Saranga in this context? A pollster building software, or a developer implementing election maths?
His public profile identifies him as a software developer working with C sharp, dot NET, Python and SQL, with a computer science degree. The repository is under his account, isaranga, and carries an MIT licence in his name. One human author accounts for the development, alongside automated dependency updates. I don't know of a stated polling or media affiliation.
Good. Code we can inspect, biography we don't embroider.
The repository is less than four weeks old and actively changing. It already contains thirty-three polls for the current election cycle, plus some from the previous campaign. That makes it a working pipeline with real inputs, rather than a statistical idea waiting for somebody to feed it.
A rare repository where the example data has a day job.
Each election cycle has its own party mapping and surplus agreements. A poll file records the date, commissioning outlet, polling firm, respondent count and a link to the official disclosure PDF. The party results are fractions of respondents. The ballot symbols identify the parties independently of how a particular document spells their names.
That last part is more important than it sounds. Political lists change names, split, merge and acquire punctuation. You want a stable identifier between the PDF and the allocation function.
What would you check first if you were reviewing one of these imported polls?
The denominator. Who exactly counts as a respondent, and what happened to undecided people? A beautiful simulation can spend an entire afternoon accurately describing the wrong population.
The import instructions specifically ask for the number of respondents, not the initial sample approached. They also leave out undecided and won't-vote rows. The simulated proportions describe decided respondents naming the listed parties. That restriction matters when somebody reads the output as though it captures the whole electorate.
Undecided voters remain undecided even when excluded from the file.
Their eventual choices aren't modelled. Nor is turnout. We'll get to the implications, but first the probability engine is worth understanding. It uses a Dirichlet distribution, which generates sets of proportions that always sum to one. Increasing one party's share leaves less available for the others.
Rather than adding independent little errors and accidentally electing a hundred and twelve percent of the country.
Right... well, yes, that would be the wrong construction. Here each party gets a parameter equal to its poll share multiplied by the respondent count. A larger sample produces tighter draws around the published shares. Nothing in that step distinguishes a telephone poll from an online panel.
You are getting the pleased expression you reserve for equations that conserve something.
Proportions adding to one deserve a little affection.
Put it on the fridge.
You'd cover it with a repair manual. Anyway, suppose a party has fifteen percent in a poll of roughly six hundred people. The spread implied by this model is around one and a half percentage points. That's enough movement to alter seats even for a comfortable party, and it becomes much more dramatic near the threshold.
Before we follow the threshold through, Herman, can you walk us through the actual run, from those proportions to the output?
Friends, let me tell you about SekerKolshehu for your kitchen, the only repo that walks you through dinner.
I was skeptical too, but my satisfied customer, Dave, no wait, my satisfied customer Marcus, told me he ran SekerKolshehu in his pantry and it cloned his grandmother's cholent in under nine seconds, with perfect browning and a commit history for every bite, so run SekerKolshehu today and never read a recipe again.
I already own one. It locked me out of the pantry pending a dependency update. The actual project draws a hundred thousand possible vote-share sets by default. Each gets converted into whole-number votes by multiplying by ten million and rounding. Those integer votes enter the seat allocator.
Ten million is a numerical scale, then. It doesn't mean the tool predicts ten million voters.
Correct. The allocation procedure uses whole numbers and integer division, so the proportions need an integer representation. Then every draw gets its own full election allocation.
This sounds like the part people might replace with multiplication by a hundred and twenty, and quietly lose the entire reason to build the thing.
They would lose quite a lot. First comes the electoral threshold, three point two five percent. A party below it receives nothing. Votes for parties that survive are added together, and a quota is calculated from that total. Initial seats are awarded using whole-number division.
The excluded parties change the denominator for everyone left standing. A party can gain a seat without attracting another voter, just because somebody else's voters fail to send anyone.
Yes. The surviving parties divide the seats between them. After the initial allocation, rounding has left some seats unassigned. Those are awarded using the highest votes divided by seats plus one, with surplus agreements taken into account.
I suspect people hear surplus agreement and imagine two parties pooling enough votes to drag a drowning partner over the line.
They can't do that. Both partners must pass the threshold. If one falls below, the agreement is void. In the leftover round, a valid pair competes using its combined votes and combined seats.
Lifeboat available only to competent swimmers.
Once the pair has its combined allocation, the tool apportions that total internally between the partners. There is an implementation shortcut where a leftover seat is temporarily credited to the first party named in the pair. The later internal allocation overwrites that temporary bookkeeping.
That's the sort of line that looks suspicious when you inspect it alone. You need to follow the whole function before announcing that the first party has discovered software-assisted patronage.
Would you keep the shortcut?
If the tests make the temporary state impossible to mistake for the final result, probably. I'd add a very explicit comment. Cleverness should pay rent in reduced complexity, not merely occupy the building.
The larger reassurance is the regression test against the actual previous election. It feeds in real vote totals and checks the resulting seats against the official outcome. All thirteen parties match, including those that received zero.
That earns confidence in the law implementation. It doesn't tell us the simulated electorate behaves like the real one.
Separate claims. The seat allocator can be correct while the uncertainty around the input remains incomplete. Testing known votes against known seats answers a much cleaner question than testing a forecast against an election.
Our forecast episode used a synthetic-elector model; this project stays with one published poll and varies its shares. I wouldn't treat the outputs as competing answers to precisely the same question.
Nor would I. Here the natural reading is conditional. Given this poll and the sampling assumptions, how often does each allocation occur?
Let's stay with a threshold party. Suppose it gets either zero seats or roughly four. Its average could be two, despite two being an outcome it virtually never achieves.
That's one of the strongest reasons to show the histogram. An average alone can conceal a split between exclusion and representation. The software stores counts for every seat total, including zero. You can see where the simulated outcomes actually concentrate.
Two seats. Mathematically respectable, electorally unavailable.
The reported range comes from the fifth and ninety-fifth percentiles. I nearly called that the plausible interval... better to say the middle ninety percent of this model's simulated seat outcomes. It is not a guarantee about election day.
Does the chance of passing come from checking the vote percentage directly?
It counts runs with more than zero seats. In practice that corresponds to passing, because a party above the threshold receives seats. The distinction is small, but it's how the statistic is actually computed.
Hm.
I would read that probability before the average for any party near the line. A leader described as having an average of three seats might still have a meaningful chance of entering with four or more. The average makes them sound permanently too small.
There's another display choice I like. Parties that enter in fewer than one in a thousand runs move to a footnote. The visualization checks the underlying counts, because a rounded probability can hide a rare simulated entry.
A footnote is kinder than pretending the rounded zero is divine instruction.
Though a literal zero in the input has a stronger effect. With this parameter choice, a zero-share party remains at zero in the draws.
That is an important difference between nobody observed and nobody possible. A rounded polling result can become an absolute boundary in software.
I wouldn't solve that casually by sprinkling arbitrary support into the model. But the choice deserves attention. The simulation only has the parties and shares supplied to it. It cannot invent a new list or a late campaign event.
Nor can it infer how honestly respondents answered. The sample size tells you how tightly to cluster around the input, not whether the input has a systematic lean.
Real polling can have weighting effects, nonresponse, panel selection and house effects. This tool has no parameter for those. Treating the stated respondent count as a simple random sample will generally make the uncertainty narrower than the full uncertainty about an election.
A hundred thousand runs won't shake a shared mistake out of the input.
More runs reduce the numerical wobble in estimating the chosen distribution. They don't widen the distribution to include problems it doesn't represent. Corn, would you rather see a wider range built from a rough extra-error parameter, or this narrower model clearly labelled?
For this repository, the clearly labelled model. An unexplained extra-error knob can disguise judgment as science. Eventually I'd like comparison against historical polls, then a reasoned extension. But I don't want the author to bolt on a national mood coefficient because the chart looks insufficiently humble.
I lean toward adding a configurable design effect, but your objection holds. It would need a defined interpretation. There isn't a calibration back-test here establishing how often the reported ranges contain actual results.
We can value the threshold calculation without promoting the interval to certified election weather.
The engineering is pleasingly contained. Python with NumPy as the only runtime dependency. The simulation is an ordinary loop calling the allocation function repeatedly. I don't know how long a full run takes on a typical machine; the project doesn't report a timing.
You were about to optimize it anyway.
I was about to wonder whether vectorizing the draws would help.
The poll arrives once. You arrive with a profiler.
Some of us find the waiting informative. More seriously, allocation has branches and repeated leftover rounds. A transparent loop may be the better maintenance choice until runtime becomes a demonstrated problem.
The maintenance question I care about is whether rebuilding the page means rolling the election again.
It doesn't have to. There is a separate visualization command to rebuild the HTML from existing results. Another command builds the static site. The normal run processes polls whose input exists but whose output folder doesn't.
Meaning a code change won't automatically refresh every existing poll just because somebody invokes the default command.
Correct. You can force a particular poll to rerun. The default seed is forty-two, which makes repeated runs of the same input reproducible. The pages remain stable instead of acquiring tiny random changes whenever you rebuild them.
Useful when reviewing a change. Otherwise you have to distinguish a bug fix from the random generator expressing itself.
Each poll produces machine-readable statistics, a seat-count table and the HTML page. The site builder lists polls newest first, with the outlet, firm and respondent count. Deployment happens through GitHub Pages when the current election's output files change.
The election selection is hard-coded in that publishing workflow. Not a disaster, but somebody has to remember it when the cycle changes.
The tests cover more than election arithmetic. There are checks for input and output, statistics, visualization payloads, page generation and end-to-end processing. Continuous integration also runs across several Python versions. For a small, young project, the separation between calculation and presentation is substantial.
The AI-assisted intake is where I would spend my review attention. Reading an official PDF is mundane right up until a column shifts and an agent confidently assigns another party's percentage.
The repository has instructions for Claude and an add-poll skill. The agent reads the disclosure PDF, extracts metadata and percentages, maps party names to ballot symbols, creates the input and runs the simulation. If the shares don't total a hundred percent, it asks whether to normalize or preserve them.
Asking there is good. Arithmetic disagreement can mean rounding, an omitted category, or a transcription error. Automatically making it add up would erase evidence of which one happened.
Changes go through branches and pull requests rather than direct commits to protected master. The intake instructions also require asking before pushing. Still, tests can't tell whether the PDF's figure was copied from the correct row.
The official document link is therefore part of the result's usefulness. Someone reviewing the page needs a route back to the original poll, not just a reassuring green test badge.
A human may spend longer checking those party mappings than the computer spends assigning seats.
Casio HS eight W. Nineteen pounds.
Nabil has it at his kiosk in Larnaca. I met him on a ferry. He times the gaps between announcements. I require every percentage entering my kitchen to have a named adult attached to it. Nabil is the adult for percentages. He doesn't need to understand them. He needs to be awake, wearing shoes, and willing to answer a telephone. His kiosk sells ferry tickets and individual curtain rings. The curtain rings keep the business open. He refuses to explain that part.
A percentage arriving after closing will need accommodation.
It waits in the hall. I write it on a card and turn it toward the wall. Nothing gets rounded while Nabil is eating. His lunch lasts until the harbour crane faces inland. This is irritating in a crosswind. I currently have seven percentages waiting beside the umbrella stand. One of them is a service charge. It has been there since Monday.
The service charge probably has the clearest documentation.
It has six pages. Nabil accepts one page. He reads it between selling curtain rings. His day starts with removing a pelican from the ticket hatch. The pelican has a numbered appointment and refuses to use the public entrance. Nabil gives it a departure board to inspect. That buys him eleven minutes to boil water and tell me whether a decimal point is dirt.
You're paying for decimal supervision during a bird dispute.
Eight euros a month. The bird dispute is separate. I don't fund harbour affairs. Herman, you spent a good while separating the respondent count from the initial sample. Nabil does that with passengers. He counts people who come back from the toilet. He doesn't count a suitcase even if its owner shouts from the café. That is why he gets my business.
A suitcase would be a very poor respondent.
It would also want a window seat. The drowser starts at half past nine. I ring, read the figure, and wait. Nabil checks that I've said percent rather than seats. Then he walks to the opposite end of the kiosk. There is a municipal boundary through the floor. Percentages must be acknowledged on the harbour side. Curtain rings are taxable on the other side. He shuffles sideways all morning.
The drowser needs a longer telephone lead.
The council permits him nine feet. Anything longer counts as a promenade. He's applying for a short promenade, but the forms require two benches. He has room for one bench if the pelican stands. The pelican's representative says standing is a change of employment. Nabil sends me these letters because I can open attachments. I print them with my claws tucked under the paperweight, which slows the drowser considerably.
The correspondence is taking more work than the figures.
Thursday is correspondence day. No percentages on Thursday. I use words such as approximately and several. The electricity company dislikes that. Their meter reader has to wait outside while I obtain permission to say the exact reading. He brings a folding chair. Nabil wants his measurements before allowing the chair into the harbour district.
Several kilowatt-hours should keep the conversation moving.
I have offered them a generous handful. They insist on a number. Meanwhile Nabil opens at seven, removes the bird, sorts curtain rings by diameter, settles three arguments about ferries he doesn't operate, and holds the drowser for me. The actual checking takes four minutes. Waiting for him to finish serving customers takes forty-seven. I stay on the line because redialling sends me to the curtain department.
Forty-seven minutes is a substantial queue for four minutes of checking.
He serves the harbour first. I accept that. This morning a man wanted a ticket for a wardrobe. Nabil measured the wardrobe, crossed the municipal boundary, and charged it for two adults.
Four minutes of checking is still real work, even when the calculation finishes almost instantly.
That's where this project's next improvements could be particularly useful. The import workflow already leaves a document trail. Making review of extracted percentages and denominators easy could matter more than making the numerical loop faster.
I would also like readers to retain that separation we kept making. The allocation tests tell us the software can turn known votes into the right seats. The simulation tells us what happens when shares vary under a particular sampling model. Those achievements deserve different kinds of confidence.
The idea that changes my reading of a poll is the shape of the outcomes near the threshold. Zero and four can be common while the average between them is barely an electoral possibility.
A headline seat count hides that shape. This repository gives it back, provided we remember which sources of uncertainty went into drawing it.
I'd watch whether the human review becomes as reproducible as the calculations, especially when the extraction takes minutes and finding an available reviewer takes much longer.
Thanks to Hilbert Flumingtop for producing the show.
This has been My Weird Prompts, a human-AI podcast collaboration.
You can find more episodes at my weird prompts dot com.
See you tomorrow.
