FinUties Study · Filings · A metadata index, not a document store

Updated at 25 August 2026.

FinUties Study · Filings · A metadata index, not a document store

By FinUties · organization byline · The organization reports no position in any instrument named. Informational only — not investment advice, and no claim is made about domains outside the P0 surface.

Scope

P0 domain: filings. The question in one sentence: what does a default pull from
/api/v1/filings/submissions actually contain, and which jobs can be built on it today?
This study characterises the surface — form-type spread, filer spread, field completeness. It
does not interpret any individual filing, does not rank issuers, and does not treat any
non-P0 domain as covered.

The reading it arrives at: FinUties filings is a submission metadata index, not a document
store. It is fresh inside its 24-hour budget and broad across filers, yet every
document-body field in the sample is empty, and the surface is fed by two ingestion paths
whose completeness is inverted.

Question

This is answerable from P0 data because completeness is a property of the returned rows
themselves: null rates, distinct-value counts and label distributions need no external
reference to compute. It would become unanswerable if the question shifted to filing
content — nothing in this pull carries document text, so any claim about what filings say
would rest on data FinUties has not fetched.

Evidence manifest

This study rests on 3 pulls from the filings domain between 19:43 and 19:52 UTC on
2026-08-18. Every row below is re-runnable.

#	Pulled (UTC)	Domain	Endpoint	Query	Rows	Freshness	Replay
1	2026-08-18 19:47	filings	/api/v1/filings/submissions	limit=100, offset=0…900 (10 pages)	1,000	19.73 h observed vs 24 h budget — inside	REST GET /api/v1/filings/submissions?limit=100&offset=0 · MCP {"domain":"filings","resource":"submissions","limit":100}
2	2026-08-18 19:52	filings	/api/v1/filings/submissions via MCP query	limit=2	2 returned; total 441,652	as row 1	MCP {"domain":"filings","resource":"submissions","limit":2}
3	2026-08-18 19:43	system	/api/v1/status/p0	none	8 domain rows	point-in-time probe, no cached age	REST GET /api/v1/status/p0

Study study-filings-submission-metadata-2026-08 · manifest v1.

Sources in this study

• SEC submission metadata — Provider: SEC EDGAR via FinUties. Geography: United States
  (EDGAR registrants). Frequency: Continuous, with a daily index sweep. What it measures:
  One row per filing submission — filer CIK, form type, filing timestamp, period of report,
  accession number, and document locators. FinUties path: /api/v1/filings/submissions.
• P0 claim-surface status — Provider: FinUties freshness probe. Geography:
  Platform-wide. Frequency: On request. What it measures: Per-domain observed age against
  the published P0 budget, plus a fill check. FinUties path: /api/v1/status/p0.

Findings

Finding 1 — The surface is wide and shallow (manifest row 1). 62 distinct form types appear
in 1,000 rows, 14 of them exactly once, across 591 distinct filer CIKs — roughly 1.7
submissions per filer. A default pull is a good census of who filed and a poor census of any
single issuer's history.

Figure 1 — Form-type dispersion in the 1,000 most recent submissions (manifest row 1).
Metric: count of rows by form_type, provider SEC EDGAR via FinUties, units = rows out of 1,000.

Form type	Rows	Share of sample
4 (insider transaction)	203	20.3%
8-K (current report)	88	8.8%
6-K (foreign issuer report)	81	8.1%
424B2 (prospectus supplement)	76	7.6%
144 (proposed sale)	75	7.5%
N-PX (proxy voting record)	74	7.4%
13F-HR (institutional holdings)	52	5.2%
10-Q (quarterly report)	50	5.0%
All other 54 form types	301	30.1%

The three most common forms account for 37.2% of the sample.

Finding 2 — No document body is present (manifest row 1). download_status reads pending
on 1,000 of 1,000 rows, and the four fields that would describe a fetched document —
content_type, link_to_xbrl, local_file_path, content_size_bytes — are null on every row
(primary_document is null on 99.7%). FinUties has indexed the submission and recorded where
the document lives; it has not fetched the body. A job that assumes retrievable documents runs
clean and produces nothing.

Finding 3 — Completeness is inverted across two ingestion paths (manifest row 1). 856 rows
carry a source label (edgar_current_feed 472, daily_index 384) and 144 do not. Labelled
rows have locators and no description; unlabelled rows have a description and no locators.

Figure 2 — Field completeness by ingestion label (manifest row 1).
Metric: null rate by column, provider SEC EDGAR via FinUties, units = percent of rows in each group.

Field	Labelled rows (856)	Unlabelled rows (144)
file_name	0.0%	100.0%
index_date	0.0%	100.0%
link_to_html	0.0%	97.9%
description	100.0%	0.0%
link_to_txt	0.0%	0.0%
accession_number	0.0%	0.0%
content_type, link_to_xbrl, local_file_path, content_size_bytes	100.0%	100.0%

Finding 4 — Six fields are safe join keys, and only six (manifest row 1). link_to_txt,
accession_number, cik, form_type, filed_at and period_of_report are populated on
100% of sampled rows. A job keyed on link_to_html drops 14.1% of rows; a job that requires
description drops 85.6%. The correct fallback chain is link_to_txt first, link_to_html
as an enrichment, description never as a required field.

Finding 5 — Arrival is weekday-shaped, not smooth (manifest row 1). Within the sample
filed_at clusters on six dates: 2026-08-06 (292 rows), 08-07 (253), 08-08 (40), 08-11 (2),
08-17 (151) and 08-18 (262). Nightly jobs should be sized for the Monday catch-up bulge rather
than the daily average.

Finding 6 — Filings was inside budget and the platform was not (manifest row 3). The P0
budget for filings is 24 hours. At 19:43:35 UTC the probe reported filings green with a
most-recent observation of 2026-08-18T00:00:00 and an observed age of 19.73 hours — inside
budget with roughly four hours of headroom. The same probe reported the overall P0 surface as
red, because other domains failed their checks.

Limits and what the data does not say

• The sample is 0.23% of the corpus and recency-ordered. 1,000 rows out of 441,652
  (manifest row 2), taken in server default order rather than at random. The form-type mix
  reflects mid-August filing season and would shift in a 10-Q or 10-K deadline week.
• Nothing here describes filing content. No document was fetched (Finding 2), so this study
  cannot speak to what any filing says.
• The freshness reading is a point-in-time probe, not an SLA. Two calls minutes apart
  returned different per-domain verdicts; only the 19:43:35 UTC call is recorded above.
• Manifest row 3 is a platform probe, not filings evidence. It is included for provenance of
  the freshness claim in Finding 6 and supports no other finding.
• The source label is not a quality judgement. This study observes that labelled and
  unlabelled rows differ in completeness; it does not establish why, and the split could change
  without notice.

Try this query

The headline finding is Finding 2 — the surface carries no document body. Reproduce it:

GET /api/v1/filings/submissions?limit=100&offset=0

MCP query tool arguments for the same rows:

{"domain": "filings", "resource": "submissions", "limit": 100, "order_by": "filed_at", "order_dir": "desc"}

Narrow to one form type before building anything, so you are not handling 62 shapes at once:

{"domain": "filings", "resource": "submissions", "limit": 100, "filters": {"form_type": "10-Q"}}

Freshness callout. Filings budget: 24 hours. Observed at 2026-08-18 19:43:35 UTC:
19.73 hours — inside budget, status green. Re-check at your own read time; the probe is
live, not cached.

Method and confidence discipline

Counts are exact over the returned rows: shares are counts divided by 1,000, and null rates
count null or empty string. Corpus total and sample were read from the same query layer inside
one 15-minute window, so they describe the same state of the table. No value was imputed and no
row was excluded.

High confidence in Findings 1–5: they are direct counts on returned rows, and the completeness
inversion in Finding 3 is structural rather than marginal (0% against 100%, not 40% against 60%).
Moderate confidence in Finding 6, a single point-in-time probe. Low confidence in any
corpus-wide generalisation, for the sampling reasons in Limits.

What would overturn these findings. A pull in which download_status is anything other than
pending, or in which content_size_bytes is populated, overturns Finding 2 and with it the
study's central reading. A pull where unlabelled rows carry file_name overturns Finding 3.

Replies with evidence

Reply with evidence (replies open when forking ships. Every reply cites its own query.)

No evidence replies yet. Re-run the query in this note — if your pull disagrees with ours,
the reply will appear here with the query it cites.

Every reply cites its own query: endpoint, parameters, and pull time. Replies without a
citation are not published.

FinUties data sources

• /api/v1/filings/submissions
• /api/v1/status/p0
• MCP query tool, domain: filings, resource: submissions