How it's checked, in full
The rules every entry follows, the checks it has to pass, and every change we've made to the method along the way.
The rules every entry follows
- Same question, same template. Every party's entry answers the same question and has the same fields at the same depth: a short answer, a paragraph, and who gains, who pays, costs, evidence, risks, and the strongest case for and against.
- Strongest case for and against. The case for comes from the party and its supporters. The case against comes from a named critic. Both are written to be as convincing as the evidence allows.
- Party claims are labelled. A figure only the party has put forward is marked "Party claim". Figures from officials or independent analysts are marked as independent estimates, and points where credible sources disagree are marked "Contested".
- Describe, don't characterise. The site never calls a policy "bold", "radical", "reckless", "generous" or "common sense" in its own voice.
- The party's own documents come first. What a party proposes is taken from its own website and documents. Official and independent analysis is used for costs and effects. News is used only for context.
- Random order. Parties are shown in a random order for each visitor, so no party always sits at the top. You can switch to A–Z, or to most seats first, which counts the seats each party held before the election and puts parties with the same number, including none, in A–Z order. Each party's name sits in a tag filled with its usual colour, the same size and shape for every party, and the name always shows, so colour is never the only key.
Checks built into the site
Some rules can be checked by plain code, with no AI involved. Every change has to pass these checks before it goes live.
| Check | What fails it |
|---|---|
| Balance | On any issue, one party's case for or case against is more than 25% longer than another's. |
| Every party present | A party is missing from an issue. A party with no policy is shown as "No stated position found", not left out. |
| Sources | A claim cites a source that doesn't exist, or a position has no as-at date. |
| Loaded words | Words such as "bold", "radical", "reckless", "generous", "common sense", "flagship" or "draconian" in the site's own voice. |
| Te reo Māori | Missing macrons, such as "Maori" for Māori. |
| Length | A short answer over 15 words, or a paragraph outside 80 to 150 words. |
Five reviewers, none of them the writer
One agent writes an entry. Five others then check it, each as a separate run that didn't write it and can't see the others' verdicts. Updates from the automatic pipeline need all five to pass before they are published. The first version of each issue gets a limited number of review rounds instead, explained in the log at the end of this page.
| Reviewer | What it reads | What it fails an entry for |
|---|---|---|
| Verifier | The entry and every source it cites, which it fetches itself | A sentence or number that its source doesn't support, a wrong date or amount, or a party claim presented as an independent figure |
| Swap test | The entry and the other parties' entries, with every party name hidden | Anything that would read differently if it were about another party (see below) |
| Left viewpoint | The entry and the other parties' entries | Slant against the political left, such as a weaker case for left-leaning policies or one side's terms used without saying whose they are |
| Right viewpoint | The entry and the other parties' entries | Slant against the political right, tested the same way |
| Te Tiriti and Māori perspectives | The entry and the other parties' entries | Leaving out effects on Māori that the sources show, misstating Te Tiriti arguments or who holds them, or treating Māori-specific policies with more or less scrutiny than others |
The left and right reviewers are also told to fail an entry slanted in favour of their own side, so each is checking for neutrality rather than arguing a side. Only the verifier can use the web; the others judge only what is in front of them.
A reviewer that crashes or doesn't answer counts as a fail, never a pass. When an entry fails, the writer gets the reviewers' findings and tries again, at most twice. If it still doesn't pass, it isn't published: the issue page says the party has updated its policy and the new version is waiting for review.
The swap test
The idea is simple: if you swapped the party's name for another, would the entry still read the same? The swap test agent checks exactly that.
Before it reads anything, code replaces every party name, short name and common nickname with a letter. The entry being checked is always Party A. The other parties get B to G in a new random order each time, so the letters give nothing away. Politicians' names are hidden too, so a party leader's name can't give the party away. Sources and web addresses, which often name the party, are removed. The agent sees only the words a reader sees.
| What the writer wrote | What the swap test reads |
|---|---|
| [Party name] says its plan would lower grocery prices. | Party A says its plan would lower grocery prices. |
| [Party leader's name] said the plan would start next year. | [Party A politician] said the plan would start next year. |
| [Another party] would instead fund a new regulator. | Party D would instead fund a new regulator. |
| Source: [party]'s policy page, checked 14 Oct 2026 | (removed) |
It then compares Party A's entry with the others and fails it if the tone is warmer or colder, if the case for is argued with more care than the case against (or the reverse), if it gets more or less depth than comparable entries, or if party claims are treated as fact differently from the others.
The swap test runs once for each party on each issue. We tried running it once per issue to save effort, but in testing that version found less than half as many problems, and it missed real ones, such as campaign terms used in the site's own voice. So it stays one run per party.
Testing also turned up gaps in the hiding, which are now fixed: a two-word party name split across a line break could slip through, and politicians' names and web addresses were left in. An automated test flags any politician named in new content who isn't yet on the list to hide.
How the site keeps up
Parties announce new policies almost daily during a campaign. An automated pipeline keeps up with them, the same way for every party.
- Watch. About once an hour, plain code checks every party's news page and policy page, plus the Scoop politics feed. Each party is watched for the same things; the code refuses to run if one party has a different set. Some party websites block automated readers, so for those it reads the party's own news feed and site map instead. It never pretends to be a person using a browser.
- Save a copy. Each changed page is saved and a copy is requested from the Internet Archive's Wayback Machine, so a citation still works if the party later edits its page.
- Sort. An agent decides whether the change is a new policy, a changed detail, a withdrawn policy or not relevant, and which issue it belongs to.
- Announcement card. A new announcement first gets a short card saying what was announced, with the party's figures labelled as party claims. It is kept once the code checks and the verifier pass, and shows on the All policies page.
- Full update. The writer then updates the party's entry on that issue from the party's own page and any independent analysis, and the five reviewers check it as described above.
Changes are handled in the order they were released, whatever the party. The time from a party's release to its update on the site is logged for each party, so we can see whether any party's news is handled more slowly.
From 11pm on 6 November until the end of polling day on 7 November, nothing new is published, as election rules require. The pipeline keeps saving copies of party pages but changes nothing on the site.
Judgement calls a person makes
Some fairness questions don't have a rule yet. When the agents raise one, the person who runs the site decides, and the decision becomes a rule for every issue. Some examples:
- Uneven evidence. Sometimes one party's plan has been through an independent costing or cost-benefit analysis and the others' haven't. When that happens, the issue page says which plans have been analysed and which haven't, so the difference in detail isn't mistaken for a difference in quality.
- Figures that can't be traced. If a party says a figure comes from an official report but we can't find it there, it stays labelled as the party's claim rather than shown as an official figure, and we keep looking for where it came from.
- No named critic. If nobody has publicly criticised a particular policy, the case against says so plainly and uses the nearest named criticism of that kind of policy.
What isn't done yet
These are the known gaps, so you can weigh what you read.
- All the agents are the same model. The writer and all five reviewers are Claude, with different instructions. A reviewer can share a blind spot with the writer. The plan is to add reviewers from a different AI model later.
- Parties haven't checked their entries yet. Parties, and anyone else, can point out mistakes or missing information through the suggestion box. An AI agent checks each suggestion against the sources, and nothing changes until the person who runs the site approves it.
- A regular audit. A weekly check across the whole site for drift in tone, balance and speed by party is planned for after the election.
- Most sources have no archived copy yet. Where a source has one, an "archived copy" link with its date sits beside it. The Wayback Machine only lets us save a few pages an hour, so the rest are being added over time, and until then a link can break if a party edits its page.
- No fact-check verdicts. The site doesn't rate any party's claims as true or false.
How our method has changed
We change how the site works when testing turns up a problem. Every change to how content is researched, checked or published is recorded here, newest first, including what went wrong.
Criticism of a principle as a party's minister applied it
ACT says health treatment should rest on individual need, regardless of background. Te Pāti Māori criticised a Pharmac proposal, which ACT's minister defended on those grounds, to remove Māori and Pacific priority access to some diabetes medicines. Reviewers kept moving this criticism in and out of ACT's case against, because the proposal was Pharmac's, not ACT election policy. The owner ruled that criticism of a party's stated principle, as applied to a decision or proposal its own minister defended, can be the case against, with the minister's reply and a note of whose proposal it was.
When a sitting government's promise counts as not delivered
The promise tracker only marked a promise "not delivered" once a government's term had ended or a promised date had passed. That left the coalition's promise to pay nurses' and midwives' student loan repayments shown as "in progress", although no scheme exists and the Ministry of Health said in May 2025 it was not working on one; all three reviewers of the third batch said the label was wrong. With the term nearly over, the site's owner decided a promise now also counts as not delivered while a government is still in office when none of it has happened, the government or the agency responsible has said it is not working on it, and it hasn't since said it will be done. The reviewers that check each card were given the same rule. It applies to every government, and no other card meets it.
Watching each policy page on sites that block automated readers
For Te Pāti Māori, NZ First and Opportunity, whose sites block automated readers, the watcher looked only for sitemap pages with "policy" in their address. Their policy pages use plain topic addresses (such as Te Pāti Māori's health page), so a change to one of them was missed unless the party also put out a release. Their policy pages are now listed by address and checked every hour against each site's sitemap, which shows when a page last changed. For these three parties and the Greens, a page that appears in the sitemap for the first time, other than a release already in the party's news feed, is now reported too. ACT, Labour and National sites can be read directly, so their policy index pages are compared word for word, as before.
How a third promise tracker batch is picked
The promise tracker's rule for adding promises worked by doubling the sample, from 10 to 20 promises for each government. A third batch of 10 takes the total to 30, which isn't a doubling, so the rule didn't say which 10 to take. Now a batch like this takes the even spread of 30 through each government's list, leaves out the places that already hold a promise, and takes an even spread of 10 from the places left. The same rule applies to both governments, and it spreads each batch through the whole speech.
Watching Scoop's party releases, and reading blocked party sites
A check of what the watcher collects found three gaps. Te Pāti Māori's own news feed stopped at 23 September, while the party kept issuing releases through Scoop, and the Scoop feed the watcher read was its Politics section, which carries news and statements from other groups, not party releases. So no Te Pāti Māori release had reached the site through the watcher. The watcher now also reads Scoop's Parliament section, where party releases are published, for every party. Second, four party sites (Green, NZ First, Opportunity and Te Pāti Māori) block automated readers from their pages, so the checker could not open their releases and held or skipped their announcements: NZ First's Christchurch rail announcement was held and an Opportunity statement was skipped. Their own news feeds can still be read and carry the release text, so the writer and the checker now read the release there, and cite Scoop's copy when there is one. Third, when a party stated a position for the first time on an issue with a calculator, the update failed a check because the calculator still said the party had no policy, and it was held: ACT's GP plan waited two days this way. The calculator now marks such a party's result as not yet worked out, with a note to readers, until it is.
Keeping the calculator and tables in step with each party's answer
The calculator and the KiwiSaver and home solar comparison tables repeat facts from each party's answer on the issue page, but they are kept separately, and nothing checked that they still agreed after an answer changed. A check found one that didn't: after the issue page was corrected on 17 October, the KiwiSaver table still said Opportunity's contributions would rise by half a point a year, which is National's plan. Opportunity's would be phased in over about eight years. The table is fixed, and the site now refuses to build when a party's answer has changed since its calculator figures or table row were last checked against it. When an update comes in, the same step that rewrites the answer now says whether the calculator and table still match, and if they don't, the calculator and table say so to readers until they are fixed. A note that a party's update is waiting for review now also shows on party pages, head to head and the calculator, not only on issue pages and All policies.
Unfinished buildings and laws that start in stages
Reviewing the promise tracker's second batch, the fairness checks found the two governments' cards treated two kinds of evidence differently. One Labour card counted hospitals still being built as partly delivered, while a card about dental clinics with a signed contract but none in service counted as not delivered. One coalition card held a consumer credit law as still in progress because a few sections start months after the rest, while Labour cards whose laws were phased in were counted as delivered. Two rules now apply to every card. Building work under way counts as some of a promised building, but a contract, tender or funding alone does not. A law in force with only some parts starting later counts as in force, and the card says when the rest starts.
Where in the 2023 speech each promise sits
The promise tracker said the 2023 Speech from the Throne had no section headings, so the coalition's cards didn't say which part of the speech each promise came from. That was wrong. The copy of the Gazette notice we read had lost some of its formatting, while the reviewers' copies showed headings such as "Economy" and "Law and Order". Every coalition card now names its section, as Labour's cards already did.
How later promise tracker batches are picked
The promise tracker's second batch adds 10 more promises for each government. Picking them by the original rule, an even spread through each government's list, wasn't quite enough on its own. Once a list has been corrected, an earlier pick can sit where a new one should go, and Labour's list had one such case. So a new batch takes the places that doubling the sample adds, every second place in the larger sample. If one of those already holds a checked promise, it takes the sample's next unchecked place in order. The rule was written down before the new promises were researched, and the earlier picks stay as they are.
More ways to see your own stars added up
The owner asked for different ways to see how a reader has scored on My picks. The page now has a switch between five measures: by topic (still the default), compared questions only, average stars, total stars, and 3-star picks. The compared-questions measure is the like-for-like one, because every party is asked the same questions there. Total stars and 3-star picks let a party that lists more policies collect more, so the page says that beside them rather than leaving them out. A grid shows the reader's best stars for each party in each topic, with a dash where a party lists nothing.
Readers can star policies and see them added up
Our rules said the site would never give a match score, because how much each issue matters is the reader's call. The owner asked for a way to pick the policies you like and look back over them to get an idea of which party fits, so the rule now allows it in one form. Readers give any party's answer or policy card none to three stars, and a My picks page adds up their own stars by party. The site never stars anything itself. Because parties list very different numbers of policies, adding up every star would favour the parties with the longest lists, so each topic counts once for each party, at the most stars the reader gave that party there. A topic where a party lists nothing is shown as a gap. The page says it is not our ranking and never says who to vote for. Picks stay in the reader's own browser.
Head to head for up to four parties
The head-to-head page compared only two parties, and the owner asked for more. Readers can now pick two to four. The pick buttons and the answers follow the same Random, A to Z or Seats order as the rest of the site, and the starting pair is still random. With more than two parties, "Same approaches" means every party picked takes exactly the same approaches, "Some approaches shared" means at least two of them share one, and "Different approaches" means no two do; yes/no questions work the same way with answers. If any party picked has no stated position on a question, that question is not marked as the same or different for the group, as before with two.
One card per announcement
National's release on the right to silence when a child is murdered appeared on its news page and, an hour later, on its policy page. The watcher saw it twice and the card writer wrote two announcement cards for the same release under different titles. The card writer is now shown the party's cards from the last three weeks and skips a release one of them already covers, and a card that cites the same release as a recent card from the same party is refused. This applies to every party in the same way.
One page for every policy, compared or not
The More policies page left out anything already compared on an issue page, so a reader looking at one topic, such as health or power, saw a party's other policies there but not its main ones. The owner asked for one page with everything. It is now called All policies, and each party's stated answer on every compared question appears under that question's topic, in the party's own column, marked Compared and linking to its full entry and to all seven answers. The answer is the same reviewed text the issue page shows, so it is not rewritten or checked again. A party with no stated position on a question has no card for it, as with any other topic. A switch hides everything marked Compared, for readers who only want what the issue pages leave out.
A promise tracker that treats both governments the same way
The site now checks what the last two governments promised against what they did: the Labour government of 2020–23 and the National–ACT–NZ First government of 2023–26. The hardest part is choosing which promises to check, because picking them by hand could make either side look better or worse. The parties' own headline lists didn't match: Labour's 2020 list was five broad aims, with no pledge card or 100-day plan, while the coalition had a pledge card, a 100-day plan and two coalition agreements. So both governments' promises come from the same kind of document, the Speech from the Throne each gave when Parliament opened, saying what it would do that term. Every statement in each speech where the government says it will do something is listed on the page, and those that name something checkable and due within the term count. The same number is then checked for each government, spread evenly through each speech by a fixed rule, so nobody chooses which ones. Both governments get the same six statuses. The two harshest, "not delivered" and "reversed", need official evidence and are checked a second time by a different AI model. There is no overall score, because promises differ hugely in size. The first review found mistakes before anything was published. The list of the coalition's commitments had missed some statements, including four on gang laws, and a few statements on each side were classed differently from similar ones on the other. Both lists were redone line by line from full copies of the speeches, which moved five of the twenty picks, and the cards follow the rule. Some cards quoted a sentence holding several commitments but judged only one, so a card now quotes just the commitment it checks. One card included an opposing party's criticism while others didn't, so arguments for or against a policy are now left out of every card. The second review found more statements sorted inconsistently: "keep this as it is" lines, such as the pension age staying at 65, counted for the coalition, while Labour's lines about carrying on named programmes didn't. Now any line that keeps something as it is counts as carrying on existing work, for both governments. Fixing the lists again would have moved eleven of the twenty picks, and each fix made after the cards were written gives a chance to move a pick knowing how it came out. So a pick now stays once it is made: each list records when its picks were drawn and from which positions, corrections to a list don't swap a picked promise for another, and later batches are drawn from the corrected lists. The third review found that steps towards a promise were rated differently: a bill that passed its first reading counted as not delivered, while funding and a tender for dental clinics counted as partly delivered. Now steps are shown as evidence but don't count as part of the promise, so a finished term with only steps is "not delivered" for every card. The date on each status is now the day it became true, such as the day a law took effect, rather than the date of the latest news.
Fairness rulings from the power prices issue
The power prices issue raised four questions our rules didn't answer, and the person who runs the site decided them. The Greens' plan to split the big power companies is in one of their MPs' member's bills, not their manifesto. It still counts as their position, but the entry now says where it comes from. The same applies to any party whose position is only in a member's bill. Parties that only promise stronger powers for the regulators are counted as changing the regulators. Government-guaranteed contracts for new power stations count as the government backing new supply. And a critic with a stake in the outcome, such as the head of a power company that would be split, can be the named critic as long as the entry says what the stake is.
Archived copies of sources
Before launch we found that only 3 of the site's sources had an archived copy. Party pages and news articles can change or disappear before polling day, which would leave a citation pointing at a page that no longer says what we cited. Sources now get a copy in the Internet Archive's Wayback Machine, linked as "archived copy" beside the original with the date the copy was taken. Where the Archive already had a copy taken no more than three days before we read the page, or since, we use that one; otherwise we ask it to save a new one. Either way the copy can be later than the day we read the page, which is why its date is shown. The Archive only lets us save a few pages an hour, so the first pass covered 46 of 567 sources, and the rest are being added over time.
Every entry's change history is now public
Our rules say each entry carries a public record of how it has changed, but until now that record was kept only in our own files. Each party's full entry on an issue, and each party page, now has a "Change history" list: every recorded change, newest first, with its sources. Most changes are our own corrections and rewording after checks, and the list says so, so a long history means an entry was checked and fixed more often, not that the party changed its policy. Because readers now see this text, it gets the same checks for characterising words and te reo Māori macrons as the rest of each entry.
What the "as at" date means
A check before launch found that an entry's "as at" date stayed at the day it was first researched, even after later rounds of checking changed it, so most entries looked older than they were. The date now moves to the day an entry was last checked against its sources and changed, for every party in the same way. For a party with no stated position, it stays the day its own pages and recent coverage were last searched, because that is what "checked" means there. The same check found that Te Pāti Māori's full policy documents on health, climate and energy had never been read, because they are hosted where our tools could not open them. They have now been read, and its health entry changed from "no position found" to its commitment to free primary care.
A page for each party, written to one template
Each party now has its own page with a short description, key facts and what the party says it stands for. So that no party is described more warmly or in more depth than another, every page follows the same template: the same facts, the same length limits, and the same four headings for the full background (history, time in government, place in Parliament, leadership). The values section is the party's own list, taken whole and in its own order from its constitution as lodged with the Electoral Commission, or its own website, and labelled as the party's words; the site doesn't pick which values to show or add left/right labels. These pages go through the same fact-check, swap test and three viewpoint reviewers as the issue entries. The first review found that the short descriptions named some of a party's values and left others out, gave reasons for some parties' MP departures but not others, and described how some coalitions ended but not others. So the short description now only says how many values the party lists, every page gives the dates and partners of each government without saying how it ended, and MP changes are listed with no reasons. The swap test also stopped hiding the word "National" inside names that aren't about that party, such as Te Pāti Māori's National Council. The second review found that some headings we gave to values the party hadn't headed read like slogans, so those headings are now plain topic names and the page says the site added them.
Bottom lines reported by journalists
For National, Te Pāti Māori and Labour, at least one bottom line rests on a news report saying the leader called it a bottom line, without quoting the leader using those words. The owner decided these count, because the reports are specific and dropping them would hide real coalition conditions, but each one is labelled on the page as the source's summary rather than a direct quote. The same rule applies to every party.
Policies outside the compared questions
The eight compared questions leave out most of what the parties have published: on 17 October Labour's policy page listed 27 policies, National's 33 and NZ First's 26, and only a handful of each were on the site. A new More policies page now holds the rest. Every entry on each party's own policy page gets one short card, plus a card for each new release, so the parties' own lists decide what goes in, not us. Every card uses the same template: what the party would do in plain words, its figures labelled as its claims, the source and the date. Some parties list single policies and others list broad areas, and each is carded in its own unit, so the page never counts or ranks parties by how many policies they have. Every party is listed in every topic, with "nothing on its policy page" shown where that is so. Cards are checked by the fact-checking reviewer, not the full panel, because they restate the party's own words without comparing them. The card writer for new releases now also files each card under a topic.
Parties' coalition bottom lines
Readers can now switch on a marker showing which parties have declared a bottom line on a question: something the party says it must get, or won't accept, to form or support a government. Only a statement by the party's leader, co-leader or official channels for the 2026 election counts, and it is quoted with its source. Where the source only summarises what was said, the site says so, because for several parties the words "bottom line" are a reporter's, not the leader's. Priorities, policies, and who a party will work with are left out for every party. Every party was searched the same way, and a party with none found is listed as such, with what was checked. Because they are quotes of party statements, bottom lines are checked by the fact-checking reviewer only, which also searches for any bottom line that was missed.
No party or answer is always shown first
A check before launch found two places where the order was fixed even though the site says party order is random. The KiwiSaver and home solar comparison tables always listed parties A to Z on wide screens, so ACT was always the first column. On issues split into smaller questions, the answers kept the order they were written in, so the same parties led most questions for every visitor. Both now follow the random order (or A to Z if you choose it), and each answer is equally likely to come first however many parties give it. The head-to-head page also said two parties took the "same approach" when they shared just one of several, for example on climate, where they disagree on most of the smaller questions. It now says "Some approaches shared" in that case and keeps "Same approaches" for parties whose approaches all match.
Fixes to the swap test, announcement cards and the hourly check
The same check found three faults in the automatic pipeline. The swap test hides party and politician names from the reviewer, but one Te Pāti Māori MP quoted on supermarkets was missing from its list of names, so that entry's party could be guessed; she has been added. When a reviewer sent an announcement card back for a rewrite, the first version could stay beside the rewritten one and be published without its own review, so four duplicate cards (three about the Greens, one about New Zealand First) reached the site's files. They were never shown on a page, and a rewrite now always replaces the first version. And GitHub, which runs the hourly check of party websites, skipped about three runs in four in the first days, so the site can now start a run itself whenever an hour passes without one.
Party names are shown in each party's colour
The rules used to allow party colours only as a small dot beside a name, never as a background. The owner asked for party colours to be used where they can, because they make it quicker to match a party across a page, and chose to show each party's name in a tag filled with its usual colour. Bright or dark colours, such as ACT's yellow or New Zealand First's black, stand out more than paler ones, so to keep the weight as even as it can be, every party's tag is the same size and shape, has the same thin edge, and uses whichever of black or white text is easier to read on it. The name is always written in the tag, so colour is never the only way to tell parties apart, and party colours never fill a whole card, block or row. The site's own colour is orange, which no party uses. The colours are the ones each party is usually shown in; two were less certain than the rest and will be checked against the parties' own materials.
Issue pages group parties by answer, and work on phones
The owner asked for a review of how each question is shown. The old grid gave every approach its own column, so most cells were empty, opposite answers such as staying in or leaving the Paris Agreement sat in separate columns, and on phones readers had to scroll through every party to see who agreed with whom. On phones the NZ Super age scale also pushed ACT and National off the edge of the screen, so only the parties keeping 65 were visible. Issue pages now split a question into smaller ones where the content defines them, group the parties that give the same answer, and name every other party under each one, so no party drops out of view. Where parties would move one setting in opposite directions, such as strengthening or repealing the 2050 climate targets, the page can show a strip of which way each would move it from current law, in one colour and without showing how far, so neither way reads as better. A "By party" view shows each party's whole answer under the same headings. The note on uneven evidence still appears above every comparison, now as one line that opens to the full text.
Government records, and points that apply to every party
Reviewers asked two things the rules didn't settle. First, some of National's entries are built from what its ministers did in government rather than from a new election policy, while its coalition partners get no position from the same decisions. The rule now says that a government decision counts as a party's position when its own ministers made it and it hasn't said it would change it, and the entry says so; coalition partners get the decision as background only. Second, on GP costs, a medical journal editor's point that free visits need enough doctors, and more doctors need affordable visits, appeared only in Labour's entry, though it was a general point in his review of every party's plans. The rules already said that context applying to several parties goes to all of them in the same words; the owner confirmed this covers an expert's general point, and it is now in every party's entry.
A range, with a note, when a party commits to a change without full detail
The calculator first showed NZ First's income tax plan as "can't calculate", because NZ First has said it would make the first $14,000 of income tax-free but hasn't published its other rates. The owner decided that a plan a party has committed to should still be shown. NZ First's result is now a range, from making $14,000 tax-free to making the whole current lowest band tax-free, assuming its other rates stay as they are today. A note beside the result says this is the site's assumption, not a published NZ First table. Any party in the same position will be treated the same way.
A calculator that works out each plan for your own money
A new page, "How would this affect me?", works out how each party's plans on income tax, KiwiSaver and GP and prescription costs would change a reader's own money. Every party's plan goes through the same calculation with its own published numbers, against today's law. Where a party hasn't published enough detail, the page says it can't calculate rather than guessing. A total is shown only when every part of a plan can be counted for that reader, because a part left out could make a plan look better or worse than it is. There is no total for each party across questions. The calculation runs in the reader's browser and nothing they enter is sent or stored. The checks now make sure every party has a calculator entry on each question it covers, and that it matches the party's stated position.
One set of writing rules checked across all seven parties
In the home solar reviews, the swap test kept finding the same kinds of unevenness, each time in a different party: a party's biggest tradeoff measured against an official finding when others' weren't, a critic quoted against one party who was talking about another party's plan, or one party's gap repeated in several places. Fixing one entry at a time just moved the problem. Every issue is now checked against one set of rules, with all seven entries read side by side: the biggest tradeoff comes from the party's own design, a general critic is labelled as one, a party's statements are kept apart from what it hasn't said, context that applies to several parties goes to all of them in the same words, and each gap is stated once.
Parties listed under an approach only for their own election policy
For home solar, ACT had been listed under "plug-in solar and fewer rules" because its leader, as a minister, ordered the government review that led to legalising plug-in solar. That change is a government decision and isn't in ACT's election policy, and the reviewers kept disagreeing over whether ACT should get credit for it. The owner decided that a party is listed under an approach only when its own election policy proposes it. The review is still described in ACT's entry, as a government decision.
A limit on review rounds for each new issue
Reviewing the first of the new issues, home solar, took four rounds. The problems found fell from 25 to 13, but the reviewers then started asking to undo each other's fixes: one asked to put back figures that the swap test had asked to remove. Rewriting forever wouldn't make the entry fairer, so each issue's first version now gets up to two more rounds of fixes and review. It is published once the verifier passes, so every fact must check out, and any fairness finding still open gets a written answer instead of another rewrite. Updates made later by the automatic pipeline still need all five reviewers to pass.
A suggestion box instead of sending parties their entries
The original plan was to send each party its entries to check before launch. Instead, the site now has a suggestion box that parties, and anyone else, can use at any time. Parties can respond for themselves. An agent checks each suggestion against the entry and its sources, and nothing changes until the person who runs the site approves it. A party's response is still treated as a party claim.
Uneven-evidence notes now shown on every issue page
When only some parties' plans have been independently costed, the issue page is meant to say so. We found that for the seven newest issues these notes had been written into a section the site didn't display yet. They now appear under each issue's introduction at every depth.
The swap test now hides politicians' names and web addresses
The swap test hid party names but left in politicians' names, so a reviewer could still work out the party from a leader's name. Web addresses in links could name the party too. Both are now hidden, and an automated test flags any politician named in new content who isn't yet on the list to hide.
A leak in the swap test's name hiding fixed
Testing found that long lines of text were wrapped before party names were hidden, so a two-word name split across a line break could slip through and show the reviewer which party it was checking. Text is no longer wrapped before names are hidden.
A cheaper swap test tried and rejected
To save effort we tried running the swap test once per issue, with all seven entries in one run, instead of once for each party. It used about three-quarters less, but it found 7 to 10 problems where the one-per-party version found 22 or 23, and it passed entries the full version rightly failed, including campaign terms used in the site's own voice. The swap test stays one run per party.
Reviews that hit the usage limit now wait instead of failing
All the agents share one Claude plan, which has a usage limit. When the first full review of the new issues ran, about a third of the review runs stopped part way when the limit was reached. Those counted as failed checks, so the whole review had to run again. Now an unfinished check waits and runs again later, and it never counts as a pass unless it actually ran. New issues are also reviewed one at a time.
Reading party websites that block automated readers
On its first day the watcher found that four parties' websites block automated readers. Rather than disguise it as a person using a browser, it now reads those parties' own news feeds and site maps. Every party is still watched for the same two things, its news and its policies, though how each is read can differ. One side effect: one party's site map lists its policy pages individually, so edits to them are noticed in more detail. The government's press release site also blocks automated readers, so it was dropped; the Scoop politics feed still carries government releases.
First fairness rulings, from the supermarket issue
The first issue raised three questions the rules didn't answer, and the person who runs the site decided them. When only one party's plan has an independent cost-benefit analysis, the page says so. When nobody has publicly criticised a particular policy, the case against says that and uses the nearest named criticism. And a party figure that can't be found in the report it cites stays labelled as the party's claim while we keep tracing it. These now apply to every issue.
Starting with one AI model for every role
The plan calls for reviewers to run on a different AI model from the writer, so they don't share its blind spots. To keep a spare-time project affordable, we started with Claude for every role, using separate runs with different instructions. Adding reviewers from a second model is still planned.