How we source and rate funds

Primary sources

• FMA Disclose register — the official New Zealand register of licensed managed investment schemes. Source of fund existence, manager identity, PDS documents, fees and returns disclosures.
• Fund manager websites — unit prices, fact sheets and supplemental performance data, fetched weekly.
• IRD — PIE tax rules, PIR thresholds and tax treatment guidance.
• Sorted Smart Investor — used as a cross-validation anchor only; not a primary source.

What we do not do

• We do not publish a "best fund" list, a star rating, or any subjective ranking.
• We do not accept paid placement or sponsored fund profiles.
• We are not a licensed financial adviser. We do not provide personalised advice.

Data freshness

• Fund register and fees: refreshed daily from FMA Disclose.
• Returns: refreshed monthly when manager fact sheets are published.
• PDS documents: re-fetched weekly; if the version differs from the last copy, we record a diff.
• Every fact-bearing data point on the site shows its source URL and the date we observed it.

Confidence tiers

Every sourced data point is tagged A, B, C or D:

• A — direct from FMA Disclose or the fund manager's signed PDS.
• B — fund manager website (fact sheet, performance page).
• C — third-party aggregator (Sorted, Morningstar) used for cross-checking.
• D — derived or estimated. Always shown explicitly.

Document provenance & file fingerprints

For the funds we crawl directly, every PDS, SIPO, OMI and Quarterly Fund Update on a fund page links to the file on the manager's own website. We don't host copies of the document — we link straight to the source.

When we fetch a document, we record:

• Source URL — the manager's own document page.
• Fetch timestamp — what we display as "last verified" on each fund page.
• File size in kilobytes.
• SHA-256 fingerprint — a 64-character hash of the exact bytes we received. The "file fingerprint recorded" label on each document means this hash is logged in our database.

If a manager re-issues a document, the new file produces a different SHA-256 fingerprint. That's how we detect document changes without holding our own copy of the file. The change history feeds into our PDS-version timeline.

PDS / SIPO / OMI extraction pipeline

For every NZ retail managed fund on the site, we extract a structured set of facts from the fund's current Product Disclosure Statement, Statement of Investment Policy and Objectives, and Other Material Information document — all lodged on the FMA Disclose register. Coverage as of last build:

• 247 structured extracts across 125 PDS + 63 SIPO + 59 OMI documents.
• 257 of 289 funds have on-file PDS facts (objective, performance-fee terms, buy/sell spread, minimum investment, distribution frequency, hedging policy, PIE status).
• 261 of 289 funds have on-file SIPO facts (strategic asset allocation target/min/max ranges, ESG approach, derivatives policy, exclusions).
• 266 of 289 funds have on-file OMI facts (licensed supervisor, auditor, custodian, related-party transactions, conflicts of interest, scheme-level performance-fee mechanics).
• Every extracted entry carries a Tier-A SourceBadge linking to the source PDF + the date the document was lodged. The full manifest is browsable at /sources/.

How extraction works. Each PDF is processed with a two-stage pipeline:

1. pdftotext first — fast, free, deterministic. Used when the PDF has a clean text layer.
2. Claude Sonnet 4.6 PDF endpoint as fallback when pdftotext returns <500 chars or graphics-only glyph noise. Sonnet receives the raw PDF bytes and a JSON-schema prompt; output is parsed and re-checked.

Both paths feed into the same JSON schema with explicit provenance metadata (pdf_hash, source_doc_date, source_tier A/B/C/D, extraction_method, cost_usd).

The 36-month freshness gate. No document older than 36 months is extracted or cited — this is a cross-Pangaea data-freshness rule. The fund manager is required by the FMC Act to refresh the PDS on material change; a 36mo-old document is no longer current.

Verbatim source paragraphs. Each extracted field carries a source_paragraph citation — a 1–3 sentence verbatim quote from the underlying PDF, capped at 350 chars. This is the audit trail: any extracted fact can be traced back to the exact wording in the source document.

Hash-versioned diff trail. Every extract is provenance-tagged with the SHA-256 hash of the source PDF (first 12 chars shown publicly). When a manager re-issues a PDS, the new file produces a different hash; our next ingestion picks up the change and the on-page SourceBadge shows the new date. The same hash is the foundation of the planned PDS-diff engine (re-run extraction quarterly, surface field-level diffs as a paid Tier-2 SKU).

Edge cases. 2 fund schemes failed extraction due to malformed PDF metadata (The Curve missing CreationDate / Mint corrupt XRef stream); these will be re-extracted once the source-PDF issues are fixed by the publisher. Several historical PDFs sit past the 36-month freshness gate (Booster supplements from 2016–2021, Smartshares ETF supplements from 2014–2020) — those funds carry only the most recent valid PDS we could extract, or fall back to Sorted Smart Investor data with a Tier-C SourceBadge.

AI output policy

The PDS extraction pipeline uses an LLM (Claude Sonnet 4.6) to parse free-text disclosure documents into structured fields. Three guardrails apply:

• Verbatim grounding. The extraction prompt requires every field to be supported by a verbatim source paragraph from the PDS. The model is explicitly instructed not to infer or fabricate.
• FMA-safe output wrapping. Our MCP server (the LLM-accessible API surface) applies a four-piece sanitiser to every tool response: prohibited-terms strip ("best", "guaranteed", "leading", "recommend"), KiwiSaver® trademark rephrase per IRD 2021 ruleset, a mandatory FMA disclosure preamble, and an advice-handoff trigger that flags any tool input containing personal-advice phrasing ("should I", "right for me") and prepends a referral notice toward a licensed Financial Advice Provider.
• Human-curated outside the extraction layer. Editorial pages, guides, and methodology text are human-written. The <meta name="ai-content-declaration"> tag on every page declares "human-curated, mechanical-extracts" — accurate to both halves of the content.

Corrections

If you find an error, email hello@managedfunds.nz. We correct verified errors within one working day and log the change.