Two People, 1,574 Candidates: Building an Independent Election Bulletin
How a two-person team published source-cited fact sheets for every candidate in Taiwan's 2026 local elections: decide what you can claim, steer agents with domain knowledge, ship per card, let people settle what models can't, and hold the design to one system.
Taiwan's official election bulletin gives each candidate a photo, a short biography, and a policy statement written by the candidate. It says nothing about their wealth, who funded their last campaign, or whether they have a criminal record, even though much of that is already public as government open data, scattered across agencies, formats, and scanned PDFs.
This year a colleague and I built an election bulletin that puts those records next to the official one: Gazette 2026. It covers all 1,574 candidates for mayor and council across 243 districts in the 2026 local elections. In the two weeks before election day it reached about 117,000 users and 2.1 million page views, with most visitors spending two to three minutes reading.
There were two of us. I was the product owner and lead developer; my colleague is a political-domain expert who contributed through pull requests. Five decisions are what made our collaboration work.
1. Decide what you are allowed to claim before you collect
The first decision was how much the page would assert. One option was to state findings about the candidate: "declared NT$48 million in land," "convicted of vote buying." The other was a dossier: list every record that matches the candidate's name and leave readers to judge whether each one is really about them. The dossier is safer for the publisher and close to useless for a voter, who has no way to tell whether a court ruling that mentions a name is about the person on the ballot. I chose to assert, and assumed the bottleneck was collection: crawl long enough and the facts will come.
A test against court judgments proved that wrong. Taiwan's judiciary masks the names of individual defendants, so a convicted person appears as 辰○○, while witnesses, victims, and people merely mentioned keep their full names. When I searched one real candidate's name, all five hits were as a counterparty, a victim, a witness, or in passing. None was a conviction. A broader name match returned 89 judgments for a single candidate, almost none about that candidate. Listing them under the candidate's name with a disclaimer would still be defamation by volume.
Crawling longer cannot fix that. Whether we found a judgment and whether its defendant is this person are two different bottlenecks. So the first architecture decision, recorded as an ADR, was to assert claim by claim, tiered by how certain each record's identity is. Records filed under real names with structured keys, such as asset declarations and political donations, can be stated as facts. Records with only a name need a second key. Masked records go in with their uncertainty shown, never as a statement about the candidate. Each claim on the site carries its own source and type, so review can approve or withhold one claim without holding back a whole page.
2. Steer agents with domain knowledge, not just guardrails
Asset declarations look like the easy part: official forms, filed by the candidates themselves. In practice they are scanned, often handwritten, and inconsistent in ways no schema anticipates. One line OCR'd as 3,200,005,763 says 320,000 on the original. Filers write "15,000,000 萬" (ten thousand times larger than they meant), use a decimal point as a thousands separator, enter the per-square-metre price where the total belongs, or list the same jointly bought house once per owner.
Each of these is an editorial decision, not a parsing bug. My colleague made most of them; my part was spotting failure patterns in the output, having agents correct them, and reworking how we value real estate from actual transaction records when a filing gives no price. The principle we settled on was use only the numbers the filer wrote: outside data such as market prices or land registries can be cited in a note, but it can never change a number. Where a figure is literally absurd, the site shows it as filed, flags it as a probable filing error, and leaves it out of the totals. By 10 November the rulebook had 64 numbered rulings.
Guardrails were not the hard part; every agent workflow has them now. The hard part was that a capable agent, left to its own judgment, makes calls that are reasonable in general and wrong for this domain. It would "fix" a suspiciously high land value against the market price, which quietly replaces what a politician declared with what we think they should have declared. It would add up two spouses' identical entries for the same flat and double a household's wealth. The rulings are how domain knowledge reached the agents: each new pattern came with an original case, the conditions it applies to, a counterexample, and a regression test, and all declaration data passes through one publication function, so no page could compute its own totals. That is what kept the agents' work in line with how people who know politics read these filings.
3. Ship per card, not per site
Through the first month the site was built as a whole: every source sealed, every card validated, then publish. That was the right order while we were still establishing what the data could support. By September the data pipeline had its own strict, whole-batch checks, and the priority shifted to iterating fast enough for an election calendar. There was no longer a reason for every candidate page to wait on the slowest data source.
So on 10 September I split the two. Data production keeps its strict checks. The website build validates each card on its own: identity, source, period, numbers, and public format. A card that fails is hidden and logged to a public data-status page; the rest of the site builds. We launched with identity cards for every candidate and added asset, donation, and criminal-record cards as each one passed, without loosening what "passed" meant.
It also forced a distinction the interface now keeps: not built (we have not connected this source), no data (we checked and found nothing), and unavailable (we have a reason not to answer). An empty criminal-record card must never be read as "no criminal record."
4. Let people settle what models can't read
OCR, checked against two models reading each page independently, resolved most of the declarations. My colleague built and ran the rest: a volunteer site that sends the remaining cells to people one at a time and accepts an answer only when enough of them agree. By late October volunteers had reached consensus on about 11,200 of roughly 13,400 cells, and where they disagreed with the models, the volunteers won.
5. Hold the design to one system
Agents are as prone to design drift as they are to editorial drift. Each session produces something locally sensible; a few dozen sessions later the site has five button styles and three ways to show an empty card. In late September and early October I rebuilt the design system as a single written spec with shared components and a live showcase page: one type scale (display, name, title, lead, body, caption), one set of controls, 44-pixel touch targets, and layouts that must survive a 320-pixel screen and 200% text.
What changed was where design work happened. From then on, a design decision was a revision to the spec, dated and numbered, and every page picked it up through the shared components. By November the spec carried more than thirty such revisions, from navigation order to where the source and share buttons sit on a card, and the site stayed visually consistent across hundreds of agent-written changes in the final weeks.
What the numbers say
Besides the landing page, readers' favorite was the asset ranking that lets them sort mayoral and council candidates by declared wealth (187,000 views; one in five visitors used its filters). Readers came for comparison more than for single profiles. Council candidates drew slightly more ranking users than mayors, which makes sense: they are the races where voters know the least.
Visitors who arrived from search engines stayed longest, about four and a half minutes against two to three for everyone else. In this context that is a signal about the readers rather than the site: someone who types a politician's name into a search box already cares more about politics than the average visitor. And almost no one used the share button, which is what we expected: people share by screenshotting a card and posting it to Threads.
The broader lesson is the one I keep relearning with agents. Two people can now produce what used to take a newsroom, but only if what the people know is written down where the agents must follow it. In this project that was the claim tiers, my colleague's rulings, and the design spec; together they are how two people's judgment reached all 1,574 pages.