# SEC Fund Prospectus → RDF Triple Dataset — Design

## Goal
A training/eval dataset for text→RDF-triple extraction where:
- the INPUT text is much larger than the OUTPUT JSON (realistic "needle in long doc"),
- the OUTPUT is a genuine **graph of entity→entity relationships**, not flat key→literal attributes,
- it is finance **reference data** (openfunds-aligned),
- there is a **non-model gold baseline** AND a path to a **strong-model baseline**.

## Why not flat XBRL fee data
XBRL Risk/Return (management fee, TER, returns) yields only `Fund --pred--> literal`
attributes — a star of literals, no entity-to-entity edges. Rejected as the *primary*
graph source because it produces no relational structure. (May be added as optional
literal-valued triples to enrich the schema, but it is not the point.)

## Evidence (sample: Fidelity Oxford Street Trust, 485BPOS 0000028540-25-000048)
Full prospectus prose = 1,061,532 chars. Relational mentions found in prose:
adviser 381, sub-adviser 25, distributor 31, custodian 59, transfer agent 24,
"Trust" 805, "series of" 13, "managed by" 38, named benchmark index.
=> The prose contains a real multi-entity-type graph.

## Target ontology (entity→entity edges)
```
Fund      seriesOf       Trust
Fund      advisedBy      InvestmentAdviser
Fund      subAdvisedBy   SubAdviser
Fund      distributedBy  Distributor
Fund      custodian      Custodian
Fund      transferAgent  TransferAgent
Fund      managedBy      PortfolioManager
Fund      tracksIndex    Index            (passive/index funds)
```
Optional literal-valued triples (attributes) to round out the openfunds record:
managementFee, netExpenseRatio, return1yr/5yr/10yr, portfolioTurnover, objectiveText.

## Baselines (the key selling point)
1. GOLD, no model:
   - N-CEN service-provider table -> advisedBy, custodian, transferAgent, distributor, auditor (with LEIs)
   - Series/Class CSV            -> seriesOf, hasShareClass (Trust->Series->Class skeleton)
2. SILVER, strong model (GPT-4 / Opus), measured against the N-CEN gold:
   - managedBy (portfolio managers), subAdvisedBy, tracksIndex (named only in prose)

## Unit of extraction
One FUND (series) = one sample:
  input  = the prospectus section(s) for that fund (kept within a long context)
  output = the fund's subgraph of triples
N-CEN/Series-Class gold is keyed at trust/series level, so per-series is the natural join.

## Data gaps to fix (local dump)
- fund_data.db was accidentally deleted (only stale -shm/-wal remain) -> rebuild from data/.
- load_xbrl_rr.py tag constants are partly stale vs the actual num.tsv vintage
  (e.g. ExpensesOverAssets, DistributionAndService12b1FeesOverAssets,
   AvgAnnlRtrPct/AnnlRtrPct, PortfolioTurnoverRate, BarChart*QuarterlyReturn).
- Local N-PORT dump is MISSING FUND_REPORTED_HOLDING.tsv and IDENTIFIERS.tsv
  (the holds/issuedBy edges) -> would need re-download if holdings edges are wanted.
- XBRL flat files contain NO text blocks -> narrative input must be fetched from EDGAR
  filings (filing_text), not from the flat files.
- N-CEN flat files are NOT yet downloaded locally -> needed for the gold edges.

## Holdings edges (deferred to a 2nd track)
Holdings (`fund holds security`, `security issuedBy issuer`) are NOT in the prospectus.
Text-bearing sources for holdings:
- Annual/Semi-Annual Report (N-CSR): "Management Discussion" (MDFP) names TOP holdings
  in prose -> real `holds` edges; gold = N-PORT holdings table. The full Schedule of
  Investments is a table, not prose.
- Fund fact sheets / PM commentary (fund-company marketing): "Top 10 Holdings" + prose,
  but off-EDGAR and no standardized machine-readable gold.
Plan: build prospectus->service-provider graph FIRST (clean N-CEN gold). Add an
MDFP(N-CSR)->holdings dataset as a SECOND finance sub-domain later — this also
strengthens the thesis's cross-domain generalization claim.

## This session: PROOF OF CONCEPT (~20-50 funds)
End-to-end on a small slice: fetch prospectus text, get N-CEN gold service-provider
edges + Series/Class structure, emit text->triples samples in the <triple_start>
RDF marker format, and a gold-baseline scorer. Validate before scaling.