Cascadia Portfolio · Fee Examiner

The same firm, named three ways, with no key to join them

A conformed spine across a frozen bankruptcy database, fee applications pulled from the public docket, and two task-code vocabularies — three sources, three owners, none sharing an identifier. Reported honestly at its actual state: 48 of 216 Appendix B-era cases searched so far.

Where the build actually stands

Nothing below claims the acquisition is finished, because it is not.

The population is 216 large-corporate Chapter 11 cases filed under the U.S. Trustee's Appendix B fee guidelines. Of those, 48 have been searched against the public docket archive and 168 have not — a rate limit measured at 50 requests per hour, on either credential tier, is the reason. 6 of the searched cases were truncated at a page cap and are excluded from the rates below rather than counted as complete.

Cases searched

48 / 216

45.8% of searched cases yielded at least one stored document; document-level reach is 2.22%.

Firm identity resolved

9 / 25

Caption-to-database pairs matched by a stated subset rule. Every unresolved pair is published with its own reason, not dropped.

Task vocabulary, exact match

9 / 22

U.S. Trustee project categories with an identical 1997 UTBMS counterpart. 3 have none at all.

Requested paired with allowed

0 / 11

The court's own award, checked against the applicant's ask, under four stated rules. Zero pairs yet — every row says which rule stopped it.

Coverage varies by forum, and Delaware — the largest court in the population — is the least reached

Chart 1 data table

Chart 1 data — coverage reach by court
Court	Cases searched	Documents stored	Reach
cacb	1	0	0.0%
moeb	3	0	0.0%
deb	19	4	21.1%
vaeb	2	1	50.0%
nysb	7	4	57.1%
txsb	13	10	76.9%
njb	1	1	100.0%
txnb	2	2	100.0%

"Not yet searched" and "not covered" are different claims, and this chart carries only the first: every bar is a court that has been searched, and its reach is the share of its cases with at least one retrievable document, out of 216 cases in the population overall. 2 searched cases returned no docket in the archive at all. The documents are public at the claims agents' own sites, which prohibit automated retrieval — this measures what is reachable by the one permitted method, RECAP, and that is the honest scope of the claim.

Two vocabularies for the same billing categories, and one has no counterpart

Chart 2 data table

Chart 2 data — every U.S. Trustee category, by match type
U.S. Trustee project category	Match type
Asset Analysis and Recovery	Exact
Asset Disposition	Exact
Avoidance Action Analysis	Exact
Business Operations	Exact
Case Administration	Exact
Claims Administration and Objections	Exact
Non-Working Travel	Exact
Plan and Disclosure Statement	Exact
Real Estate	Exact
Assumption and Rejection of Leases and Contracts	Rename
Employee Benefits and Pensions	Rename
Employment and Fee Application Objections	Rename
Employment and Fee Applications	Rename
Financing and Cash Collateral	Rename
Meetings and Communications with Creditors	Rename
Relief from Stay and Adequate Protection	Rename
Tax	Rename
Corporate Governance and Board Matters	Broader
Litigation	Broader
Budgeting (Case)	None
Reporting	None
Valuation	None

The U.S. Trustee's own notice says seven of its categories were extracted from an older vocabulary, "all but one" of which is in the 1997 UTBMS set — and does not say which one. This build tested that claim rather than repeating it: the one is Reporting. The three categories with no UTBMS counterpart at all are Budgeting (Case), Reporting, Valuation.

Nine of twenty-five caption-to-database firm pairs resolve by a stated rule

Chart 3 data table

Chart 3 data — firm-identity resolution
Outcome	Caption/firm pairs
Resolved by a stated rule	9
Unresolved, published with a reason	16

Chart 3 detail — the 9 resolved pairs

Chart 3 detail — the 9 resolved pairs, by matched tokens
Database firm	Docket caption string	Matched on
Greenberg Traurig	Greenberg Traurig, LLP	greenberg traurig
Kirkland Ellis	Kirkland & Ellis LLP and Kirkland & Ellis International LLP	ellis kirkland
Kirkland Ellis	Kirkland & Ellis LLP and Kirkland & Ellis International LLP	ellis kirkland
Kirkland Ellis	Kirkland & Ellis LLP and Kirkland & Ellis International LLP	ellis kirkland
Kirkland Ellis	Kirkland & Ellis LLP and Kirland & Ellis International LLP	ellis kirkland
Kramer Levin	Kramer Levin Naftalis & Frankel LLP	kramer levin
McDermott Will	McDermott Will & Emery LLP	mcdermott will
Pachulski Stang	Expenses of Pachulski Stang Ziehl & Jones LLP	pachulski stang
Skadden Arps	Skadden, Arps, Slate, Meagher & Flom LLP	arps skadden

A 100% match rate would be evidence the matching rule is too loose, not that the data is clean. The rule matches when the database's short editorial name is a token subset of the docket caption's full name — never a similarity score, never a single shared token unless it is unique in that docket. The 16 pairs that do not match are published rather than dropped — but they share one stated mechanism, not sixteen distinct diagnoses: the subset rule did not match, most often because the database records only the lead firm and the docket names a co-counsel, local, or conflicts firm the database never recorded — e.g. Barley Snyder LLP and Expenses of Giuliano Miller and Company LLC. Telling those 16 apart by their actual cause (an abbreviation mismatch versus a genuinely unrecorded firm, say) is not yet built; recorded as a build-forward candidate for the resolver itself, not answered by this chart.

One application's own staffing split, plus two smaller filings that add only paraprofessional time

Chart 4 data table

Chart 4 data — staffing mix, hours and fees share by role
Role	Applications reporting	Hours share	Fees share	Rate min	Rate max
partner	1	49.7%	62.5%	$875	$875
associate	1	46.9%	34.2%	$482	$482
of_counsel	1	1.8%	1.9%	$825	$825
paraprofessional	3	1.7%	1.4%	$200	$545

This is not a staffing mix across 3 applications — a reading panel asked the right question and the answer changes the chart. Partner, associate and of-counsel hours all come from the same single document; the other two in-scope filings contribute only a paraprofessional line each, pooled into the paraprofessional bar (3 documents). So the partner/associate/of-counsel split is one application's own internal role breakdown, not a pattern observed across a sample — and none of the 3 documents carries an extracted firm name. That is not a resolution failure across the board: a different filing on the same docket (deb 20-11177) resolved cleanly to a named firm — it is these specific documents whose caption text did not match the extraction pattern. Read this chart against the coverage figures above, never as a claim about the population.

Zero of eleven applications could be checked against what the court allowed

Chart 5 data table

Chart 5 data — why requested could not be paired with allowed
Blocking rule	Applications
No parsed order on this docket	8
Order is omnibus	2
Filer strings do not resolve to one firm	1

This is the module's one figure against an answer key it did not produce — the court's fee order, written by a different author than the application it responds to. An omnibus order publishes no single allowed figure by design: several professionals' awards in one order cannot be attributed without a rule this module has not written, so the award count is published and the amount is left blank rather than summed or guessed.

The largest bar is neither a timing gap nor a parser miss — a reading panel asked which, and the corpus can answer. Across the 5 dockets behind the "no parsed order" applications, the classifier found 77 order-shaped docket entries on 4 of them — the orders exist and were correctly identified. Every one of those 77 entries is marked unavailable in RECAP: the court docket carries the entry, but nobody has purchased and uploaded the PDF. This is the same coverage mechanism Chart 1 measures for applications, confirmed here for orders.

How the population sizes on this page relate, stated once rather than left for a reader to reconstruct. 216 cases make up the population; 48 have been searched so far (Chart 1); 63 documents have been retrieved from those searches, among which 25 distinct caption-firm pairs were seen (Chart 3) and 11 documents were successfully parsed; 11 parsed applications had a matching docket to attempt a court-order pairing against (Chart 5). Each number is a different filtering stage of the same corpus, not an independent sample — a case counted in one is not guaranteed to appear in another.

How it stays right

The figures above are read from the module's own health surface, written by the same gates that must pass before anything is published — not restated by hand.

Build state

All three publish gates passed: true
Last acquisition run: stopped: rate window clears in 3605s; stopping rather than idling
Docket entries classified: 3,667 (2.62% unclassified)
Documents parsed: 11 of 63 (17.5%)
Spine frozen as of: 2023-01-12
Health surface generated: 2026-09-10 00:49:45 UTC

What this module refuses to publish, and why

A module that only ever shows what it measured has said nothing about what it declined to measure. Every refusal below is a decision, not a gap that was never noticed.

Measures deliberately not certified
What is refused	Why
Total professional fees per case	The archive holds a minority of the applications in most cases; a per-case total would sum an unknown fraction and read as complete. Coverage is published as its own measure instead.
Anything at per-person grain	Rates by named timekeeper, seniority curves, leverage by individual. The source data supports all of them. No per-timekeeper row exists anywhere in this module's pipeline — the parser aggregates to application × firm × role as it reads, before anything is written.
Anything cut by debtor identity	No output carries a party column. Telecommunications debtors are excluded at the spine, before any measure, by an industry-code test rather than a name match.
Awarded amounts as an input to any other measure	They are the reconciliation's answer key. Using an answer key as an input is how a check stops being one.

Disclosure

An independent portfolio project by Aaron Robbins. This module uses data from the Florida-UCLA-LoPucki Bankruptcy Research Database, as its use conditions require. The UTBMS bankruptcy code set is the intellectual property of the American Bar Association, reproduced by the LEDES Oversight Committee with the ABA's permission. Court filings are public records. No natural person's name appears anywhere on this page or in this module's data — not a timekeeper, not an attorney of record, not a judge — and no telecommunications debtor is analysed. Nothing here is legal advice.

Source, governance documents and build scripts: github.com/RobbinsAnalytics/cascadia-fee-examiner. Every figure on this page is computed at build time from a certified artifact; the independent re-derivation that gates publication is src/validate_measures.py.

Spine frozen as of 2023-01-12 · health surface generated 2026-09-10 00:49:45 UTC · cost layer retrieved per document, incomplete at 48 of 216 cases