Data collection order

1. The brand's own site first. Product pages, pricing pages, warranty/terms pages, and FAQs on the brand's own domain are the default source for every fact.
2. An authorised Singapore retailer, flagged. When a brand doesn't publish a figure itself (common for pricing — several brands render prices client-side or sell purchase-only via a sales consultation), an authorised SG retailer's listing is used instead, and the model's data notes say so explicitly.
3. Never a third country's site. A brand's regional site (e.g. a Malaysia storefront) is not used for Singapore-market facts, even when it has more detail — those fields are left null instead.
4. Never AI-summarised or unverifiable web claims. If a claim can't be traced to a specific page checked directly, it isn't recorded.

Nullability policy — “a null beats a guess”

Every identity and structural field (slug, brand, name, status, product URL, website) and every source block is required. Every other fact — dimensions, warranty length, filtration technology, a filter's price or replacement interval — may be null when the brand genuinely hasn't published it.

A null is never filled in with a plausible-sounding number, and it's never rendered as 0 or left blank on the page — it always shows as an explicit “Not published”, usually with the specific reason from that model's data notes. Real brands genuinely withhold real facts (several sell filters only through a service contract, not itemised; several don't publish warranty length at all) — this dataset records that honestly rather than hiding it.

The cost-of-ownership formula

Computed, never hand-entered. Exactly:

Prices are the brand's standard published price as last verified — GST-inclusive where the brand states so, flagged in notes where it's unclear (see, for example, Wells' own pricing metadata, which explicitly marks GST as excluded). Time-limited promotions are ignored in favour of the standing list price; a prominent, long-running promo is still noted, not used as the recorded figure. Consumable replacement is scheduled on the brand's stated interval, not actual household usage.

The incomplete-TCO rule

If any filter is missing a published price or replacement interval, that filter is excluded from the subtotal and the whole total is flagged incomplete — marked with a † wherever it appears, linking to the model's cost breakdown, which lists exactly what's missing.

An incomplete total is always a floor, never an overcount — the real cost is at least that much. It is never presented as a like-for-like number against a model with a fully published cost schedule. A model with no purchase price published at all (or no filters and no servicing published) simply doesn't get a computed total — it renders as “Not published”, not S$0.

Re-verification windows and update cadence

A staleness checker (npm run stale) lists every sourced fact past its window:

• 100 days — pricing, filter prices, and rental terms. These change often.
• 365 days — physical specs, warranty/terms, and certifications. These change rarely.

Every entry on that worklist is re-checked on a quarterly cadence: re-verify against the source, append a dated price observation rather than overwrite the old one, and bump verified_at. Prices are append-only observations, so a model's price history is never lost to an edit.

Dimension leaders, never an overall score

This site never publishes a single ranking of “best” purifier. Facts and analysis are kept visibly separate: specs and prices are sourced facts; anything framed as a verdict is labelled as analysis from stated criteria. Where a superlative is shown (cheapest over 5 years, smallest footprint), it is one reproducible comparison over one field, restricted to models that actually have that field published — never a weighted composite score. See /about for the ownership disclosure this stance depends on.