Public-Data Studies: Analysis Protocol

Public copy. Internal tool and reviewer names are removed. Every rule, amendment and date is unchanged from the frozen internal file.

Status: FROZEN. This version supersedes v1, which two outside reviewers rejected (both verdict REVISE). Every change after this point is a dated amendment. Each amendment states what changed, why, and whether any results had been seen.

What had been seen before freezing:

• File structures: sheet names, labels, question and answer IDs, scope definitions and series codes.
• SBA counts by approval year and age code, and by approval year and status. No charge-off outcomes were split by group.
• Three overall figures that were already public or incidental:
  • BTOS AI use, 17.9% (published by Census);
  • the overall MOPS score, 0.549 in 2015 and 0.532 in 2021, printed by an inspection script.

No size-class, sector or group estimate was viewed.

A. Rules for all studies

1. Sources. US federal statistical works only: Census, BLS and SBA. These are not subject to copyright (17 USC 105), and the SBA FOIA files are public records.
   • Every figure cites the agency, the table or series ID, the release, and the download date.
   • Every article carries the line: “The agencies did not review or endorse this analysis.”
   • No agency logos or seals.
   • SBA row data is never republished: aggregates only, and no borrower, lender or franchise brand is named.
2. Statistics.
   • Each study has exactly ONE primary contrast, named below.
   • Alpha is 0.05, two-sided. Only 95% intervals are used; 90% is never computed.
   • Everything else is descriptive and labelled that way.
   • No multiplicity adjustment is made; that is stated.
   • The headline verb comes from the primary contrast, including when it is null.
3. No unregistered estimates.
   • Nothing appears in an article unless it is registered here, or sits in a box labelled “Exploratory, not pre-registered” with the reason it was run.
   • Exploratory content is barred from title, deck, conclusion and social copy.
4. Population identity.
   • The agency’s own unit and scope wording is quoted in each article.
   • No loose “small businesses”.
   • Size classes are named exactly.
   • Different sources never share one definition of “small”.
5. Composition caveat, mandatory where sizes are compared. “Size classes differ in industry and occupation mix; differences by size reflect both. Nothing here isolates an effect of size.”
   • No “because they are small” wording.
6. Prior-art box. Every Method section opens with “Already published by the agency: … New here: …”, backed by a dated search memo.
7. Self-serving framing.
   • Disclosure line: “Kamyar Shah sells fractional COO and CMO services.”
   • No pitch inside the findings.
   • No “businesses like yours are failing” phrasing.
   • Gaps are presented as neutral benchmarks.
   • The contact line is a single neutral sentence after Method.
8. Vintages. the claims file records file names, download dates, SHA-256 hashes, series IDs and the seasonal-adjustment vintage. Articles state “as of [release/extract date]”.
9. Gates.
   • Recompute: an independent reimplementation from the raw files, written independently by a second model and executed, must reproduce every number in the claims file.
   • Agency check: each primary number is checked against the agency’s own published table or release where one exists.
   • Outside review: two independent outside reviewers on the plan plus claims, and again on the final article.
   • Other checks: plan and article checks against the evidence, then a grammar check.
   • No shared template: each article has its own structure. Only the mandatory lines (rule 1 no-endorsement, rule 5 composition caveat, rule 7 disclosure) are standard and may not be softened.

P1 (COO, manufacturers): has the management-structure gap between small and large plants moved?

Data and verified concordance

• Census MOPS public tables, 2015 and 2021, sheet “Employment”: score and SE by establishment employment class.
• The class labels are identical in both years: 1-4, 5-9, 10-19, 20-49, 50-99, 100-249, 250-499, 500-999, 1000-2499, 2500+.
• Both years list the same 16 scored questions, and both footnotes describe the same two-step scoring rule.
• Only these two waves exist. The www2 directory lists 2015/ and 2021/ only, and no 2017 or 2019 tables exist.
• Concordance gate, run in the analysis script: the scoring-rule footnote text from both years is stored, and if it differs materially P1 is dropped.

Population: “manufacturing establishments in the Annual Survey of Manufactures mail sample”. This is stated in the title or deck. No sentence generalizes beyond manufacturing establishments.

Primary contrast (P1-PRIMARY)

• D = (S₂₅₀₀₊ − S₁₋₄)₂₀₂₁ − (S₂₅₀₀₊ − S₁₋₄)₂₀₁₅
• SE = √(sum of the four squared SEs). This assumes independent samples within and across years, which is disclosed.
• 95% CI.
• If the CI covers zero, the lede is “no clear change in the gap”.

Descriptive

• Every class score in both years, with 95% CIs.
• The within-year gap for each year.
• The 2021 Response Shares for all 16 questions, in an appendix table.
• The text may call out the three least-adopted structured answers. That selection rule is fixed here: the lowest share selecting the most-structured answer.

Removed: H-P1a (that the score rises with size). Census already published it, so it is cited as Census’s finding.

P2 (COO): Hiring, quitting and layoff rates by establishment size (BLS JOLTS)

Data: JOLTS flat files, total private, seasonally adjusted (“S” series only).

Classes: 01 (1–9), 02 (10–49) and 03 (50–249), plus a large comparator.

• The comparator is 1,000+, pooled from classes 05 and 06 using levels.
• Implied employment per class and month = level ÷ rate × 100, taken from the TS series.
• Pooled rate = sum of levels ÷ sum of implied employment.

Elements: QU, LD, HI and TS. Derived: quit share of separations = QU level ÷ TS level. JO and OS are not used.

Windows, locked:

• W1: the latest 36 months in the file downloaded 2026-10-03.
• W0: Jan 2017–Dec 2019, the pre-pandemic baseline, also 36 months.

Primary contrast (descriptive): the quits rate of 1–9 minus the quits rate of 1,000+ in W1.

• Reported with the share of months with the same sign and the monthly min–max.
• No significance is claimed. JOLTS publishes no sampling error at this level, and that is disclosed.

Descriptive: every element for every listed class in both windows, and the change from W0 to W1.

Wording:

• “Per 100 employees per month at establishments with N–M employees”.
• Never “who leaves”. JOLTS has no demographics.

Kill rule: if any class series is missing in a window, that class is dropped from the window with disclosure.

P2b (COO, separate short article): Benefit access by establishment size (BLS NCS)

Data: NCS nb series, private industry workers, access rates.

• Size classes: less than 50 workers (subcell 02) and 500 or more (07).
• The latest reference year in the file.
• Benefits:
  • medical care;
  • retirement (any);
  • paid sick leave;
  • paid vacation.
• Series IDs are fixed in the analysis script by exact title match. If a title is not found, that benefit is dropped with disclosure.

Primary contrast: the paid sick leave access gap (<50 minus 500+).

• With a 95% CI from published SEs if BLS publishes an SE series for both.
• Otherwise point estimates only, with “no published standard error” disclosed.

Wording:

• “X% of private industry workers in establishments with fewer than 50 workers had access to…”. This is a worker share, never a share of establishments.
• P2 and P2b never cite each other in the title or lede, and no causal link is drawn.

P3 (CMO/COO): What AI users of different sizes changed, and why non-users hold off (Census BTOS AI supplement)

Data: BTOS AI supplement 2025–2026, collected Nov 17, 2025 to Feb 8, 2026. Sheets “Employment Response Estimates” and “Employment Standard Errors”.

• Size classes, verbatim:
  • A = 1 to 4;
  • B = 5–9;
  • C = 10–19;
  • D = 20–49;
  • E = 50–99;
  • F = 100–249;
  • G = 250 or more employees.
• Scope (Census data dictionary):
  • 1 = all companies;
  • 2 = companies that used AI in any business function;
  • 3 = companies using AI to perform tasks previously done by employees;
  • 4 = generative-AI users;
  • 5 = companies planning to use AI;
  • 6 = companies NOT planning to use AI in the next six months.
• Every share prints its scope.

Primary contrasts (two, both reported; α = 0.05 each):

• P3-PRIMARY-1: Q7 Answer 1, “Trained current staff to use AI” (scope 2, AI users), class A minus class G.
• P3-PRIMARY-2: Q13 Answer 3, “Lack of knowledge on the capabilities of AI” (scope 6, non-planners), class A minus class G.
• z = difference ÷ √(SE_A² + SE_G²), assuming zero covariance between disjoint domains (disclosed).
• The estimate column is the one the sheet populates for that answer: “Estimate” or “Estimate – Yes”. The column used is recorded in the claims file.

Descriptive (registered): by all seven classes, with SEs:

• Q2 functions and Q3 uses (scope 2);
• Q4 task substitution (scope 3);
• Q6 employment effect;
• every answer to Q7 (scope 2) and to Q13 (scope 6);
• Q1 use, as context citing Census.

Suppression: a comparison with any required cell marked “S” is reported as suppressed. There is no pooling, no class substitution, and no borrowing from the 2024/25 supplement.

No cross-wave comparison.

Q6 wording rule: “businesses reporting that AI use [increased/decreased] employment (self-report)”. No causal headline.

Census scope sentence quoted: “representative of all employer businesses in the U.S. economy, excluding farms”.

P4 (exit planning): Five-year charge-off for SBA 7(a) loans to buy a business vs loans to existing businesses

Data: SBA 7(a) FOIA files as of June 30, 2026 (FY2010–FY2019 and FY2020–Present), plus the data dictionary.

Verified codebook facts

• The BusinessAge value “Change of Ownership” first appears in FY2018. Earlier cohorts cannot identify purchases, so the window is FY2018–FY2020.
• LoanStatus values: PIF, CHGOFF, EXEMPT (disbursed and open, status withheld), CANCLD and COMMIT.
• ChargeOffDate and GrossChargeOffAmount exist.

Estimand: five-year cumulative charge-off incidence. A loan counts as an event if its ChargeOffDate is no more than 1,826 days after its ApprovalDate.

• Every FY2018–FY2020 loan has at least 5 years of follow-up by the extract date, so no loan is censored before day 1,826.
• Open (EXEMPT) and paid loans count as non-events.
• Excluded:
  • CANCLD and COMMIT (never disbursed);
  • BusinessAge Unanswered or blank. Their count and share per cohort are disclosed.

Groups

• CoO = “Change of Ownership”.
• EXISTING = “Existing or more than 2 years old”, “Existing, 5 or more years”, “Less than 5 years old but at least 4”, “Less than 4 years old but at least 3” and “Less than 3 years old but at least 2”.
• STARTUP = “Startup, Loan Funds will Open Business”. Context only; this comparison is prior art.

Primary contrast (P4-PRIMARY): CoO vs EXISTING five-year count incidence, all sectors, FY2018–FY2020 pooled. Wilson 95% CI for each, plus a difference CI (Newcombe).

Registered descriptive

• By 2-digit NAICS sector where each arm has at least 1,000 loans. If fewer than 3 sectors qualify, only the all-sector result is shown.
• CoO franchise vs non-franchise, all sectors, if each arm has at least 500 loans.
  • Franchise = FranchiseCode not in {blank, 0, 00000}.
  • The misclassification risk is disclosed.
• Dollar version for the primary contrast: charged-off dollars ÷ approved dollars, with a 2,000-draw bootstrap percentile 95% CI.
• Loan-size bands (<$150k, $150k–$500k, $500k–$1M, >$1M) for CoO.

Mandatory disclosures

• These are SBA-guaranteed loans only.
• A charge-off is a loan outcome, not business failure.
• The FY2020 cohort received CARES Act Section 1112 payment relief. A sensitivity on FY2018–FY2019 only is registered and reported.
• Rates are not adjusted for lender clustering.
• Prior art: acquisition vs startup is already published elsewhere.

Kill rule: if fewer than 2,000 CoO loans remain after exclusions, P4 is dropped.

Order

1. Analysis scripts write the claims file for P1, P2, P2b, P3 and P4.
2. Independent recompute.
3. Agency checks.
4. Outside review on claims plus plan.
5. Plans and articles.
6. Plan and article checks against the evidence.
7. Final outside review.
8. Grammar check.
9. Author approval. Nothing is published before that.

Change log v1 → v2

• 90% intervals dropped.
• A single primary contrast per study; an exploratory box rule added.
• P1: concordance verified; difference-in-differences primary with stated SE formula; H-P1a removed; all 16 questions shown.
• P2 split into P2 (JOLTS) and P2b (NCS).
• The unpublished 1,000+ class is pooled correctly from levels.
• JOLTS tests replaced with a descriptive rule.
• Windows locked.
• P3:
  • size classes and scopes pasted verbatim;
  • two named primaries;
  • suppression rule;
  • Q6 wording rule;
  • no cross-wave comparison.
• P4: rebuilt as five-year cumulative incidence, because resolved-only rates were biased. The window moved to FY2018–FY2020 because “Change of Ownership” does not exist earlier.
  • A reviewer’s FY2010–2014 suggestion was not feasible.
  • A reviewer’s claim of 2017/2019 MOPS waves was checked and is false.
• Rules added: prior-art box, self-serving framing, vintage pinning and legal attribution.

AMENDMENT 1 (2026-10-03, before any size or group estimate was computed)

Change. The P1 call-out rule is replaced.

• Old rule: “lowest share selecting the most-structured answer”.
• Why it was dropped: it requires a per-question mapping to MOPS’s scoring order that Census does not publish in the tables. For example, it is ambiguous for the display-board question.
• New rule: the call-out names the three practices with the HIGHEST 2021 share choosing the explicit absence answer, from this fixed list:

Results had not been seen. Only answer texts had been read.

AMENDMENT 2 (2026-10-03, mechanical lookup fix)

Change. For P2b, the retirement series title in the BLS file reads “access to all retirement benefits”, not “access to retirement benefits”. The exact-title match now uses the BLS wording.

Results seen. The other three benefits’ values had been printed. The retirement value had not.

Why this is not outcome-dependent. It is a lookup correction of the series title. The estimand is the same.

AMENDMENT 3 (2026-10-03, after results were seen, during gate 4 and gate 6 checks)

(a) Contradiction resolved: P3 primaries. Rule A2 says one primary contrast per study. Section P3 registered two primary contrasts before any estimate was computed.

• The P3 section governs: P3 has two co-primary contrasts.
• Both are reported.
• No headline uses the null co-primary as if it were positive.
• This resolves a contradiction inside the frozen text. It is not a new comparison.

(b) P1 appendix scope. The article’s appendix lists the 2021 absence-answer share for each of the 12 questions in Amendment 1. It does not reproduce all 90 response rows. The full 16-question distribution is cited to the Census table.

• Reason: readability.
• Results: had been seen.
• Effect: no figure is added or removed relative to the registered call-out rule.

(c) P1 standard error. Census publishes a standard error of 0 for the 2,500-or-more class in both years. It is used as published, and the article discloses it.

AMENDMENT 4 (2026-10-03, after results were seen)

P4 loan-size band boundaries. The registered labels were “<$150k, $150k–$500k, $500k–$1M, >$1M”. They were implemented, and verified in gate 2, as half-open intervals:

A loan of exactly $1,000,000 therefore falls in the top band. The article uses these exact boundaries.

Status. The registered text was ambiguous at the boundaries. The implementation is unchanged.

Disclosure. This amendment is disclosed because the results had been seen.

AMENDMENT 5 (2026-10-03, after results were seen; wording correction raised by an outside reviewer)

The registered disclosure was incomplete. It said “The FY2020 cohort received CARES Act Section 1112 payment relief.” Section 1112 payments also covered 7(a) loans already in regular servicing in 2020, including loans approved in FY2018 and FY2019.

What changes

• The article states that no cohort is free of the relief.
• The FY2018–FY2019 sensitivity is still reported because it was registered, but not as a remedy.

What does not change: no estimand and no figure.

AMENDMENT 6 (2026-10-03, after results were seen; from a second outside review)

(a) P2b interval, applying the registered rule. BLS publishes standard errors for the March 2026 benefit estimates in a separate table set (“Employee Benefits in the United States, Standard Errors for News Release”, Table 6 for paid leave, Tables 1 and 2 for retirement and medical care). The flat files used for the estimates do not carry them. The registered rule (“a 95% CI from published SEs if BLS publishes an SE series for both”) therefore applies. The primary interval is computed from those published SEs as difference ± 1.96 × √(SE₁² + SE₂²), assuming zero covariance between the two size classes, which is disclosed. Private industry SEs used: paid sick leave 1.3 (1 to 49) and 1.0 (500 or more). The other three gaps get the same calculation as descriptive figures. Results had been seen.

(b) P2 exploratory box. Quit rates for all six published BLS size classes, both windows, from the published seasonally adjusted rates (including 250 to 999, which the design left out). Reason: an outside reviewer noted that the registered comparison is an extreme-group design and that the pooled 1,000-or-more group mixes two classes with very different rates. Shown only in a box labelled “Exploratory, not pre-registered”, and kept out of the title and deck.

(c) P4 exploratory box. Days from approval to first disbursement (SBA field FirstDisbursementDate) by group, and a five-year charge-off rate counted from first disbursement for loans first disbursed by June 30, 2021. Reason: outside reviewers raised the possibility that change-of-ownership loans disburse later, which would shorten their time at risk under the approval clock. Exploratory box only.

(d) Wording, with no change to any estimand.

• P1 absence call-outs. Two of the 12 registered answers are “Rarely or never”, so the heading reads “lacking or rarely using”.
• P1 waves. MOPS began as a 2010 pilot. The published tables used here cover 2015 and 2021, and the line “Only these two waves exist” in section P1 refers to published tables.
• Interval labels:
  • P1 and P3 intervals are labelled approximate under the registered zero-covariance assumption.
  • P4 intervals are described as model-based uncertainty for loans of this kind, not as sampling error for the listed loans.
• P4 dollar measure. It is named a ratio (charged-off amount ÷ approved amount), and overlap of two intervals is no longer used as an inference.

(e) Public copy. A public copy of this protocol is prepared with internal tool and reviewer names removed. Every rule, amendment and date is unchanged.

AMENDMENT 7 (2026-10-03, after results were seen; from a third outside review)

(a) P4 exploratory addition.

• What was added: the five-year charge-off rate, counted from approval, for the 20,446 loans excluded for blank or Unanswered business age, plus their median approved amount, plus an extreme bound (all excluded loans assigned to change of ownership).
• Reason: outside reviewers raised selection risk from the exclusion.
• Placement: exploratory box only.

(b) Wording, with no estimand changed:

• P1 response shares are stated as Census’s “Percent of Respondents”.
• The P1 SE formula is described as a zero-covariance assumption.
• P2 quit share is renamed the quit-to-separations ratio, described as approximate because seasonally adjusted series are not additive.
• P2b states that the four benefits were fixed in this protocol before any figure was computed.

AMENDMENT 8 (2026-10-03, after results were seen; author decision after outside review)

P2 framing.

• The six published BLS size classes are promoted from the exploratory box to the lead descriptive result, and the title and deck now describe that pattern. This applies to quit rates only, for both windows.
• The pattern is supported by the 36-month averages and by a month-by-month count:
  • In all 36 months of each window, the highest of the six class rates was at 10 to 49 or 50 to 249 employees.
  • In all 36 months of each window, the lowest was at 5,000 or more.
• Why. Two outside reviewers judged that leading with the registered extreme-group comparison against a constructed pooled class would misrepresent the published data, because the pool is dominated by the 5,000-or-more class.
• What does not change. The registered primary comparison (1 to 9 minus pooled 1,000 or more, W1) is unchanged and is reported in full, with its sign count and range. No significance is claimed for anything.
• Disclosure. The article states that this framing was adopted after the results were seen.
• Rule override. This amendment overrides rule A3’s bar on exploratory content in the title and deck, for P2 only, and the override is disclosed.

AMENDMENT 9 (2026-10-03, after publication; author decision)

Guidance sections.

• What changes. Each article adds sections on what the figures mean and how an owner might act on them, followed by a “What this data does not tell you” list.
• Labelling. Each guidance section is labelled as interpretation and guidance from operating practice, not a finding, and states that the study did not test the recommendations and that the agency does not endorse them.
• Numbers. Numbers used in guidance come only from figures already reported in the article or from registered descriptive tables.
• External facts. Statements about laws, plan types, or definitions point the reader to a professional or to applicable law.
• Review standard. Reviewers check guidance only for three things: misused numbers, cause-and-effect claims attributed to the data, and legal, tax, or HR statements presented as settled fact. Findings sections are unchanged and keep the full standard.

Correction to Amendment 7(a), recorded under Amendment 9.

• Amendment 7(a) called the all-to-change-of-ownership reassignment an “extreme bound”. It is not a worst-case bound, because group membership among the excluded loans could be related to their outcomes.
• The article now calls it a simple reassignment scenario and says it is not a worst case.

Wave 2 (frozen October 4, 2026): owners, injuries, separations, new-hire earnings, S corporation income and family ownership

Public copy. Internal tool and reviewer names are removed. Every rule, amendment and date is unchanged from the frozen internal file.

Status: FROZEN October 4, 2026.

• This version revises a first draft that two outside reviewers returned for revision.
• Every later change is a dated amendment that states whether results had been seen.
• Before freezing, only file layouts and labels had been inspected. Two kinds of totals were printed incidentally during that inspection:
  • IRS 2022 header rows;
  • the all-industries size-class injury rates and totals that BLS had already published.
• No study estimate was computed or viewed before freezing.

Rules for every wave 2 study (in addition to the wave 1 rules)

1. One registered comparison per study. Everything else is descriptive and published in full, including null, tied, reversed and unavailable results. No industry, year or class is highlighted because its value is extreme.
2. Intervals only where valid. A 95% interval is shown only for a quantity the agency publishes with a standard error. Ratios, sums and differences built from separately published cells are point estimates. Each says that a covariance-aware interval is not available from the public tables.
3. Missing is never zero. Suppression and nonpublication markers, and QWI status codes other than 1 or 6, are treated as missing. If a registered cell is missing, the comparison is “not estimable” and is never silently replaced.
4. Evidence boundary before guidance. Every article places this sentence immediately before its guidance: “These are descriptive aggregates for the stated federal-data population. They do not identify causal effects, diagnose an individual company, or estimate the result of adopting a practice. The guidance that follows is operating judgment, not a finding from the data.”
5. Title rules. A title must pass all of these, and accuracy is pass or fail, not a score:
6. name the correct unit (owners, firms, establishments or S corporation returns);
7. keep every qualifier that is essential to the truth;
8. use a number only if it is the registered comparison, with its comparison stated;
9. avoid causal verbs, “typical”, superlatives beyond the data, forecasts and prescriptions.

At least five candidate titles are drafted after the findings are fixed. The reason for each choice is recorded.

6. Agency wording. Each article quotes the agency’s own scope and unit wording, verified against agency documentation.

Studies and registered comparisons

• W1. How owners’ previous businesses ended.
• Data: Census Annual Business Survey, owner characteristics, reference year 2023. Unit: owners of responding employer firms.
• Registered comparison: among owners who report a previous business they started, the share whose previous business was purchased (by an individual or a company) minus the share whose previous business is no longer operating. Point estimates.
• W2. Where injury rates peak across establishment-size classes.
• Data: BLS Survey of Occupational Injuries and Illnesses, 2024, private industry.
• Analysis set: the most detailed non-overlapping industries, fixed from labels and codes before any rate was read.
• Registered comparison: among industries with all five size-class rates published, the count whose single highest total recordable case rate is at 50 to 249 or 250 to 999 employees, with the full distribution and ties shown.
• W3, W5 and W7. Census Quarterly Workforce Indicators.
• Scope: national, private, all industries, not seasonally adjusted.
• Window: the latest four quarters, at or before 2025Q3, in which every value used has status 1 or 6. Baseline 2017 to 2019.
• Each comparison is the mean of four quarterly contrasts, with all four reported.
• W3: separation rate, firms aged 0 to 1 year minus firms aged 11 or more.
• W5: stable-job separation rate, firms with 0 to 19 employees minus firms with 500 or more.
• W7: average monthly earnings of new hires into stable jobs, firms aged 0 to 1 divided by firms aged 11 or more.
• W4. Where the receipts dollar goes in S corporation returns.
• Data: IRS Statistics of Income, Table 3.2, tax year 2022.
• Registered comparison: aggregate net income less deficit from a trade or business, as a share of aggregate business receipts, for returns with $1 million to under $2.5 million in receipts against $10 million to under $50 million. Years 2014 to 2021 are descriptive.
• W6. Majority family ownership by firm size.
• Data: Census Annual Business Survey, business characteristics by employment size, reference year 2023.
• Registered comparison: the published majority family-owned share at 1 to 4 employees against 250 to 499 employees. Each share has its published interval; the difference has none.

Amendment 1 (October 4, 2026, before any study estimate was computed)

• (a) W1 question. The question asks about the status of the previous business the owner started most recently, not including the current business. It allows one answer, so the comparison is unchanged.
• (b) W6 denominator. Census’s percentages for this table were confirmed, using a question outside the study, to use the firms that answered as their denominator. The registered comparison therefore uses the published percentages.
• (c) W6 definition. Census counts a firm as majority family-owned only when two or more members of one family own the majority. A one-owner firm is therefore not counted, and every article says so.
• (d) W2 analysis set. The set was fixed from labels and codes before any rate was read:
• 1,020 private-industry rows with a six-digit industry code;
• less 31 rows that BLS footnotes as drawn from mine-safety or railroad regulators, which BLS states are not comparable;
• leaving 652 non-overlapping rows.

Amendment 2 (October 4, 2026, after results were seen; raised by outside review)

W4 common-size statement. The protocol called for “the full common-size statement”. The analysis code, written before any ratio was computed, covered only these lines, and no other line was computed or viewed:

• cost of goods sold;
• compensation of officers;
• salaries and wages;
• rents paid;
• advertising;
• employee benefit programs;
• net income less deficit;
• two sums of those lines.

“Full” is corrected to describe exactly those lines.

Amendment 3 (October 4, 2026, after results were seen)

W5 was completed but is not published as an article.

• Every check passed.
• The registered result: the stable-job separation rate at firms with 0 to 19 employees minus firms with 500 or more averaged 1.4 percentage points over 2024Q3 to 2025Q2. By quarter it was 1.1, 3.5, 0.9 and 0.1.
• The author decided not to publish an article on a contrast this small and unstable. The result is stated here so that the wave is not selectively reported.

Checks completed (October 4, 2026)

• Independent recompute. Every published figure was recomputed by an independent auditor writing its own code from the raw files. Every difference was traced to its cause and recorded. Where the auditor’s own code was wrong, the correction is logged with evidence. All seven studies match on every published number.
• Agency cross-checks.
• The W1 and W6 figures match the Census data service exactly.
• The 2023 W2 rates match those BLS published.
• The W3, W5 and W7 quarterly results were rebuilt from state-level Census figures and agree with the national files to within 0.1 point, or 0.7 point for the earnings ratio.

Wave 3 (frozen October 4, 2026): loan charge-offs beside firm deaths, advertising shares beside job growth, and separations beside job destruction

Public copy. Internal tool and reviewer names are removed. Every rule, amendment and date is unchanged from the frozen internal file.

Status: FROZEN October 4, 2026.

• Wave 3 was designed to answer a criticism of waves 1 and 2: that each article republished one agency table. Every wave 3 headline is built from two federal sources joined on a sector, firm-age or year key, and no agency publishes the resulting figure.
• The plan was red-teamed before drafting, and the draft protocol went through three rounds of outside review before freezing. Two studies in the original plan were cut before freezing: one because the owner-hours question it needed no longer exists in the Census owner tables, and one because wages are not published by employer size, so the study could only have been a scenario. A measurement instrument and a distribution step were also removed before freezing as outside the scope of a publishing protocol.
• Before freezing, file layouts, variable lists and status codes had been inspected. One feasibility query returned values rather than structure: national Business Dynamics Statistics rows by firm age for 2023 (job destruction and death rates), the all-sector net job creation row, two sector rows, and one construction-sector age-1 row nationally and for California. These are the Census side of W3-5 and part of the growth side of W3-3. The SBA, IRS and QWI sides, and every registered comparison, were unseen.
• The raw Census API responses the studies use were retrieved and archived with hashes before freezing and are the controlling data; the live API was not re-read for any registered quantity.

Rules for every wave 3 study (in addition to the wave 1 and wave 2 rules)

1. Buyer question first. Each study is registered with the question an owner is asking when deciding whether to hire operations or marketing leadership, phrased so the registered comparison can answer it.
2. Pre-freeze checks are part of the protocol, not amendments. Prior-art searches, crosswalks, variable definitions, eligible-cell sets, age-code maps and source vintages were completed and archived before freezing.
3. Fixed sources, archived. Every study names one agency release, reference period, file or API vintage, download URL and extraction date. The raw API responses are archived and controlling.
4. Precision follows the join. A headline built from two sources with different units is a range, a rank or a side-by-side, never a point estimate at a precision the join cannot support, and never a quotient of unlike events. Every unit mismatch is stated in the sentence that carries the number.
5. One registered comparison per study. Everything else is descriptive and published in full, including null, reversed, suppressed and unavailable results, with every sector or class marked available, limited coverage, suppressed or not comparable.
6. No comparison to the author’s service. No study computes or implies the cost, value or result of fractional work. The disclosure line appears immediately below the byline and again with the contact line.
7. Outside-review stopping rule. Review ends when the sentence gate is clear, one outside reviewer from a different model family returns a publish verdict, and no unresolved blocking or major finding remains. Later findings about data validity, denominators, disclosure, legal substantiation, privacy or a false agency claim reopen the article.
8. Intervals only where valid. The SBA file, the BDS and the QWI carry no sampling error. IRS sole proprietorship figures are sample estimates with no coefficient-of-variation table for the income statement used, so shares are shown with a stated caution and no interval.
9. Missing is never zero.
10. Title rules as in wave 2, with the unit and denominator in the title. Five candidates per article after the findings are fixed.
11. Mandatory unit quote as in wave 2.
12. Publish gate as in waves 1 and 2.

Studies and registered comparisons

• W3-4. Where SBA loans to buy a business charged off, beside where firms die.
• Data: SBA 7(a) loan-level FOIA files as of June 30, 2026; Census Business Dynamics Statistics, firms and firm deaths by sector, 2019 to 2023.
• Cohort: non-cancelled loans approved October 1, 2017 through September 30, 2020 with the business-age code “Change of Ownership”. A loan is charged off within five years when its charge-off date is on or before the fifth anniversary of approval; a later charge-off scores zero. Inconsistent records are set aside and counted. If inconsistent records exceed 1% of eligible loans, the study is dropped pending an amendment.
• Registered comparison: Spearman rank correlation across sectors with at least 200 eligible loans and complete Census figures, between the five-year charge-off rate and the mean annual firm death rate 2019 to 2023. Descriptive and ecological; no probability value; no quotient between loan events and firm events. Dropped if fewer than eight sectors qualify or fewer than 2,000 loans remain.
• W3-3. Advertising dollars per dollar of receipts on sole proprietorship returns, by sector, beside the sector’s employer job growth.
• Data: IRS Statistics of Income, nonfarm sole proprietorship returns, tax year 2023, Table 2 income statements by sector (and Table 1 return counts); Census BDS net job creation and denominator by sector, 2021 to 2023 (2020 excluded as a pandemic outlier, stated with the result).
• The share is advertising dollars divided by aggregate receipts on sample-weighted returns: receipts-weighted, not a typical return’s share, not a size benchmark.
• Registered comparison: Spearman rank correlation across matched sectors between the 2023 advertising share and the 2021 to 2023 net job creation rate. No causal reading and no temporal direction. Dropped if fewer than ten sectors match.
• W3-5. Worker separations and job destruction at firms aged 0 to 1, side by side.
• Data: Census QWI, 2023 quarters 1 to 4, private firms aged 0 to 1 (and 11 or more as a contrast), separations and beginning-of-quarter employment with status flags; Census BDS 2023 by state, sector and firm age (ages 0 and 1 summed; 11 and over), job destruction, job destruction from closing establishments, and the Census denominator.
• Frames: for all industries, the 49 jurisdictions with usable QWI cells in all four quarters and matching BDS cells (Alaska and Michigan returned no QWI data). Each sector has its own complete-case frame; sectors with fewer than 40 jurisdictions are marked limited coverage.
• Registered comparison: three separate measures reported side by side, the QWI annual separation rate per 100 average beginning-of-quarter jobs, the BDS job destruction rate and the BDS closures-only rate. They are not subtracted, no ratio is formed, and no remainder is described, under any condition.

Amendment 1 (October 4, 2026, after results were seen; raised while building the registered subsector appendix for W3-3)

• The IRS Table 2 column “Administrative and support and waste management and remediation services” includes “Management of companies and enterprises” among its subsectors, so its Census counterpart is sectors 55 and 56 together, with the Census rate recomputed from the two sectors’ summed counts under the frozen rule for grouped sectors. The pre-freeze crosswalk had mapped the column to sector 56 alone. The registered rank correlation moved from minus 0.10 (wrong mapping) to minus 0.09; both are reported and the null is unchanged.
• Table 2 has three header tiers, not two. The registered descriptive appendix covers every labelled column: 153 rows, two suppressed in the IRS file.
• Both changes were recomputed by a freshly generated independent program and matched.

Checks completed (October 4, 2026)

• Independent recompute. Every published figure was recomputed from the raw files by separately written programs built without access to the studies’ results. The first programs failed on file paths and were patched only for paths; the patches are logged. All 25 original claim groups and the 12 amended W3-3 claim groups matched.
• Sentence gate. Each article passed a sentence-by-sentence check against the claims, the protocol, the recompute record and the verified agency quotes on its final text.
• Outside review. Each article received a publish verdict from an outside reviewer and had every blocking or major finding from the other reviewers applied. Material corrections included wrong sector rankings in a first draft, ten shares rounded from rounded values instead of raw ratios, and the W3-3 crosswalk correction above.
• Grammar check and full human read before publication.

Title amendment, all waves (October 5, 2026, after publication; author decision)

The page title (H1) of every article in waves 1, 2 and 3 was shortened to its search title, which passed the same title gate (unit and denominator named, no causal verb, no banned word). The full registered titles were 20 to 42 words long and filled the first screen of each article. They remain in the protocol record and in each article’s deck and method; no body text, figure or cover changed. The fourteen page titles now read: