Constructing the record

Revision, 2026-09-11. Two external reviews found this chapter's methods discussion complete on everything except the design itself: it stated how the corpus was built and what is wrong with it, and never stated why this case, or what one case can establish. Section VII gains a passage separating reproducibility from validity — the exhibits guarantee a figure follows from the database and say nothing about whether the coding behind it is sound — and section IX gains limitations 9 (case selection) and 10 (the null result is consistent with, not evidence of, the explanation offered). No figure changes; the counts remain as of 2026-08-28.

Revision, 2026-09-06. Two things this chapter reported as outstanding have been done, and saying so is part of what the chapter is for. The appellate layer is now first-class — section II is rewritten and thesis/exhibits/appellate_layer.py measures it — and all three of the canonicalisation defects named in section IX have been repaired, along with six more of the same kind found while repairing them. Section IX's limitation 8 gains the reason those repairs matter to a reader: the corpus grows by discovery as well as by filing, so a figure in this chapter cannot be recomputed from a later database by filtering on dates. Every count in the body is left at its 2026-08-28 value and is dated rather than wrong.

I. The claim this chapter has to earn

Chapter 1 asserted that ninety-two originating actions exist and that no court has counted them. Chapter 2 argued that no American filing-restriction mechanism is built to. Both claims depend entirely on a prior one, which has so far been asserted and not defended: that ninety-two is the right number.

That is not a trivial thing to establish, and the ways of getting it wrong all inflate it. Count a removed case twice and the total climbs. Count an appeal as a new action and it climbs further. Count a caption change, a consolidated companion, or a case filed against this litigant, and it climbs again. Every one of those errors pushes in the direction the argument wants, which is exactly why the method has to be stated before the number is used.

This chapter states it: what was collected, what could not be, how docket records were collapsed into actions, how party names were resolved, how evidence of different kinds was kept apart, and how the one coded variable the argument leans on was defined and checked. It ends with the limitations, at length, because a methods chapter that reports only the parts that worked is not a methods chapter.

II. Collection, and what is not in it

The corpus holds 102 docket records across five trial forums in three states.

Appeals are not counted among those records, and since 2026-09-06 they are no longer buried in prose either: the appeals table holds one row per appellate proceeding, keyed to the trial docket it was taken from. It now holds 30 appellate dockets arising from 22 trial-court docket records — 12 in the Third Circuit, 5 each in the Superior Court of Pennsylvania and the Ninth Circuit, 3 each in the Second Circuit and California's Second Appellate District, and 2 whose court is recorded but whose number was never recovered. It held 27 from 20 when the table was built on 2026-09-06; the verification pass of 2026-09-10 found three more, none of them newly filed except one docketed the day before.

Keeping them out of the docket-record count is a modelling decision, not an oversight. A removal keeps both dockets as cases rows joined by SameActionAsCaseID, because both are trial-level records of one action; an appeal is a different grain, and folding it in would inflate the published record count without a new lawsuit existing. The site therefore names three grains and never mixes them — originating actions, trial-court docket records, and appellate dockets — and this chapter uses the middle one throughout.

The layer is measured by python thesis/exhibits/appellate_layer.py, and its first result is worth stating here because it bears on chapter 4's question about how one dispute multiplies into many records. Not one of the 30 crosses a sovereign boundary. Lackawanna appeals go to the Superior Court of Pennsylvania, federal appeals to the circuit for the district the case sits in at the time (following a transfer where one happened, as with C046 and C058, both of which moved to California districts and both of which appeal to the Ninth Circuit), and the Los Angeles Superior Court appeals to California's Second Appellate District. The appellate layer adds docket records without adding cross-forum reach.

Verification followed on 2026-09-10. 19 of the 30 have now been checked against the reviewing court's own docket, up from 7; 8 carry a recorded docket number whose current status has not been independently confirmed, and 3 record an appeal established to exist whose number was never recovered — those keep a blank number rather than a guess. Posture counts drawn from this table are therefore still a floor, but a much higher one.

The pass is also its own small lesson about where appellate records live. Pennsylvania's Unified Judicial System portal publishes a full appeal docket sheet, disposition text and all, for anyone who asks; the federal side publishes almost nothing free. CourtListener's mirrors of the circuit dockets carry the case-opening entries and stop, so a federal appeal's disposition had to be read from the district docket below, where the mandate is entered when it returns — and only where the district mirror was current enough to have received it. Eight of the eleven appeals still unverified or unnumbered are federal for exactly that reason.

It was assembled from free public sources. There is no PACER account behind it, which is a methodological choice with a cost: PACER charges per page, and a corpus of this size read fully through PACER would run to a significant sum, so the project reads what CourtListener, GovInfo, the RECAP archive, the Los Angeles County Court Access Portal, New York's iApps e-filing portal and the Lackawanna County Prothonotary's own portal publish for free. Where a document could not be obtained, the corpus says so in the row rather than paraphrasing from an index.

The cost of that choice is unevenly distributed, and it shapes what the previous chapters could say. Of 502 citations carrying a URL, 80 — 15.9% — still point at hosts a reader cannot open: DocketBird (20), Justia dockets (19), PacerMonitor (17), Casemine (11), Trellis (8), Rulings.law (3), Docket Alarm (2). The remaining 84% resolve on CourtListener (143 plus 11 RECAP documents), the Lackawanna portal (90 on ic-access.com, 48 on lpa-homes.org), New York's iApps (53), GovInfo (46) and the Los Angeles portal (14).

The asymmetry runs along the forum boundary that Chapter 2 is about. Federal opinions are free, full-text and searchable. Lackawanna County publishes a document list, one case at a time, behind a session-based portal; reading a Lackawanna order means opening that case's own page and retrieving that document. The consequence shows up in the evidence density: across 37 M.D. Pa. originating actions the corpus records 21 filing-restriction indicators, 10 of them Strong, and across the larger block of 38 Lackawanna actions it records 7, two of them Strong. That ratio is a fact about document accessibility, not about conduct, and any chapter that uses indicator counts comparatively has to say so.

III. De-duplication: the originating action

The unit that matters is not the docket record. It is the originating action — one underlying lawsuit, however many dockets it has passed through.

The mechanism is a union-find over the SameActionAsCaseID column on cases. Where one record is marked as the same action as another, the two are merged; the merged action is counted once and dated from its earliest filing. Both records survive in the corpus and both keep their own case page, because a reader auditing the claim needs to see the state docket and the federal docket separately. They are one action only for counting.

Eighteen records carry a SameActionAsCaseID value, collapsing into ten actions that span more than one docket record. All ten are Lackawanna County complaints removed into the Middle District of Pennsylvania: C011/C052, C013/C054, C014/C053, C015/C057, C018/C058, C030/C064, C032/C074, C034/C065, C035/C066 and C036/C068.

102 records − 10 duplicated pairs = 92 originating actions.

Two further reductions matter for any claim about what this litigant did, as opposed to what exists. Four of the 102 records are actions brought against him — C100 (Saladino v. Tufano, S.D.N.Y.), C022 (Fidelity Bank v. Free Range Meat LLC) and C031 and C037 (both American Express National Bank v. Tufano, Lackawanna). So 88 of the 92 actions were commenced by this plaintiff, and that is the denominator the repeat-filing variable uses. Direct appeals are not separate actions at all; they are recorded in the parent case's status.

This is the same de-duplication the public site applies. tools/build_seo.py implements it, assets/app.js reimplements it in JavaScript for the legacy dashboard, and thesis/exhibits/detection_gap.py implements it a third time for the thesis. Three implementations of one rule is a maintenance hazard and is named as such in CLAUDE.md; the thesis's protocols import the third rather than adding a fourth.

IV. Canonicalisation: 710 appearances, 334 parties

Defendant names arrive as free text in a semicolon-separated field, which means the same company appears under several spellings across several complaints.

Across the 102 records there are 710 defendant appearances, carrying 366 distinct name-as-filed strings, which resolve through defendant_aliases to 334 canonical parties. A separate attorneys table records the six lawyers and law firms that appear as parties rather than as counsel — Jeffrey Davis, David D. Lin, Lavely & Singer, Kjar McKenna & Stockalper, Lewis & Lin, and Cipriani & Wener (names given as the attorneys table records them, A001–A006).

Canonicalisation is where the corpus's most consequential defects live, and Part VIII returns to them. Three are known and unrepaired.

V. Five evidence layers, kept apart

The corpus does not have a table called "evidence." It has five categories that are structurally separated, and the separation is a hard constraint on the site's generator and on this thesis, not a presentational preference.

A court finding is recorded only when it can be tied to an order, opinion or report and recommendation. That rule is why the two vexatious-litigant requests live in vexatious_motions and not in court_findings: both asked a court to declare this litigant vexatious, and neither obtained a ruling, so neither is a finding of anything.

The secondary_sources layer is the smallest and the best illustration of why the separation is not fussiness. Both of its two rows are blog posts by Jeffrey Davis, a lawyer who represented this litigant, wrote publicly about the arbitration and the Saladino default, and is himself a defendant in two of the tracked actions — C004 in Lackawanna County, filed 12 February 2025, and C094 in New York State Supreme Court, filed six days later. His account is useful for chronology and for the arbitration language it quotes. It is also the account of an adverse party. The corpus records that in the row's own EvidentiaryTreatment field rather than leaving a reader to discover it.

Around these sit the working layers: 140 research notes, 158 dated timeline events grouped into 25 chronology sequences, 11 DocketBird index records, and 94 catalogued videos.

VI. The coding rule

One variable does real work in the argument, so it has a written rule, published at thesis/protocols/01-refiling.md and applied to all 92 actions on 2026-08-28.

An action is a refiling when, at the moment it was commenced, the plaintiff had already commenced a prior, distinct originating action arising from the same underlying dispute, and the new action re-asserts substantially the same claims.

All three conditions are required. Condition 1 excludes a removal or transfer of the same action (that is what the union-find is for), a direct appeal within the same action, an action not commenced by this plaintiff, and a same-day companion — where two actions share a filing date the data cannot order them, so neither can be the other's prior. Condition 2 is satisfied by a shared DisputeID or, for any pair, by two claims summaries describing the same real-world transaction. Condition 3 requires the same core wrong against at least one shared canonical defendant.

Adverse termination is deliberately not an element. This is the design decision that matters most, and it is what separates the measure from res judicata. A duplicate filed while the first action is still pending is precisely the conduct this thesis is about, and a rule that required a prior loss would exclude it. Prior adverse termination instead drives the strength grade, where Strong requires either a judicial characterisation of the pair, or prior adverse termination plus the same claims and defendants, or documented verbatim carryover between pleadings.

The result, at originating-action grain: 43 of the 88 commenced actions repeat an earlier one — 13 Strong, 25 Moderate, 5 Context — against 46 coded No and 3 Unclear. By year of filing the repeats are 1 in 2022, 2 in 2023, 2 in 2024, 20 in 2025 and 18 in the first eight months of 2026.

Mind the grain here above all. Querying the cases table directly returns 52 Yes out of 102, because the coded value is written identically to every record in a same-action group so that the column is usable at either grain. 52/102 and 43/88 are the same finding at two units. Quoting one for the other is the easiest error available in this project and has already been made once, in an earlier draft of the proposal.

VII. Validation

A method that produces a number nobody else has produced cannot be checked against anything — unless it can first be made to produce a number somebody else already produced.

That is what thesis/exhibits/detection_gap.py does. It takes the two occasions on which a court stated a count of these filings, applies the union-find, restricts the result to that court's own forum, cuts it off at that date, and compares:

Two dates, two judges, two different values, exact reproduction both times. Nothing in the de-duplication was tuned to those figures — the same union-find had already been written for the public site — and the script exits non-zero if the reproduction ever breaks, so a future data edit that invalidates the thesis's headline claim fails loudly rather than silently. It is not wired into CI; it is run by hand before the numbers are cited.

This is the licence for the all-forums figure, and it is a narrow licence. It establishes that the counting method matches judicial practice when given judicial inputs. It does not establish that the corpus is complete — an action nobody has found is invisible to a validation built on found actions.

And reproducibility is not validity. That distinction is worth labouring, because this project's most visible methodological feature is the thing most likely to be mistaken for something stronger. Every figure in this book is regenerated by a script; the scripts are read-only, stdlib-only, and each exits non-zero when a premise it validates stops holding. A reader can therefore check that the numbers follow from the database, exactly, every time.

None of that says anything about whether the database is right. A script reproduces a classification perfectly whether the classification was sound or biased, and the exhibits here are auditing arithmetic, not judgment. Where the input is a column somebody else recorded — a filing date, a docket number, a court's own disposition — reproducibility is most of what matters. Where the input is this project's own coding, it is close to none of it: RefilingOfPriorCase, the filing-restriction indicators, the statement tiers and factors and chapter 4's rung labels are all contestable readings that a script will reproduce to the last decimal.

So the gates are a guarantee against one class of error and silent about another. They stop a figure moving without anyone noticing. They cannot stop a coder from being wrong in a consistent direction, and a consistent direction is exactly what a single coder with a hypothesis produces. The reliability evidence that would speak to that is set out in section VIII and limitation 1 below, and it is thinner than the reproducibility layer by a wide margin.

VIII. What the audit found

thesis/README.md asked for an inter-rater reliability check on the refiling variable. There was no second coder, and this project does not claim inter-rater reliability.

What ran instead is a rule-versus-judgment audit, which is a weaker instrument and is reported as one. refiling_rule_coder.py implements the mechanically decidable part of the rule — condition 1, condition 2 restricted to a shared DisputeID, canonical-defendant overlap, and a lookup of whether a court had already characterised the pair. It cannot read two claims summaries against each other, so it cannot implement condition 3 at all. Its predictions were generated before the coding pass and compared afterwards:

python thesis/protocols/refiling_rule_coder.py --compare
  agreement: 74/92 = 80.4%   (18 to adjudicate)

Because the screen is deliberately weaker than the rule, a disagreement is a finding about what reading the claims text added — not an error in either column. All 18 are adjudicated in the protocol, and the pattern in them is the most useful thing the audit produced:

The screen's failures are almost entirely failures of the corpus's own metadata, not of the rule. Two disagreements arose because a defendant roster omitted a party the case's own claims summary named as a lead defendant, or because two spellings of one company carried different canonical ids with no alias joining them. Five arose because neither case in the pair carried a DisputeID, leaving the screen no way to test condition 2 — and those five resolved two to Yes, two to No and one to Context, which is the useful result: a missing dispute group is genuinely uninformative in both directions, so the 29 DisputeID-blank cases could not have been coded by screen alone. Eight were strength changes that required reading the pleadings' summaries against each other. Two were cases where the pleading is unavailable and the honest code was Unclear even though party overlap alone would have supported Yes.

An audit that mostly indicts the data rather than the rule is a better outcome than one that finds nothing, and worse than one run by a second human being. Both things are true.

IX. Limitations

Stated plainly, and in the order of how much they cost the argument.

1. No second coder. The audit above substitutes for the inter-rater check and is not equivalent. Any published use of the repeat-filing variable must say so.

1. Condition 3 rests on claims summaries, not on pleadings. 101 of 102 cases carry a claims summary, but a substantial minority of those are sourced to docket coding rather than to the complaint's own text; the notes say which, case by case. Where that made comparison impossible the code is Unclear (3 actions); where it made it thinner but possible, the grade is capped at Moderate. C092 has no claims summary at all.

1. Defendant canonicalisation had three known defects when this chapter was written; all three were repaired on 2026-09-06, and repairing them found six more. The original three: C040's roster omitted the Shopify entities its own claims summary names as lead defendants; Fenix Internet, LLC and Fenix Internet LLC carried different canonical ids with no alias joining them; and Taban was absent from C018's roster though its claims summary names him. Each had been worked around by hand.

The six found afterwards are one defect wearing one disguise: a party listed surname-first on a state docket and first-name-first on the federal docket it was removed to, with no alias joining the two spellings, so one person counted as two entities. Five sat inside removal pairs (C030/C064, C036/C068, C015/C057) and were found by changing a gate in escalation_ladder.py from a liveness check into a detection rule; the sixth, Jeffrey K. Davis, Esq., had survived the morning's repair pass because that pass aliased Jeffrey K. Davis without the , Esq. suffix the data actually carries, so it matched nothing. The canonical-entity count fell from 335 to 327. A canonicalisation repair is only as good as the exact string it targets, and a defect of this shape is invisible to inspection — it takes a rule that recomputes rather than a constant that remembers.

Repair remains scoped: the wider Shopify family holds 18 distinct canonical entities, 13 with their own published page, so consolidating them is a separate editorial decision and not a tidy-up.

1. Dispute assignment is incomplete and, in one case, internally inconsistent. 29 of 102 records carry no DisputeID. Within the single action C014/C053, the state record carries DSP018 and its own federal removal carries DSP001. Chapter 2's cross-forum dispute figures rest on the 73 records that are assigned.

1. ProSe is unrecorded for 44 of 102 records, including 19 of the 38 Lackawanna records. This is why Chapter 2 states the Rule 233.1 eligibility point in terms of the rule's target population rather than a coded count, and why the pro se claim cannot yet carry weight in the chapters that need it. Only 8 of the 44 blanks have any pro-se language anywhere in the database and 4 more resolve by the same-action twin rule; the remaining ~34 need a docket research pass before any coding rule can apply.

1. 15.9% of citations sit on hosts a reader cannot open, and the blocked share is concentrated in the state-court material.

1. Absence of a row is not absence of an event. Chapter 2's negative findings — that no invocation of Pa. R.C.P. 233.1 or of 22 NYCRR 130-1.1 appears anywhere in this record — are claims about what this corpus contains. A motion denied without a written opinion in a county court would not be visible here. What is safe is the narrower claim that no filing restriction was entered anywhere, because an entered restriction appears on a docket.

1. A moving target, and an as-of date does not make a figure reproducible. 33 of the 92 actions were filed in the first eight months of 2026. Every count needs its as-of date, and the exhibits should be re-run rather than quoted from a previous chapter.

The subtler half of this was only noticed on 2026-09-06, and it limits what a reader can check. The corpus grows by discovery as well as by filing, and the two are not the same clock. C103 and C104 were filed in September and November 2025 but only found on 2026-08-30, in a search run under a company name rather than an individual's; C105 was filed 2026-09-01 and found on 2026-09-06. So a reader who takes a later database and filters it to "actions filed on or before 2026-08-28" does not get this chapter's 92 and 102 — they get 94 and 104, because two actions that existed on that date were not yet in anyone's record of it.

Nothing in the database records when a row was added, so those earlier states cannot be reconstructed from it. The as-of date on each chapter therefore says when the corpus was read, not what a date filter would return, and the two diverge by exactly the discovery lag. This is the thesis's own subject turned on the thesis: a corpus, like a docket, shows what somebody went looking for, and the count depends on who looked and when. It is also the reason the figures in chapters 1 to 3 are left as written rather than recomputed — recomputing them would silently replace what was known with what is known now.

1. The case was selected because it already looked like the thing the book argues about. This is the limitation an argument about invisibility can least afford to leave implicit. The subject was not drawn from a sample frame of high-volume filers; he came to notice because a large cross-forum record was already visible, and the project then set out to assemble it. Everything this book says about detection therefore runs on a case that was, by construction, detected.

Two consequences follow. The corpus cannot support any claim about how often this happens — not a rate, not a base rate, not a statement that the American system routinely fails this way. And the features that made the case findable are the same features that make it atypical: an unusually high filing rate, an unusually complete public commentary track kept by the filer himself, and a geography in which one county court and one federal district sit on top of each other so that ten removals all land in the same place. A reader should assume this record is extreme rather than representative.

What the selection does not damage is the doctrinal argument. Chapter 2 reads four regimes against their own text, and that reading holds or fails whether or not this filer exists. The empirical chapters are an existence proof with a mechanism attached, which is a weaker claim than a measurement and is the claim the book makes.

1. The null result is consistent with the explanation offered, and does not establish it. Zero filing restrictions across this record is compatible with an institutional aggregation gap. It is also compatible with judges exercising discretion against restriction, with the evidence actually put before any particular court being insufficient, with the procedural posture never presenting the question, and with defendants having sound strategic reasons not to move. This project cannot rule those out, does not try, and states the finding at the strength the record supports. Chapter 1 names the alternatives where the null result is first stated; chapter 8 turns instead on the narrower fact that survives all of them, which is that two parties did ask and neither request was decided.

X. Why the separation is load-bearing

It would be possible to build all of this and still ruin it in the last step, by letting the layers of Part V blend on the way into prose.

The argument of this thesis is that no court has determined that this litigant is a vexatious litigant, and that the reason is structural. That argument is destroyed, not supported, by a corpus that quietly treats a research classification as a judicial finding. If a RefilingOfPriorCase code graded Strong were allowed to read as though a court had said so, the thesis would be asserting the very determination whose absence is its subject.

So the discipline runs in the direction that costs the argument something. 34 filing-restriction indicators are recorded and none of them is a finding. 43 actions are coded as repeats and no court has called them that. Two motions asked for a vexatious-litigant declaration and both are filed under party filings, where the assertions in them are not evidence of the facts they assert. The corpus is at its most careful precisely where a looser rule would help most.