I will be posting quite a bit of content over the next couple of weeks. I have been working on several projects this summer, and they are beginning to come together. Let’s start with a new database that is now available: Updated Dissemination Time. (Updated Dissemination Time). When I first needed to access acceptance time values I originally pulled these from an archive of the content that was delivered through the SEC PDS feed and frankly I was in a hurry and trying to be responsive to a request. As of earlier this year the need to automate the update of the original db was languishing in my to-do list. I finally started working on it a couple of months ago.

This time, rather than just vacuum in the acceptance time I decided to create something more comprehensive that would allow you to consume as much or as little of the details as you wanted. If you follow the link above you can review the field definitions in the database. In addition you will find a brief description of two Python programs that are available in the S:\PythonCode directory that you can run to collect data at scale.

The problem was (and in some ways still is) the SEC documentation of the meaning of the control block delivered in the feed is sparse. Here is the explanation of the fields from the PDS Technical Specification with my interpretation of their practical meaning.

|   Bytes | Length | SEC field               | Interpreted meaning                                                                                                                                                                                                                                                                                                           |
| ------: | -----: | ----------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
|    1–20 |     20 | Accession Number        | The EDGAR accession number assigned to the submission. It should be stored as text, including the hyphens. For a post-acceptance correction, this field identifies the filing to which the correction applies.                                                                                                                |
|   21–40 |     20 | Submission Type         | The form or submission type, such as `10-K`, `8-K`, or `DEF 14A`. The value is padded with trailing spaces.                                                                                                                                                                                                                   |
|  41–100 |     60 | Submission Company Name | The conformed name of the company associated with the submission. This is descriptive information and should not be treated as a substitute for the company’s CIK.                                                                                                                                                            |
| 101–160 |     60 | Subject Company Name    | The conformed name of a separate subject company, when applicable. This field is often blank.                                                                                                                                                                                                                                 |
| 161–168 |      8 | Receipt Date            | The date EDGAR received the submission, in `YYYYMMDD` format. This date can differ from the acceptance date.                                                                                                                                                                                                                  |
| 169–176 |      8 | Acceptance Date         | The date EDGAR accepted the submission, in `YYYYMMDD` format. Together with Acceptance Time, it identifies when EDGAR formally accepted the filing.                                                                                                                                                                           |
| 177–182 |      6 | Acceptance Time         | The time the submission was accepted, in `HHMMSS` format. The SEC specification labels this time as EST.                                                                                                                                                                                                                      |
| 183–188 |      6 | Build Time              | The time EDGAR completed the build of the dissemination record, in `HHMMSS` format. 
| 189–194 |      6 | Reserved                | An unused field that may be blank or may repeat Build Time. It should not be assigned a substantive meaning.                                                                                                                                                                                                                  |
|     195 |      1 | Priority                | A binary value, not a printable ASCII digit. The SEC defines the values as `3` for a post-acceptance correction, `4` for low priority, `8` for medium priority, `12` for high priority, and `16` for a high-priority.                                                                                                 |
|     196 |      1 | Public/Private Flag     | A binary value indicating whether the dissemination is public or private. The current published table lists the possible meanings but does not visibly provide the corresponding numeric codes. The raw byte should therefore be retained.                                                                                    |
|     197 |      1 | Transaction Type        | A binary value representing a submission, retransmission, correction, deletion, submission-form-type transfer, or document-type transfer. The current published table lists these categories but does not visibly provide the numeric mapping. The raw byte should be retained rather than used alone to classify a deletion. |
| 198–201 |      4 | SRO Distribution List   | An obsolete and unused distribution field. It has no current substantive filing meaning.                                                                                                                                                                                                                                      |
| 202–266 |     65 | Reserved                | Unused blank padding at the end of the control block. Unexpected nonblank content may indicate a parsing-offset problem or a format variation.                                                                                                                                                                                |

One important qualification is that the SEC labels one control-block field Build Time, not Dissemination Time. For ordinary filings, the dissemination timestamp in the database is therefore a derived value based on the available control-block fields and archive date. For post-acceptance corrections, the SGML header can contain a separate <TIMESTAMP> that the SEC explicitly defines as the correction’s dissemination date and time.

The table describes the 266-byte layout in Section 2.1 of the March 2025 PDS specification. Older specifications define a 256-byte layout in which Submission Type occupies 10 rather than 20 bytes. The March 2025 draft itself retains a reference to a 256-byte block in its introductory discussion. The database records the observed control-block width and accommodates both layouts.

Another challenge appears when the field definitions are applied across all of the control blocks. The SEC defines Build Time as the time at which a submission was built on EDGAR, so one might expect it to follow Acceptance Time. Yet 6,722 records contain a Build Time value that is earlier, as a time of day, than the Acceptance Time value. Because the control block supplies a Build Time but no separate Build Date, these records should be treated as apparent timing inconsistencies rather than definitive cases in which a filing was built before it was accepted.

An especially useful feature of this work is that we can more carefully identify filing changes—post-acceptance corrections, or PACs—and whole-filing deletions. From January 1, 2023, through June 30, 2026, the database identifies six Form 8-K filings that were later deleted, one associated with each of Tractor Supply, Guidewire Software, Southwest Airlines, Agrify Corp., ClearOne, and Insight Enterprises. The deletion classification is based on the top-level <DELETION> tag in the post-acceptance correction header, rather than on the control block’s transaction-type byte alone.

As always, you do not need Python to begin exploring these data. Select the database and build an expression in our Query Tool. The screenshot below shows the query I used to identify the six deleted Form 8-K filings.

Until the new database is complete, if you want acceptance time for the complete span from 1994-2026 you will have to use both databases. As I add additional years I will be changing the name to reflect the content. Once we have fully captured the full history I will retire older database.

One thing that is on my list with this data is to start looking more closely at how the build time varies based on high volume filing days. While the SEC reports that the cut-off time for acceptances of most forms is 5:30 PM Eastern Time we have seen filings ‘appear’ after that time which we generally chalk up to a backlog in their build queue. There is quite a lot that has to happen after acceptance before a filing is disseminated and I have been hypothesizing that this has increased the delay on peak filing frequency days. Since this time is after the US markets are closed it would not seem to affect daily trading. I am just curious.

Let me wrap this up by reporting that I don’t know what I don’t know. I am still doing research on some features of this data. I hope to learn more and be able to update. I do have another piece I want to add to this, but it needs to wait until this first pass is complete.

Earlier this week, I reported finding content in PPL’s 10-K that was present in the filing’s HTML but not visible when the filing was rendered in a browser. After confirming that the problem was not caused by directEDGAR’s parsing engine, I searched our archive for the same pattern.

The search identified 19 filing documents, representing 16 distinct disclosure instances. The difference between 16 versus 19 arises because one instance involved parallel filings by PPL and three subsidiaries with separate reporting obligations..

In six of the 19 filing documents, the hidden material was either repeated elsewhere in the filing or unnecessary to the document in which it appeared. LivePerson’s amended 10-K provides an interesting example. The amendment was filed to provide the Part III information required by Items 10 through 14. However, its source HTML also contained hidden copies of nearly all the financial statement notes, beginning with the second paragraph of the notes in the original 10-K. That material was unnecessary to the amendment and had already been reported in the original filing.

The hidden material can be inspected through EDGAR by opening the filing as HTML and then viewing the page source. The styling applied to the relevant HTML container prevents the content from appearing in the normally rendered page. I have it displayed below. This view represents the filing before I removed the hidden material from directEDGAR’s processed copy.

For the six filings containing duplicative or unnecessary material, I removed that material from directEDGAR’s processed display and searchable copy. In the remaining 13 filing documents, the hidden material appeared potentially relevant. Therefore, I adjusted directEDGAR’s processed HTML so that the material would be visible and searchable. I applied a deliberately broad standard: if the material could reasonably have appeared in the filing, I preserved it.

GoPro’s 10-K, for example, contains a hidden Rule 10b5-1 trading-arrangement disclosure. The disclosure is nearly identical to one included in GoPro’s Form 10-Q for the quarter ended September 30, 2025, available here. The principal difference is that the final sentence in the 10-Q refers to the third quarter, while the hidden 10-K text refers to the fourth quarter ended December 31, 2025. Frankly, because the disclosure substantially duplicates the Form 10-Q disclosure, I could reasonably have treated it as duplicative. However, because the final sentence refers to the fourth quarter, I retained the material to be safe. It is now visible and fully searchable.

The metadata for all the affected filings identifies the Workiva platform. The available evidence suggests that the hidden material may have remained in the HTML from an editing or review stage. Some of the hidden content, such as the material from Consensus Cloud Solutions shown below, even retained yellow-background formatting that appears consistent with editing or review markup. This is suggestive, although it does not establish precisely how the material came to remain in the filed HTML.

In summary, when the hidden material was duplicative or unnecessary, I excluded it from directEDGAR’s processed version. When it appeared potentially relevant, I made it visible and searchable. I did not attempt to move any disclosure to another location in the filing because doing so would require guessing where the registrant intended it to appear. That would introduce more judgment than I am comfortable applying.

I need to run a scan across prior years as well as other filings to see if this has occurred elsewhere. I probably will not report the results of that scan unless there is a significant finding. The issue was interesting, but I do not think any of the disclosures were necessarily significant. My overall sense was that the disclosures that were not visible were unlikely to affect a reader’s interpretation of the filings.

I am redoing the code we use to parse ITEM sections for indexing. I was reviewing the exceptions today and I came across something fascinating (to an EDGAR geek at least).

The lease note in the 10-K for PPL Corp (Ticker PPL; CIK 922224) is present but not visible in their 10-K for the last two fiscal years. Here is a link to their latest 10-K (PPL 12/31/2025 10-K). If you do a search in the document there is an index to the notes to the financial statements on page 99. There is no lease related note in that schedule.

The lease note is visible in our platform version of the file we processed for PPL and each of the subsidiaries but it appears before the actual cover page for the 10-K. Below is a screenshot of the results of a search for 10-K filings – the filing in the document pane is PPL CORP’s 10-K. Although this is difficult to confirm from the screenshot, the document viewer is positioned at the very top of the filing. The lease disclosure appears in our processed version before the filing’s cover page.

This is not visible in the filing accessed through EDGAR – the content exists but is placed before the cover page content and it is tagged to be hidden. Here is a simplified sketch of the html structure that causes this content to be invisible:

<div style="display:none">
  <ix:header>
    <ix:hidden>
      <ix:nonnumeric
        name="us-gaap:LesseeOperatingLeasesTextBlock"
        id="f-3742"
        contextref="c-1"
        escape="true">
        ...
      </ix:nonnumeric>
    </ix:hidden>
  </ix:header>
</div>

Processing these filings is complicated. Because this content was contained within the Inline XBRL header, and our processing attempts to preserve header content, the lease note remained in our version of the filing.

The characterization of this content as a lease note is also supported by PPL’s 2023 Form 10-K. That filing included a visibly rendered “Note 10. Leases,” and the disclosure embedded—but not visibly rendered—in the subsequent filings is substantially similar to the visibly rendered Note 10 in the 2023 Form 10-K.

This surfaced because I kept seeing an exception while trying to improve our TOC engine and initially I thought I had a logic problem. I finally looked closer at our version of the filing and started off thinking that we must have done something wrong during parsing. It was only after screen printing the EDGAR version code while comparing it to our version and looking more closely at the tagging structure that I relaxed – we didn’t mess up!

This appears to be a filing-construction error rather than an absence of the disclosure. The lease note exists in the filed Inline XBRL document, but the document’s HTML structure prevents it from being displayed through EDGAR. It also illustrates why processing EDGAR filings involves more than simply extracting the text that appears on the screen: sometimes important content is present in the source even though a reader cannot see it.

There are a number of papers in the accounting literature that have described using directEDGAR to collect inventory write down data beginning with the Allen, Larson and Sloan 2012 Journal of Accounting and Economics paper. Historically, the collection for this data has been based on running key word searches in 10-K filings to identify potentially relevant content and then reviewing the extracted context to determine if it should be included in the sample.

Another paper that used very similar data was Ming, Markov and Shu’s 2023 Accounting Review, Motivational Optimism and Short-Term Investment Efficiency, I constructed a search that I think was close to theirs (DOCTYPE contains(10k)) and inventory w/5 (write* or obsolescence or obsolete or charge* or revaluation or provision*) pre/15 === and applied it to the 2025 10-K archive (the === signs force the search to only return results that are followed by at least three digits.). The results seemed consistent with what they reported (even though their data collection covered an earlier time period). Here is a screenshot of the search results.

Based on reading their paper and how I would collect the same data I would then run a SummaryExtraction to get the context into a CSV file to then scan and code the results.

I then wondered, could we accelerate this using the XBRL databases we have created? One challenge with using the databases is that I would either have to identify the relevant row labels or try to identify the way the data should be tagged. I decided to try the tag route first. Because I could not possibly expect filers to use a constrained set of tags I decided to make an assumption that the actual XBRL tags had to include the word inventory and then at least one of the words in my (their) original search. Here is the search I ran with the Query tool ((name LIKE ‘%charge%’) OR (name LIKE ‘%provision%’) OR (name LIKE ‘%obsol%’) or (name LIKE ‘%write%’) or (name LIKE ‘%reval%’) ) AND (name LIKE ‘%inventory%’). To be clear, I am requiring the tag used for the data to have the word inventory and at least one of the other character sequences.

Here is a screenshot of the results

I saved the results to review them and compare to the full-text search results. There were 2,298 total observations with 212 unique tags. Not all of the results were relevant. Here is a sample of some of the tags.

aaon:InventoryValuationReservesWriteOffs
iex:CostOfGoodsAndServicesSoldExcludingInventoryStepUpCharges
iex:FairValueInventoryStepUpCharges
cgnx:ExcessAndObsoleteInventoryCharges
icui:WarrantyAndReturnReserveInventoryChargedToOtherAccounts
icui:WarrantyAndReturnReserveInventoryWriteOffs
ftk:ProvisionForExcessAndObsoleteInventory
gti:CostOfGoodsAndServicesSoldExcludingInventoryWriteDown
nvax:ProvisionForExcessAndObsoleteInventory
lctc:ProvisionForInventoryObsolescence
duot:InventoryWriteoff
ngs:InventoryAllowanceAllowanceForObsolescence
ngs:InventoryWriteOffs
ew:DisposalGroupIncludingDiscontinuedOperationInventoryWriteDown
cvgw:ProvisionForAdvancesOnInventoryPurchases
arw:InventoryReversalOfWriteDown
clne:ProvisionForDoubtfulAccountsNotesAndInventory
acls:ProvisionForExcessAndObsoleteInventory

I see a really clean work-flow following from this. I think the first thing I would do is separate the tags, identify the irrelevant ones (cvgw:ProvisionForAdvancesOnInventoryPurchases) and filter those out. I won’t go further because I imagine many of you can already imagine the next steps. The point is that accessing this data significantly accelerated when the data is tagged.

The key is to be creative with how you initially filter for relevant tags. If I had to consider every tag with the word INVENTORY (there were 21,985 tags with the word inventory) the review would be overwhelming. But by requiring the word INVENTORY and then at least one other of the word roots that describe the possibility of a write-down I am evaluating a much smaller set of results for relevance.

Of course, the results are not likely to be exhaustive. I found some interesting omissions. Here is a screenshot from Blink Charging’s full-text search results:

This data was not in the results. I searched for it directly using the Query Tool and discovered that the tag was ProvisionForOtherLosses. But I noticed that they did use relevant text in the original row label as you can see in the screenshot above. I think the next step is to get creative in a similar manner using the orig_row_label field which is also fully searchable. I ran this search (orig_row_label LIKE ‘%inven%’) AND ( (orig_row_label LIKE ‘%obsol%’) OR (orig_row_label LIKE ‘%slow%’) or (orig_row_label LIKE ‘%write%’) ).

While there is sure to be duplication in the second run, that is easy to handle because you can filter out all of the tags that were in the first run.

I am not going to assert that this is going to provide complete results. But I believe that if a third round of filtering is needed it should be minimal. Finally, if there is still data to be collected, we can fallback to the full-text search after filtering out the CIKs from the filers whose tagged data was useful. Since only the financial statements and notes are tagged, if the filer made the disclosure only in the MDA without separately reporting the amount in the financials or notes then a full-text search would be the only way to identify the value.

I want to wrap this up by adding a bit of marketing. The real benefit of our platform comes from the visibility we give you into the filings. With that initial search I could easily move through the collection of filings to get a sense of my results. I instantly know if I need to expand (or contract) the search. And then within the same interface I can immediately access the tagged data. The ability to filter tags without requiring their adherence to the taxonomy provided significant benefits. Again, once I saved the results and then looked back at my search results I was ready to filter on the row labels.

When Paul Atkins, Chair of the U.S. Securities and Exchange Commission, spoke at the Texas A&M University School of Law Corporate Law Symposium, he discussed the growing complexity of SEC disclosure rules.

Referring to executive compensation disclosure under Item 402 of Regulation S-K, he said:

“The rule today has morphed into a Frankenstein monster beyond recognition… disclosure intended to inform can instead overwhelm.”

Today, Berkshire Hathaway filed its DEF 14A proxy statement — and it provides a fascinating example of that concern. Because of Item 402(u), the company must disclose its CEO pay ratio.

Yet Warren Buffett has earned $100,000 annually for more than 40 years, with no bonus and no equity compensation.

Nevertheless, Berkshire still had to estimate the median compensation of 387,815 employees across more than 60 operating groups in order to compute the required ratio.

The result:

• CEO compensation: $389,488 ($289,488 of this is security)
• Median employee compensation: $93,709
• Pay ratio: 4.16 to 1

In the proxy statement, Berkshire explicitly questioned whether the cost/benefit of complying precisely with Item 402(u) provides meaningful information to shareholders. While the cost is probably inconsequential to the company, I suspect the attention required for this is non-trivial since they run such a tight ship.

If the SEC is serious about rationalizing, simplifying, and modernizing Regulation S-K, this raises a straightforward question:

Should CEO pay ratio disclosure be required when total CEO compensation is below $1 million?

For companies like Berkshire, the disclosure may add complexity without materially improving investor understanding.

And Berkshire is not unique — more broadly, there are a non-trivial number of public companies where CEO compensation is below $1 million, suggesting this issue may not be unique.

For researchers interested in examining these disclosures, Mr. Buffett’s compensation — along with the all of the other proxy filers who made their filing today — will be available around midnight tonight for analysis in the directEDGAR Executive Compensation database.

We completed the out-of-cycle update to the Tax-Recon database. Manish did a heroic job.

First off, the distribution so far – 988 firms that have filed their 10-K as of yesterday (3/2) have adopted the requirements of ASU-2023-009 on a prospective basis. 760 adopted on a retrospective basis. When you access the database – remember that ASU-2023-009 was required for FY beginning after 12/15/202. If you filter on 2026 filings you will find no ASU_STANDARD flag for 214 accession numbers – so far one of these is a REIT, the other 213 had fiscal years ending in October, November and even early December so they are just now subject to the new disclosure requirements.

While we are investigating ways to remove some of the flags – their existence should not keep you from moving forward. For example, there are 245 cases where we have inserted a CHECK_TOTAL flag. It looks like the vast majority of these are because the registrant reports a subtotal – that is not labeled as such. Here is an example from GATX’s 10-K –

The Foreign tax effects line represents the net foreign tax effect on the dollars and rate. When we test the addition (and we do so to confirm that we are not missing something that is hidden) we would double count those values as we would include all of the values above the subtotal. We are not changing anything – the reported totals are available in the database with the values you see. But because we don’t know we have flagged this table as one that requires review.

I have to be honest – I am not sure yet how we should handle these. One option to to add an additional flag to indicate that row is a subtotal. Another option is to delete the row.

The MULTI-TABLES flag is harmless – it is the starting place for our analytical analysis of the ASU-2023-009 adoption method. We do much more than test for the existence of multiple tables but it is part of that routine. We left that flag in the client facing database because it can be helpful if you report a concern.

With respect to the ASU_2023_09 flags. The flag ASU_2023_PROSPECTIVE identifies the data in the filing that conforms to the new disclosure standard for the CURRENT year. The flag ASU_2023_PROSPECTIVE_HISTORICAL identifies the historical (generally prior two years) data that will generally be presented in only one unit. But again, be careful, there are filers that presented data in both percentage and monetary units that adopted the standard prospectively, Boeing is a good example). Boeing has two tables, one that conforms to the new standard (and has both dollars and percentages) and their historical disclosure that also reports both dollars and percentages. The flag ASU_2023_RETROSPECTIVE identifies the filings where the registrant adopted the ASU on a retrospective basis.

I ran some very rough analysis – there are about 30% more labeled data rows when firms meet the ASU 2023-009 requirements as compared to their historical disclosures. There are lots of other interesting features in the data but I know some of our clients are focused on those.

The next big push will be when the 12/31 Accelerated filers start filing in earnest. Their deadline is 3/16 so that is not too far away.

As a reminder the dimensional flags will be really helpful so don’t forget those as you are working with this data.

We are continuing to test and refine our approach to capturing the impact of ASU 2023-09 on the structure of the Effective Tax Rate Reconciliation table. Our work is grounded in the actual table structures that companies have used to implement this expanded disclosure.

One of the key attributes we believe is important to capture is the adoption method (retrospective or prospective). In many filings, the adoption method is not explicitly stated. Instead, it must be inferred from the structure and content of the reconciliation tables. To address this, we developed an algorithm that analyzes table structure and determines the adoption method based on observable disclosure patterns. As a result, we have introduced a new field, ASU_STANDARD, to identify disclosures that conform to ASU 2023-09.

If the ASU_STANDARD field is blank, the associated rows were not disclosed in conformity with ASU 2023-09.

If the field reports ASU_2023_09_PROSPECTIVE, two implications follow for that accession number:

1. All rows labeled with ASU_2023_09_PROSPECTIVE represent the ASU 2023-09–mandated disclosure.
2. A second reconciliation table will typically exist for the same accession number that presents the Effective Tax Rate reconciliation using the prior disclosure format. This reflects the prospective adoption method, under which ASU 2023-09 applies only to the current period, while prior periods continue to be presented under the previous disclosure requirements.

As a practical matter, this assumes the filer is not a new registrant, since a newly reporting entity may lack prior-period operating history and therefore may not present a legacy-format reconciliation table.

If the ASU_STANDARD field has the value ASU_2023_09_RETROSPECTIVE, only one logical reconciliation table is expected for that ACCESSION, and that table will typically report up to three years of data.

The term “table” is used here in a logical rather than strictly physical sense. In many filings, the ASU 2023-09 reconciliation disclosure spans multiple pages in the Form 10-K and is therefore physically presented as two or three separate tables. However, these tables collectively represent a single reconciliation disclosure.

When this occurs, we treat the sequence as one logical table. All rows from the component physical tables are assigned the same TABLE_NO, which corresponds to the index of the first table in the sequence. This distinction between physical tables and logical tables reflects the presentation structure of SEC filings while preserving a consistent analytical unit for downstream processing. This approach ensures that the full multi-page disclosure can be reconstructed and analyzed as a single, continuous reconciliation table.

There is a structural issue I would like to address but I think it is too early to do so. The new standard requires separate disclosures by both the jurisdiction and nature of the tax/benefit if they are 5% or more of the total income taxes paid. The jurisdiction can be identified in two ways. First by the country name in the row label as you can see in the next image:

The problem with relying strictly on the row label is that the row label may be a caption for a subsection. To understand my meaning review these rows in the database from the 10-K filing submitted by TEVA PHARMACEUTICAL INDUSTRIES INC (CIK: 818686)

If you look closely you will see that none of the rows with the names of foreign jurisdictions have data. That is because those are section identifiers. To determine the country related tax effect it is necessary to use the dimensional information provided by the filer. When this information is disclosed we capture it in the explicit_member_dim (describes nature of the intended disclosure) and then the explicit_member_text field which provides the identifying information as you can see in the next image.

The structural issue I would like to address is to add country name fields with a value of 1/Null. However, I don’t think we have quite enough information yet to do so. We need a much larger sample of these disclosures to develop the right system to map the disclosures with the right country.

Speaking of a much larger sample, March 2 is the filing deadline for 12/31 Large Accelerated Filers. For those of you eager for this data to beat the crowd we are updating our data nightly and our client facing data weekly. However, we will make a special update early in the morning of March 3rd with the filings through 3/2. However, let me warn you that about 2-4% of the filers will be one or two days late. Further, the Accelerated Filers deadline follows on 3/16.

One of the promises of inline XBRL was simplicity.
If a data point appears on the cover page of a 10-K and it’s tagged, then—at least in theory—it should be easy to extract, compare, and analyze.

Public float is a good example.

The SEC requires filers to report their public float on the cover page, and with tagged cover pages that value is now explicitly labeled and machine-readable. Many researchers and data users understandably rely on the SEC’s structured data feeds to capture it.

But there is a quiet problem.

In a subset of filings, the tagged public float value does not match what the company actually reports on the cover page.

Here’s what that looks like in practice (Packaging Corp of America 12/31/2024 10-K):

• Reported on the HTML cover page
  • $16,124,636,098
• Tagged Value in XBRL:
  • 16,124,636,098,000,000

The widely used SEC Structured data files contain the same value as the tagged XBRL. To explore you can download the February 2025 data folder and after opening look at line 7,553 in the sub.tsv file.

That’s not a rounding issue or a scaling choice. It’s a conversion error—effectively inflating the reported public float by a factor of one million.

What’s especially interesting is that these filings share a common fingerprint. They appear to have been prepared using a specific filing platform, as indicated by metadata embedded directly in the submission document.

<!-- DFIN New ActiveDisclosure (SM) Inline XBRL Document -->
 note - creation date varied across the 30 so it was not a specific date issue
<!-- Copyright (c) 2025 Donnelley Financial Solutions, Inc. -->

Based on the above, this doesn’t look like a company-specific mistake. It looks like a systematic transformation issue introduced during the filing conversion process. However, it’s more complicated than that. Based on my analysis, this software was used to file 892 10-K (I excluded 10-K/A) in 2025 but the problem seems to have occurred in roughly 30 or so.

Why this matters

If you rely on the SEC’s structured data endpoint you’ll ingest the overstated value without any obvious warning. The number is valid XBRL; it parses cleanly. Unless you validate it against the rendered filing, it looks perfectly legitimate. Why would you question it?

That’s a problem for:

• researchers using filer size thresholds in their analysis
• analysts filtering by public float,
• and anyone using public float as a screening criteria

A simple safeguard

In our own workflows we actually start with the HTML filing itself and then map tagged XBRLvalues back to what the company actually reports on the cover page. This makes discrepancies easier to flag because even for our client facing databases we report to you both the original value as reported in the html and the tagged value as you can see in the next image

This imposes additional work but it also gives you a point of reference that is lost with the raw XBRL data no matter how you access it. Ingesting just the XBRL give no reference to truth. The Inline XBRL makes more data accessible, the HTML provides a clear way of evaluating the results.

Conclusion

When the first XBRL filings were introduced, we were ready—and genuinely excited. The promise was compelling: standardized, machine-readable financial statements that could be reliably constructed directly from tagged data.

That excitement didn’t last long.

As we began validating what we were deriving from the filings, it became clear that the work required was far more exhaustive than anticipated. In one early effort, we constructed roughly 5,000 income statements from XBRL data and compared those to the original financial statements. Approximately 7% contained problems.

What made this especially frustrating was not the error rate itself, but the nature of the errors. We were unable to devise any rule, heuristic, or algorithm that could reliably identify which statements were wrong. The issues were only visible when a constructed statement was compared to the one actually presented in the filing. So XBRL could not be a starting and ending point.

Inline XBRL has materially improved this situation. By embedding tagged data directly within the rendered document, it provides the necessary context to test, validate, and reconcile structured data against what filers actually report. That context doesn’t eliminate errors—but it makes them observable.

The lesson hasn’t changed: structured data creates opportunity, but only when it is paired with validation, traceability, and context. Inline XBRL doesn’t solve data quality problems on its own—but it finally gives us the tools to see them.

We are preparing to drop a fairly comprehensive Audit Fee data base. Long time coming but one of the real challenges is that there can be some wonky data and it is what it is. I am spending time trying to devise tests to evaluate the quality of our parsing and data assignment code. One of the tests I dreamed up is checking if the Date of the Audit Report precedes the dissemination data by more than 10 days. Here is a screenshot of the audit report for CSW Industrials (CIK 1624794) for the year ended 3/31/2022.

Here is a screenshot of the auditor’s signature:

It is not a huge issue, the filing was made on 5/18/2022 and perhaps someone felt harried when doing the final checks. For us though, it raises a significant question – when do we have liberty to change data after it is reported? I am so reluctant to make changes because that seems like a slippery slope. What is the source of truth? As an aside, the auditor submitted an Exhibit-18 describing the change from LIFO to FIFO that was filed with the 10-K. That has the date 5/18/2021.

We have identified about 100 different cases of similar/adjacent issues. Trying to validate that what we are doing is correct is our primary focus. A secondary issue though is to decide what to do. As I write this I am more inclined to use the 5/18/2021 date just because the 10-K should be the source of truth.