Empirical Validation¶

Version 10.0 | 2026-06-11 — every machine-testable official edit on the FR Y-9C and the Call Report, validated against 208 million + 1.9 billion rows of real filings; plus the UBPR derivation formulas validated against real published UBPR values, and a conditional/compound testing engine that drives every registry row to an explained verdict.

Most data dictionaries assert that "assets equal liabilities plus equity" and stop there. This one tests it. Every reconciliation identity, bound, and official edit check we publish is checked, as an actual aggregate query, against the full history of real bank filings — and we publish the pass rate, the observation count, and an honest verdict for each.

In v8.0 that meant the FR Y-9C, tested against 208 million rows of bank-holding-company filings. v9.0 adds the Call Report: we acquired the official FFIEC Call Report edit checks and calculation linkbases (machine-readable, current cycle), structured them into the registry, and evaluated every machine-testable one against 1.917 billion rows of Call Report filings (2001 Q1 – 2026 Q1). We also ran a cross-source concordance that checks 16 core concepts between the Call Reports and the FDIC's independently-collected Summary of Deposits / Statistics on Depository Institutions (SDI), and corrected two definitional mappings in the process.

The combined Relationship Registry now holds 7,539 relationships, every one carrying a status and — where machine-testable — its empirical observation count and pass rate. Zero rows are left pending or unexplained.

v10.0 adds two things on top of v9. First, a conditional/compound testing engine (IF/THEN rewritten to SQL CASE, AND-conjunction splitting, and constant coefficients) made 448 of the 2,606 previously documented-but-untested conditional/compound rows machine-testable, and every one of them was adjudicated to an explained verdict; the rest stay honestly documented (2,158 GRAMMAR_DOC: null-logic and ratio-division families). Second, the UBPR — the FFIEC's analytical report of derived ratios — was acquired from the official taxonomy v181 and User's Guide, its 4,207 derivation formulas parsed with zero silent drops, and its 31 headline ratios recomputed and empirically validated at 99.77–100% against real published UBPR values. The full v10 status distribution is in the Results section below.

Where the official edits come from¶

The v9.0 Call Report layer is built on the official FFIEC Call Report edit checks and calculation linkbases, distributed in the CDR XBRL taxonomy (cdr.ffiec.gov). These are the same machine-readable formula linkbases the regulators use to validate filings: 3,615 edits per form plus 243 calculation-linkbase identities, for all three Call Report forms (FFIEC 031 / 041 / 051) at the current cycle (taxonomy v294, 2026-03-31). We additionally archived 30 historical cycles (year-ends 2001–2024 plus every 2025–2026 quarter) so the edit set can be aligned to the vintage of any filing we test.

This is the authoritative source: rather than re-deriving edits from the human-readable instructions, v9.0 parses the regulators' own formula linkbases directly, so every tested Call Report relationship traces to an official edit id or calculation-linkbase parent.

What was done¶

The methodology is the same on both sides: acquire the official edits → parse the formula linkbases → structure into the registry → test every machine-testable relationship against bulk filings → adjudicate every non-passing result, never silently dropping one.

Acquire and parse the official edits. For the Call Report, the formula linkbases were taken from the CDR XBRL taxonomy and parsed into linear MDRM expressions (sums, differences, bounds). For the FR Y-9C, the official edit checklist was parsed the same way in v8.0. Together with the dictionary's curated identities and component hierarchies, this produced the single 7,539-relationship Relationship Registry (v10.0 added the 31 UBPR derivation rows and the conditional/compound rows promoted to testable).
Test every testable relationship against real filings.
FR Y-9C: every Y-9C relationship that resolves to codes present in the bulk data was evaluated as an aggregate query over Federal Reserve bulk BHCF (FR Y-9C) data, every quarter 1986 Q3 – 2025 Q4, 13,668 holding companies, ~208 million reported cell-values.
Call Report: all 2,464 machine-testable call relationships were evaluated against 1.917 billion rows of Call Report filings (2001 Q1 – 2026 Q1) — 1,392 sign/bound tests via one bulk scan plus 1,072 identity/bound tests over a per-code extract.
Y-9C intraseries: 63 period-over-period (YTD monotonicity) edits were tested with LAG() over the filing history (62 CONFIRMED + 1 CONFIRMED_CURRENT); a further 221 are documented as conditional (they hold only within a reporting sub-population and are not cross-sectional identities).

A relationship "passes" for a filing when the two sides agree within max($1,000, 0.1%) of the larger side — enough to absorb thousands-rounding without masking real breaks.

Run a cross-source concordance. For 16 core concepts, the Call Report value was compared against the FDIC SDI value for the same bank-quarter — an independent collection — over up to 667,551 bank-quarters each. All 16 concord at 99.5–99.97%. Two definitional mappings were corrected in the process (see below).
Adjudicate every non-passing result. Each relationship that did not pass cleanly was investigated against the official edit text, the field specification, and the MDRM — not silently dropped. The Call-side adjudication is documented in full in _rebuild/empirical/ADJUDICATION_V9.md. As in v8.0, this caught bugs in our own test harness (see Adversarial rigor below), several sparse-data artifacts, and a small set of genuine official-edit violations kept as research findings. The result is a registry in which every tested row has a defended verdict and zero rows are left in an unexplained state.

Sources. Federal Reserve / FFIEC bulk data: the bulk BHCF (FR Y-9C) data (1986–2025), the bulk Call Report data (2001–2026), and the FDIC SDI time series; the official FFIEC Call Report edit checks and calculation linkbases (CDR XBRL taxonomy, cdr.ffiec.gov); the FR Y-9C edit checklist and field specifications; and the Federal Reserve MDRM. No private or derived data is involved.

The verdict taxonomy¶

Verdict	Meaning
CONFIRMED	Holds in ≥99% of all observed filings across the full available window. A structural identity or a robust bound.
CONFIRMED_CURRENT	Holds in ≥99% of filings in the current reporting regime; older vintages differ because the form line-up was renumbered or split. Correct for today's form.
QUALITY_TOLERANCE	A regulator "should"-edit or a valid economic bound with a small, explained tail. We publish the actual rate rather than rounding it to "100%".
DATA_GAP	A genuine identity whose component splits are too sparsely populated in bulk to test as an all-plus sum (banks file the parent total but leave child line-items blank; coalesce-to-zero then breaks the sum).
NOT_IN_BULK	One side uses a code that is text/confidential/narrative or was discontinued/renumbered, so the two sides never co-occur in the bulk data. Documented, not testable.
CONDITIONAL_DOC	Applies only to a sub-population or is a conditional/non-arithmetic edit (completeness test, string value-domain, function), so it is not a cross-sectional identity. Verified by documentation.
OFFICIAL_EDIT_UNMET	An official edit that does not hold even on the correct window/population. Kept deliberately as a research finding about edit enforcement, not deleted.
CONDITION_NEVER_MET	A conditional edit whose triggering condition is never satisfied in any public bulk filing — the codes are confidential supervisory items absent from all bulk data (the RC-O PD-bucket family). Verified absent, not silently dropped.
VINTAGE_CONFIRMED	A relationship confirmed on the vintage window where its codes are actually collected, distinguished from `CONFIRMED_CURRENT` by an explicit vintage anchor.
LOW_N_PASS	Passes within tolerance but on a small observation count; flagged so the thin sample is not read as a full-history confirmation.
GRAMMAR_DOC	A conditional/compound row outside the testable grammar subset (null-logic, ratio-division, multi-branch families). Documented with its parsed form rather than forced into a linear test.

Results — the full registry¶

Status	Count
CONFIRMED	2,841
CONFIRMED_CURRENT	144
LOW_N_PASS	4
VINTAGE_CONFIRMED	7
QUALITY_TOLERANCE	43
DATA_GAP	122
OFFICIAL_EDIT_UNMET	8
CONDITION_NEVER_MET	48
NOT_IN_BULK	134
CONDITIONAL_DOC	224
GRAMMAR_DOC (documented null-logic / ratio-division families)	2,158
not_testable (documented non-arithmetic classes)	1,806
Total	7,539

By scope: 2,330 FR Y-9C, 5,162 Call Report (deduplicated across the three forms), 16 cross-source, and 31 UBPR. Zero rows are pending or unexplained. The not_testable and GRAMMAR_DOC classes are not gaps — they are the documented set of non-arithmetic edits (completeness / null tests, string value-domain checks, ratio-division and null-logic conditionals, and function edits) that cannot be expressed as a linear identity over reported values; each is labelled with its class.

Showcase: confirmed Call Report identities¶

A selection of the official Call Report edits and calculation-linkbase identities, with their real pass rates and observation counts against the 1.9-billion-row bulk Call data. Pass rate is the share of filings in which the relationship held within tolerance.

Relationship	What it checks	Observations	Pass rate	Verdict
`RCON2200 = RCON6631 + RCON6636`	Total domestic deposits = noninterest-bearing + interest-bearing	666,695	100.00%	CONFIRMED
`RIAD4074 = RIAD4107 − RIAD4073`	Net interest income = total interest income − total interest expense	669,296	99.99%	CONFIRMED
`RIAD4300 = RIAD4301 − RIAD4302`	Income before taxes ties to the pre-tax/applicable-tax decomposition	669,338	99.99%	CONFIRMED
`RCON2215 = RCON2202 + RCON2203 + RCON2213 + RCON2216 + RCONB549 + RCONB551`	Total transaction-account deposits build up from their components	667,320	99.94%	CONFIRMED
`RCON3548 = RCON3546 + RCON3547 + RCONF624`	Total trading liabilities tie to their reported components	671,820	99.13%	CONFIRMED
`(RCON3493 + RCON3494 + RCON3495) ≤ RCON1420`	Past-due/nonaccrual real-estate loans bounded by total RE loans	666,695	100.00%	CONFIRMED
`(RCONB575 + RCONB576 + RCONB577) ≤ RCONB538`	Past-due credit-card components bounded by their total	656,623	100.00%	CONFIRMED
`RCON3164 ≤ RCONA590 + 25,000`	Premises bound (official edit, with its $25k allowance)	656,623	100.00%	CONFIRMED

Each row traces to an official Call edit id (e.g. R5620.2637 across forms 031/041/051) or a calculation-linkbase parent (official_calc:RCON…), recorded in the registry's sources column.

Cross-source concordance (Call Report vs FDIC SDI)¶

Sixteen core concepts were checked between the Call Reports and the FDIC's independently-collected SDI, bank-quarter by bank-quarter. All 16 concord at 99.5–99.97%. The full table is in _rebuild/empirical/RESULTS_cross_source.csv; a selection:

Concept	FDIC variable	Call expression	Bank-quarters	Agreement
Total assets	`ASSET`	`RCFD2170` / `RCON2170`	667,551	99.84%
Total liabilities	`LIAB`	`RCFD2948` / `RCON2948`	666,523	99.55%
Total equity	`EQ`	`RCFD3210` / `RCON3210`	666,679	99.72%
Total deposits	`DEP`	`RCON2200 + RCFN2200`	659,789	99.97%
Net loans & leases	`LNLSNET`	`RCFD2122 − RCFD3123`	656,563	99.89%
Net income	`NETINC`	`RIAD4340`	666,230	99.75%
Interest income	`INTINC`	`RIAD4107`	666,245	99.87%
Noninterest income	`NONII`	`RIAD4079`	664,271	99.85%

Two corrected definitional mappings (findings)¶

The concordance is not just a pass/fail check — it surfaced two mappings that the naïve crosswalk gets wrong, and we corrected them:

Total equity → Call item 3210, not G105. FDIC EQ maps to RCFD/RCON3210 (total bank equity, excluding minority interest). The minority-interest-inclusive total G105 scored only 98.6% against EQ; 3210 scores 99.72%. The correct FDIC↔Call equity mapping therefore excludes minority interest.
Net loans → 2122 − 3123, not B529. FDIC LNLSNET maps to loans net of unearned income (2122) minus the allowance (3123), not the Schedule RC-C item-12 net total B529. The B529 basis scored only 77.8%; the 2122 − 3123 expression scores 99.89%. The two "net loans" totals are defined differently, and the SDI uses the 2122 − 3123 basis.

Both corrections are recorded in json/cross_form_mapping.json under the total_equity and total_loans concepts.

Showcase: confirmed FR Y-9C identities¶

The v8.0 FR Y-9C results carry forward unchanged. A selection, against the 208-million-row bulk BHCF data:

Relationship	What it checks	Observations	Pass rate	Verdict
`BHCK3300 = BHCK2170`	Total liabilities + equity = total assets (form item 29 = item 12)	187,845	100.00%	CONFIRMED
`BHCK4074 = BHCK4107 − BHCK4073`	Net interest income = total interest income − total interest expense	187,842	100.00%	CONFIRMED
`BHCKB529 = BHCKB528 − BHCK3123`	Net loans = loans held for investment − allowance	100,102	100.00%	CONFIRMED
HI-A equity roll-forward	The full Schedule HI-A changes-in-equity build-up	55,766	99.95%	CONFIRMED
`BHCK2170 = BHCK2948 + BHCK3210`	Assets = liabilities + holding-company equity (textbook A = L + E)	187,845	92.60%	QUALITY_TOLERANCE

The last row is instructive: the textbook identity holds exactly only on the item-29 form (BHCK3300), which ties to total assets at 100.00%. The bare "liabilities + holding-company equity" form omits noncontrolling interest and so holds in 92.6% of filings — we publish the 92.6% honestly rather than presenting the textbook form as exact.

Conditional & compound edits (v10.0)¶

Many official edits are not bare arithmetic identities — they are conditional (IF <cond> THEN <test>) or compound (T1 AND T2 AND …). Through v9 these were carried as documented-but-untested rows. v10.0 added a conditional/compound testing engine that extends the linear evaluator with three constructs while keeping the linear-only token rule (multiplication/division by a variable still hard-fails rather than being guessed at):

IF/THEN → CASE. IF <cond> THEN <test> is evaluated as a SQL CASE WHEN cond THEN (test), with the pass rate computed over the sub-population where the condition holds and the population size (n_cond) reported alongside.
AND-conjunction splitting. A compound T1 AND T2 AND … of simple comparisons is evaluated as a single all-clauses-hold row predicate; OR-compounds are evaluated as written.
Constant coefficients. Multiplication by a literal constant (e.g. the × 0 zero-conversion edit) is evaluated faithfully — the fix that grew out of the v9 dropped-multiplication catch.

Coverage. Of the 2,606 conditional/compound rows, 448 became machine-testable and were run; the remaining 2,158 stay GRAMMAR_DOC (null-logic and ratio-division families, documented with their parsed form — ratio derivations are handled properly in the UBPR work instead rather than forced through a linear test). Every newly-testable row was adjudicated to an explained verdict: 246 CONFIRMED + 61 CONFIRMED_CURRENT + 4 LOW_N_PASS, then to zero unexplained — 81 DATA_GAP, 48 CONDITION_NEVER_MET, 7 QUALITY_TOLERANCE, 7 VINTAGE_CONFIRMED, and 1 new OFFICIAL_EDIT_UNMET (the 8th overall).

Two stories make the adjudication concrete:

CECL transitions are a DATA_GAP, not a broken identity. The largest block of conditional "failures" are ASU 2016-13 (CECL) allowance and provision edits. They hold cleanly once the test is restricted to the post-adoption sub-population — a bank that has adopted CECL reports the new allowance build-up, while pre-adoption filings leave the new line-items blank and the coalesce-to- zero sum breaks. The identity is correct; the public data spans the regime change. These are classified DATA_GAP with the transition fully explained on the post-adoption subset.
The RC-O PD-bucket family is CONDITION_NEVER_MET. A family of conditional edits over Schedule RC-O probability-of-default buckets references 280 codes that are verified absent from all public bulk data — they are confidential supervisory items (large-bank PD/LGD inputs) collected but never published in bulk. The triggering condition is therefore never satisfied in any public filing. Rather than report a meaningless pass rate, these 48 rows are classified CONDITION_NEVER_MET with the 280-code absence documented.

UBPR derivation validation (v10.0)¶

The Uniform Bank Performance Report (UBPR) is the FFIEC's analytical report: it turns each bank's raw Call Report into a standardized set of derived ratios (UBPR#### concepts), each defined by an official formula over Call items and/or other UBPR intermediates. v10.0 acquired the official CDR XBRL UBPR taxonomy v181 and the FFIEC UBPR User's Guide (v181), parsed 4,207 derivation formulas with zero silent drops, and empirically validated the 31 headline ratios by recomputing each from Call/UBPR inputs and comparing to the real published UBPR value, bank-quarter by bank-quarter, over 2015 Q1 – 2026 Q1 (up to ~44,000 bank-quarters per ratio). All 31 validated at 99.77–100%. The deep companion is docs/UBPR_GUIDE.md; the concept catalogue is csv/UBPR_CONCEPTS.csv (4,099 concepts; 2,186 with derivations).

UBPR code	ratio	derivation (normalized)	agreement
`UBPRE013`	Return on assets (ROA)	`100 · RIAD4340 / UBPRD659`, annualized	99.79%
`UBPRE630`	Return on equity (ROE)	`100 · RIAD4340 / UBPRD342`, annualized	99.77%
`UBPRE018`	Net interest margin (TE)	`100 · UBPR4074 / UBPRD362`, annualized	99.77%
`UBPRE088`	Efficiency ratio	`100 · UBPRE037 / UBPRE036`	100.0%
`UBPRE006`	Provision / avg assets	`100 · UBPRD483 / UBPRD659`, annualized	99.87%
`UBPRE019`	Net loss / avg total loans	`100 · UBPR1795 / UBPRE386`, annualized	99.93%
`UBPRE541`	90+ days past due / gross loans	`100 · UBPRD667 / UBPRE131`	100.0%
`UBPRE542`	Nonaccrual / gross loans	`100 · UBPRD669 / UBPRE131`	100.0%
`UBPR7414`	Noncurrent loans / gross loans	`100 · UBPR1400 / UBPRE131`	100.0%
`UBPRE523`	Loss allowance / total loans	`100 · UBPRD095 / UBPRD146`	100.0%
`UBPRE024`	Net loans & leases / total assets	`100 · UBPRE119 / UBPR2170`	100.0%
`UBPRD486`	Tier 1 leverage ratio	regulator passthrough (±0.5 pp)	100.0%*
`UBPRD487`	Tier 1 risk-based capital ratio	regulator passthrough (±0.5 pp)	100.0%*
`UBPRD488`	Total risk-based capital ratio	regulator passthrough (±0.5 pp)	100.0%*

* capital ratios are regulator passthroughs stored 2-decimal-rounded as a fraction; they validate as the documented ×100 relationship within the ±0.5 pp rounding band.

Three findings from the UBPR work¶

The validation was not a rubber stamp; calibrating it surfaced conventions and caption errors that a naïve reading gets wrong (full method in docs/UBPR_GUIDE.md):

Units and annualization. Call dollar items (RIAD/RCON/…) are in thousands, while UBPR average/level concepts (UBPRD659) are in dollars — a ratio mixing the two must divide the UBPR value by 1,000. Year-to-date flows annualize by ×(4/q) (Q1 ×4, Q2 ×2, Q3 ×4⁄3, Q4 ×1). Both were calibrated empirically on ROA and then held unchanged across every family.
The taxonomy's presentation captions mislabel some concepts — the User's Guide is authoritative. UBPR7408 is captioned in a way that reads like a capital ratio but the User's Guide makes clear it is a 12-month growth rate ("Tier One Capital 12-month growth rate"), as is UBPRE635. They validate as parsed but must not be read as adequacy measures.
Capital ratios are regulator passthroughs, stored 2-dp rounded. UBPRD486/487/488 are not recomputed from RC-R primitives (RWA is a complex regulatory computation outside this scope); they are the regulator-reported ratios, and UBPRD48x = UBPR720x × 100 holds within the ±0.5 pp fraction-rounding band.

Intraseries (period-over-period) edits¶

A class of official edits is not cross-sectional but temporal: a year-to-date flow should be monotone non-decreasing across the quarters of a reporting year (Q1 ≤ Q2 ≤ Q3 ≤ Q4). These cannot be tested by a single-row aggregate; they require comparing a filing to the same filer's previous quarter. We tested 63 such FR Y-9C YTD-monotonicity edits with a LAG() window over the filing history: 62 CONFIRMED + 1 CONFIRMED_CURRENT, all passing ≥99.7% (the small tail is amended filings, within tolerance). A further 221 intraseries edits are documented as CONDITIONAL_DOC — they apply only within a reporting sub-population and are not unconditional monotonicity identities.

Adversarial rigor — testing our own tester¶

The v8.0 adjudication caught a harness parse-bug where A − (B + C) was evaluated as A + B + C (a dropped distributed-negation). v9.0 caught a second silent parser hazard of the same family: the side-parser tokenized only + and -, so a multiplication was silently dropped — X = Y * 0 was read as X = Y. The concrete case was the official 0%-conversion credit-equivalent edit RCFDS541 = RCFDS540 * 0, which the buggy parse evaluated as S541 = S540 (≈0.35–0.67 pass) when the official edit simply requires S541 = 0. Rewriting the affected registry expressions to the faithful form re-tested at 1.0000 (REL_5717, REL_5718). This continues the v8.0 pattern of adversarially testing our own tester: a failing empirical result is tracked to its true source before any verdict is recorded.

OFFICIAL_EDIT_UNMET — eight edits kept as findings¶

Eight official edits do not hold even on the correct window, population, and all-components- reported subset. We keep them deliberately, as research findings about how the official edit set behaves against real public data rather than deleting them. The eighth was surfaced by the v10.0 conditional engine: a second RCON-side phantom decomposition of the same family as the headline case below — an official decomposition whose child codes have no public observations, so it is documented and not enforced. The two headline cases:

A phantom decomposition. The official calculation linkbase decomposes RCONM288 into RCONL191 + RCONL192 — but L191 and L192 have zero observations warehouse-wide in the public bulk data, while M288 (51,992 filings) is reported directly. The decomposition is not empirically derivable from anything public; we document it and do not enforce it.
A securitization bound that mostly fails among the population it's about. An RC-R Part II securitization bound (R7020.6507) holds at a headline 0.983 — but that rate is dominated by zero-filers (banks with no securitization, for whom both sides are 0). Among active securitizers (the population the edit is actually about, n=3,493) it holds at only 0.452: a real, substantial explained-violation population.

The remaining five are soft "should-be" quality edits (an RC-O brokered-deposit bound at 0.71 and four RI-C/RC-C CECL loan-category bounds at 0.66–0.94) whose violation rates are themselves the finding.

Limitations (read these)¶

This validation is rigorous within its scope, and that scope has edges worth stating plainly.

The not_testable class is honest, not hidden. 4,412 registry rows are non-arithmetic official edits — completeness/null tests, string value-domain checks, conditional edits, and function edits — that cannot be expressed as a linear identity over reported values. They are labelled by class, not asserted as confirmed.
"Should" edits are not "must" edits. The 36 QUALITY_TOLERANCE relationships are valid economic bounds or regulator "should"-class edits — flagged for review but not filing-rejecting. Their published pass rates (≈0.94–0.985) are real and the small tails are explained (minority interest, netting, securitization tail, allowance-vs-balance timing), not data errors.
DATA_GAP is sparse data, not a broken identity. The 48 DATA_GAP rows are official calculation-linkbase identities that hold perfectly where the full component breakdown is reported (allcomp ≈ 1.0000) but "fail" in bulk because filers report the parent total and leave many child line-items blank, and coalesce-to-zero then makes parent ≠ partial-sum. The calculation is correct; the public data is sparse.
NOT_IN_BULK is documented, not "tested away". The 134 NOT_IN_BULK rows reference codes that are text/confidential/narrative items never published in bulk, or that were renumbered/retired across the window, so the two sides never co-occur.
Vintage matters. The 83 CONFIRMED_CURRENT relationships hold for today's forms but not uniformly back to the start of the window, because both forms have been re-laid-out many times. Each is confirmed for the regime in which its codes are actually collected; the 30 archived historical taxonomy cycles let us align edits to filing vintage.

Reproducing this¶

Everything needed to reproduce the verdicts is public:

The Relationship Registry — every relationship, its expression, source (official edit id and/or calculation-linkbase parent), status, observation count (empirical_n), pass rate (empirical_pass_rate), and test window — is in csv/RELATIONSHIP_REGISTRY.csv.
The bulk FR Y-9C (BHCF) and Call Report filings are published by the Federal Reserve / FFIEC and are downloadable and queryable via the companion FreeNIC package.
The official edits are the FFIEC Call Report edit checks and calculation linkbases, distributed in the CDR XBRL taxonomy (cdr.ffiec.gov) and the FR Y-9C edit checklist.
scripts/validate_reconciliation.py evaluates the registry against any supplied filing; use --scope {y9c,call,all,ubpr} to select the filing universe. Token validity is additionally CI-enforced by scripts/ci_audit.py, which fails the build on any invalid token.

Take any CONFIRMED row, evaluate its expression against the bulk filings with a max($1,000, 0.1%) tolerance, and you will recover its published pass rate.

A researcher-facing summary¶

For a narrative walk-through of the empirical findings — the eight official edits that don't hold in filed data, the two corrected FDIC↔Call definitional mappings, the rulebook's 25-year quintupling, and the UBPR construction conventions — see the Research Note — Empirical Findings. It distils the results on this page into a citable summary, with a one-line reproduction recipe for each finding.