Data & API

Every filed value carries a cross-check status and the EDGAR accession behind it; values that fail plausibility screens are quarantined before they publish.

Weekly pipeline run; a fund's figures move when it files, and derived keys are recomputed from the canonical facts on every run.

• 90,457 published filed facts as of 2026-07-28: 23,219 (25.7%) are cross-checked across two SEC endpoints, 67,238 (74.3%) are single-source — not wrong, unchecked — and 0 conflict. The status column carries that label on every row.
• Cross-checking measures agreement between SEC's own views of the same filing, not independent confirmation of what the fund filed.
• 631 of 645 published funds carry at least one figure as of 2026-07-28; 14 carry none.
• 47 of 113,723 raw facts were quarantined by plausibility screens as of 2026-07-28 and never published; quarantine counts appear in the data_quality dataset.
• Derived keys are computed, not filed: every derived row carries primary_source = DERIVED, no accession and no filing date — its provenance is the published derivation formula, not a filing. Derived keys exist only where their filed inputs do, so coverage inherits every gap in the canonical facts.
• A ratio whose denominator makes it meaningless — leverage against negative ramp-up equity, a fee rate against a sliver of reported net assets — is suppressed, never published as a number that merely looks like one.
• Filed always wins the one key both layers can emit: the derived nav_per_share is a fallback present only where no filed value exists for the exact grain, so metric_history never carries a filed and a derived value for the same key (0 collisions measured 2026-07-28).
• canonical_value is an exact decimal, not a float: numeric(38,12) in the warehouse, Decimal128(38,12) in the Parquet exports, and a quoted decimal string in JSON. A filed 0.145 is 0.145 — never the nearest double to it — and it stays that way through every hop of the delivery.

Weekly run; N-PORT contributes monthly period grain, BDC Schedules of Investments follow annual and quarterly filings.

• 516 of 645 published funds have a book loaded as of 2026-07-28 — loaded is not complete.
• 192 of 516 loaded books do not tie to the fund's own reported total in both directions, and 8 of 516 report no total to check against. The verdict ships on every row as recon_status / recon_ratio.
• N-PORT books are stored top-100 truncated (is_partial); weights on truncated books use the filer's own reported percentage and are never normalised against a total we do not hold.
• 81.3% of positions (53,797 of 66,142 as of 2026-07-28) resolve to an issuer by normalised name only — the weakest join and the majority one, with measured two-way error of ~10.8% over-merge / ~9.9% under-merge against CUSIP-keyed ground truth. match_method ships per row so name-grade joins can be filtered out.
• position_seq is a deterministic construction, not a filed identifier: filers repeat position rows verbatim, so the position exports number each book's rows by a total order of every column. Fully-identical rows may swap seq between rebuilds — they are identical, so nothing about them changes.

Rebuilt weekly from the loaded books.

• Issuer identity inherits the holdings dataset's name-match error; every issuer-level figure carries it.
• Filed industry exists on 3,886 of 63,363 BDC Schedule-of-Investments positions (6.1%) and 0 of 20,466 N-PORT positions (0.0%) as of 2026-07-28 — the N-PORT form has no industry field. Dollar-weighted it is thinner still: $27B of $594B (4.5%). At the issuer grain the two filed bases together label 6,176 of 44,100 rows (position-carried 2,307, name-feed 3,869) and 3,349 of 25,673 issuers, holding 20.5% of aggregate fair value — measured 2026-07-28.
• Industry coverage can never approach 100%: private borrowers inside a BDC book do not file on their own, so there is no SIC/NAICS fallback. The column ships where filed — position-carried, or matched from the filed name→industry feed with industry_source saying which — and is otherwise null: never inferred, never rendered as zero.

Every fund-to-manager link names the evidence that produced it, so a filed link and a guess never ship under the same label.

Rebuilt weekly; adviser filings move on annual cycles, so most weeks are no-ops.

• Brand curation is a maintained mapping, not a filed fact; manager_link_source and name_source are the honesty mechanism and are part of the schema, not an apology.
• BDCs file no N-CEN adviser section, so their adviser fields come from each BDC's own 10-K matched exactly against SEC's IAPD registry, and are null where no exact match exists.

Recomputed weekly over the whole universe's current state.

• Crowding is key-sensitive: the index moves 46.9% to 79.6% purely on how issuer identity is resolved. Read crowded_share as a band anchor, not a point fact.
• denominator_is_truncated_holdings flags funds graded against a truncated book (no reported total_assets to use as the denominator).
• Percentiles exist only where the peer set clears the configured minimum; peer_count ships on every row so the denominator is never implied.
• Latest views only: these tables are rebuilt in place each week and carry no history. A weekly snapshot replaces the last, and deltas between snapshots are not meaningful row-for-row because rebuilt tables have no stable row identity.
• The relationship-graph edge list (mart_graph_edges) was removed from the exports 2026-07-28: it is a presentation artifact — its weight floors and caps are display judgement, not data — and everything it encodes derives from the holdings and manager_registry datasets.

Weekly pipeline run, alongside the holdings the marks are computed from.

• mark_cost = fair value / amortized cost AS FILED, which includes OID/PIK accretion.
• is_debt_proxy is a principal-present proxy, not a filed flag: 82% precision on the labeled subset as of 2026-07-28 — a boolean, never a credit opinion.
• Fund aggregates in fund_marks are computed over debt-proxy marks inside the trim window [0.01, 2] only; positions_counted, positions_trimmed and positions_excluded partition the fund-period's whole book, so a position dropped from an aggregate is never dropped from the count.
• Seniority is classified label-first with a raw-identifier-text fallback, both routes through the same strict pattern seed; seniority_basis / fv_seniority_from_text record which route classified each row / dollar. Filed labels alone cover 17.7% of fair value and the fallback lifts classification to 64.9% (2026-07-28); the four seniority shares remain of STATED fair value (fv_seniority_stated / seniority_stated_share on every fund_marks row), and FIRST_LIEN still requires explicit lien wording from either route, so the published shares remain floors, not full counts.
• mark_spread in issuer_mark_dispersion is a range across reported debt positions in the same borrower — same borrower is not same loan. Instrument labels exist on ~17% of positions as of 2026-07-28, so tranche matching is not possible at that coverage; instrument_labeled_count / instrument_mix_known ship on every row so a reader can see how much of the range could even in principle be explained by tranche.
• position_seq is a deterministic construction, not a filed identifier (see the holdings dataset caveat): filers repeat position rows verbatim, so the position export numbers each book's rows by a total order of every column.
• nonAccrual is NOT included in this dataset because it is not genuinely extracted from the source filings.

Non-traded funds have no exchange quote, so flows are the demand signal a quoted share would otherwise give.

Weekly pipeline run; N-PORT flows arrive on the monthly filing grain, ~60 days after period end.

• REGISTERED funds only (interval funds, tender-offer funds, listed closed-ends): BDCs do not file N-PORT, so a BDC's tender-offer repurchases are absent — a filing-regime boundary, not evidence that nobody redeemed.
• redemption_rate_1m divides by the most recent net_assets strictly BEFORE the flow period — the balance investors actually redeemed out of, not the end-of-period remainder — and is deliberately not annualised: flows are lumpy and tranche-based, and a one-off liquidation month multiplied by twelve implies a recurrence that isn't there.
• A measured zero-redemption month is a 0% rate, not a gap; the rate is null only where the derived layer could not rate the month (no prior net_assets on file). 101 of 344 flow-filing funds have a rated latest month as of 2026-07-28; all 344 keep their row — flows render even where the rate is unmeasurable.
• The 3-month figure is a SUM of monthly rates, each already measured against its own prior NAV — no new denominator is chosen; months_in_3m / months_rated_3m state how much of the window is observed and how much is rated, so a 1-of-3-rated sum can't pose as three months of measurement.
• The band edges (elevated_band / high_band) ride along as columns so the stated meaning is always the real configured value.

Every other dataset here is rebuilt from the current filings each week, so it always describes today; this one is the only record of what any of these numbers WERE. Each row carries its own numerator, denominator, fund count, the configured band edges in force, and the methodology marker under which it was computed.

One row per metric per build. The weekly pipeline run appends a new date; a same-day rebuild replaces that date's rows rather than adding a second reading.

• AS COMPUTED AT THE TIME, not a current best estimate of a past date. A row states what was published on as_of_date under the methodology then in force. Rows are never revised: nothing in this dataset is ever updated in place.
• A METHODOLOGY CHANGE APPEARS AS A NEW ROW, NOT A RESTATED ONE. These aggregates move when our methods improve — better issuer identity resolution shifts the crowding index with no filing having changed — so a series can step for reasons that are ours rather than the market's. methodology_version is bumped by hand when a definition changes and code_rev records the git revision that built the row, so a data change and a method change are distinguishable; comparing two rows with different methodology_version values compares two definitions.
• HISTORY STARTS ON THE FIRST BUILD (2026-07-29) AND NOTHING BEFORE IT WAS SYNTHESISED. No mart stored these figures previously, so no earlier point is recoverable; the series is short by construction and will only lengthen forward.
• A same-day rebuild REPLACES that date's rows — the last build of a date is what stands for it, the way an index publishes one value per date. Intra-day drafts are not retained; computed_at says which build survived.
• Shares ship with their denominators because a share whose denominator moved is a different measurement — a crowding index that fell because two funds stopped filing is not the same event as one that fell because books diverged. numerator, denominator and funds_counted are on every row that has them.
• Band edges ride on the row in config_edges, read from the serving relations' own stamped values rather than restated. A band change moves the number, and two adjacent rows differing only in an edge explain themselves.
• Every issuer-level measure inherits the identity-resolution error the data_quality dataset measures: the normalised-name join both over-merges (inflating crowding and overlap) and under-merges (deflating them). Dollar figures are footprints of what is filed and visible, summed across funds whose fiscal periods are different dates — never AUM totals and never same-date aggregates.
• numerator and denominator are exact decimals wherever their inputs are, with one stated exception: the crowding index numerator weights a presentation-grade share by exact dollars and therefore carries that share's precision. Each row's own caveat column names the limits that apply to it.
• The largest-platform row identifies its platform in scope, and the largest platform on one date need not be the largest on the next: that series is only one platform's history where scope is equal across the dates compared.

The same tables we grade ourselves with.

Recomputed weekly alongside everything it grades.

• Every measure carries its own numerator, denominator and a caveat sentence stating what it cannot prove. A coverage figure quoted without its denominator is misquoted.
• is_asserted marks sources with no configured freshness band — an invisible exemption reads as a passing check, so it is surfaced as data rather than hidden.