The distributed filer

This chapter cites law as well as corpus rows. The two are kept in separate tables at the end — Sources for the record, Legal authorities for the statutes — so that the rule this folder runs on, that no fact about the litigation originates in it, stays checkable by inspection.

Revision, later on 2026-09-06. C105 entered the corpus after this chapter was first drafted, on the same day, and moved the totals every section below reports: 94 originating actions became 95, and the 90 he commenced became 91. The figures here are the post-C105 ones. Nothing structural changed — all four pre-registered verdicts stand, and the quartile shape is still a dip and a rebound — but one cluster gained a second state, because the new action put a third forum on DSP006.

Revision, later still on 2026-09-06. The two canonicalisation defects section IX listed as sitting inside the measurement were repaired, and repairing them found five more of the same kind. The canonical-entity count fell from 335 to 327 (one entity has been added since, so the corpus now holds 328) and the evidence-anchoring audit from 50-of-69 to 49-of-68; the rung distribution, the cluster figures and all four pre-registered verdicts are unchanged. Section IX's limitation 3 is retired outright — the appellate layer became first-class the same day and is now measured by its own exhibit. Three cells in the Sources table that had not been carried forward from the C105 update are corrected here.

Revision, 2026-09-11. Two external reviews found two defects in section V, and both are repaired here. E2 and E4 were reported as REFUTED on non-significant rank correlations (ρ = 0.111, p = 0.29 and ρ = 0.076, p = 0.46, n = 93); a failure to find is not a refutation, they are now reported as NOT SUPPORTED, and the section states what the whole-sequence correlation was poorly placed to see — the quartile table's step change between Q1 and Q2. And E1's refutation turned on an untested definitional choice: C099 reaches rung 6 only because a private arbitral institution is coded as an adjudicator alongside sitting judges, and the four original perturbation axes never split that rung. Two axes now do. Under one of them E1 holds, and the section says so. Section II gains the cluster-boundary sensitivity the same reviews asked for. No corpus figure changes; the chapter's counts remain as of 2026-09-10.

Revision, 2026-09-10. C106 and C107 entered the corpus on 2026-09-09 and moved every total below: 95 originating actions became 97, the 91 he commenced became 93, and 105 docket records became 107. Both arrived through escalation_ladder.py failing rather than through anyone noticing — C106's sole defendant, the City of Carbondale, tripped gate G3 as an entity the screen elevated and no one had adjudicated. It is now coded Government on the strength of the caption alone, because RECAP holds a docket shell and no complaint to read, so the anchoring audit reads 49-of-69 rather than 49-of-68. Nothing structural changed: all four pre-registered verdicts stand, the rung-opening table is untouched (C106 was filed long after the Government rung opened at #24), and the quartile shape is still a dip and a rebound. The counterparty rung also fell from 263 entities to 257, which the 2026-09-06 note above did not carry into the table: that revision recorded the rung distribution as unchanged by the canonicalisation repair, and on this row it was not. C107 also gave the corpus its first natively-Californian district filing, which is why the filed-court table below now carries a C.D. Cal. row.

I. A cluster is not a case

Chapter 2 found that none of the four regimes obliges any person or office to assemble a filer's record across forums, however far their predicates are permitted to count. This chapter is the first of four demonstrations of what that omission produces in a real record — here, in the shape of the litigation itself.

Chapter 1 counted actions and found that two judges had each counted correctly and each counted only what was in front of them. Chapter 3 explained how the actions were de-duplicated. This chapter asks a different question: not how many, but what shape.

The unit here is the dispute cluster — a set of originating actions arising from one underlying quarrel, recorded in the corpus as a DisputeID on each docket record. As of 2026-09-10 there are 22 clusters, covering 71 of the 97 originating actions. The largest is DSP003, Saladino-related litigation: 8 actions over 1,928 days, from 2020-11-07 to 2026-02-17. Two clusters contain a single action. Twenty-six actions carry no cluster assignment at all.

That last number is a quarter of the corpus and it is not a rounding error, so it goes at the front rather than in a footnote. Everything this chapter says about cluster structure is a statement about the seventy-one actions that carry an assignment. The refiling protocol found, when it hit the same gap, that unassigned actions are uninformative in both directions — some turned out to repeat an earlier action and some did not — so the gap is better described as missing than as evidence of dispersal.

There is a second limit, and it is structural rather than clerical. cases.DisputeID is a single column, so the corpus can record an action in one cluster and only one. Some actions genuinely belong to two. C014/C053, Tufano v. TikTok, names TikTok and the surgeon at the centre of the Taban cluster; on 2026-09-06 the two halves of that one removed action were found to be carrying two different cluster labels, and were reconciled onto the TikTok/ByteDance cluster because that is the dispute the complaint leads with. The reconciliation picks a defensible label; it does not make the overlap go away. The clustering is a partition imposed on something that is not partitioned.

II. Eighteen of twenty-two clusters cross a forum boundary

With that stated, the shape is unambiguous. Eighteen of the 22 clusters span more than one court. Seven span more than one state.

Four of the seven largest clusters reach into two states. The Taban cluster reaches four courts — a California superior court, a Pennsylvania county court, a federal district in Pennsylvania and a federal district in California — across six actions in sixteen months.

Eighteen is the count of clusters that touch more than one court, and it is the number this section is named for. It is worth separating from the narrower one, because the two answer different questions. Sixteen of the 22 clusters reach a second court because he commenced actions in more than one forum. The remaining two reach one only because a defendant removed the single action in them — DSP007 and DSP009, each a lone Lackawanna case pulled into a federal district. Chapter 2 draws the same distinction and runs its argument on the narrower figure; the difference between the two is entirely a question of whose choice put the case in the second forum. cross_forum_web.py reports the narrower figure over the 20 clusters holding at least two actions, where no removal-only crossing appears at all: the two that qualify are precisely the two single-action clusters that rule excludes.

Because §I concedes that the partition is imposed on something that is not partitioned, the eighteen should be re-run with the known overlap reassigned the other way, and it is: escalation_ladder.py recomputes every cluster figure with C014/C053 moved off the TikTok/ByteDance cluster and onto the Taban cluster, the defensible alternative. The headline moves by one: eighteen cross-forum clusters become seventeen, and the six that cross a state line are unchanged. That is the price of the partition on the ambiguity this project knows about. It does not price the ambiguities it does not know about, and §IX limitation 2 says so.

This is the fact chapter 1 approached from the other side. A judge asked how many actions this plaintiff had filed in this court was, on the Taban cluster, looking at one of four places the same quarrel was being litigated. Nothing was hidden. The clusters are visible in the corpus because the corpus was built to make them visible; they are invisible to a court because no court has an instrument that reads across four dockets in two sovereigns.

III. Removal, and the arithmetic it creates

Ten of the 97 originating actions exist as two docket records rather than one. Each is a state case removed to federal court, and each therefore appears twice in any count that counts dockets.

The mechanism is statutory and unremarkable. Under 28 U.S.C. § 1441(a), a civil action brought in a state court of which the federal district courts have original jurisdiction "may be removed by the defendant or the defendants, to the district court of the United States for the district and division embracing the place where such action is pending." Under § 1446(b)(1) the notice of removal must generally be filed "within 30 days after the receipt by the defendant, through service or otherwise, of a copy of the initial pleading."

Both features show up in the record exactly as written. Nine of the ten removals fall between 21 and 38 days after the state filing — the thirty-day clock plus service — with a median of 30; the tenth, C032 to C074, took 83. And every one of them lands in the Middle District of Pennsylvania, because § 1441(a) sends a Lackawanna County case to the federal district that geographically embraces Lackawanna County and nowhere else.

That second point is the one that matters, and it is worth stating plainly before the numbers: removal does not scatter a case. It concentrates it. The defendant's choice of forum is not a choice at all; the statute names the destination. Ten times, a state action was mechanically relocated onto the docket of the one court that would later try to count.

IV. The one court where the duplication shows

Counting dockets rather than actions overstates the corpus by ten — 107 records against 97 actions, +10.3%. The interesting question is where that overstatement lives.

If a docket is counted only in the court it was filed in, the answer is nowhere:

Every forum is internally consistent and the aggregate is not. That is what a removal does: it creates the second record in a different court from the first, so no single court's own filing register is inflated by it, and the inflation exists only in the union.

But a court does not search its own register by where a case was filed. It searches by what is on its docket, which includes what was removed onto it. Counted that way — the definition detection_gap.in_forum() uses, filed-in or now-sitting-in — the picture changes in exactly one place:

The Middle District of Pennsylvania is the only forum in the corpus where docket-counting and action-counting disagree. It is also the only forum in the corpus where a judge ever stated a count. The court positioned to notice the pattern is the court whose own record most overstates it.

And the overstatement is recent. Cumulatively, records equalled actions through the end of 2024 — 19 and 19. The gap opens in 2025 (66 records, 61 actions) and reaches ten in 2026. Before 2025 there was no de-duplication problem to solve. It arrived with the filing rate.

What the judges did with it

Here the record cuts against the easy version of this chapter's argument, and the concession belongs in the text.

At the two dates a court stated a count, a naive docket count would have been wrong:

Both judges stated the de-duplicated figure. Neither stated the record count, which would have overstated by 2 and then by 7. Whatever instrument produced those footnotes — the clerk's records in Munley's case — it collapsed the removed pairs correctly.

So the de-duplication problem is not the courts' failure. It is a problem they demonstrably solved, inside their own forum, twice. What they could not solve was the other direction: the same de-duplication run across all forums returns 65 and 85. The court got the hard arithmetic right and the easy aggregation wrong, because only the arithmetic was in front of it.

Two ways the corpus records the same movement

A count of removals taken from the union-find returns ten. The corpus also records cross-forum movement a second way, in CurrentCourt, and the two do not coincide. Twelve records carry a CurrentCourt; two of them — C046 (M.D. Pa. → C.D. Cal.) and C098 (N.Y. Supreme → S.D.N.Y.) — have no paired record and no same-action link, so they are invisible to a removal count taken from SameActionAsCaseID alone. A third, C074, carries a CurrentCourt identical to its Court, which is a no-op and a data defect.

The corpus therefore represents cross-forum movement twice, incompletely, in two columns that disagree — which is a small, concrete instance of the thing this thesis is about. Even a dataset built to track a filer across forums records his movements in two places and needs a script to reconcile them.

V. The ladder the outline expected

The proposal for this chapter asserted an escalation ladder: that the litigation moved outward from the parties to a transaction, through opposing counsel named as parties, to licensing boards, to platforms, to the county, and finally to eight sitting judges and the Commonwealth. (The last phrase is stronger than the record supports, and the correction is worth making before the argument starts: the C038 court finding dismisses "the eight individually-named Lackawanna County judicial defendants" and identifies one of them as a sitting county judge — which is what forced the full-bench recusal.)

Testing that requires a variable the corpus does not have. The dataset records who was sued; it has no column for what kind of thing they are. thesis/protocols/02-escalation.md supplies one under a published rule: six rungs ordered by distance from the underlying transaction — Counterparty, Counsel, Platform, Regulator, Government, Adjudicator — assigned to each of the 328 canonical entities, with every elevation above rung 1 required to point at an attorneys row, a court_findings row or a claims-summary sentence naming the role. Nothing is written to the database and nothing appears on the site; the values live in the exhibit.

The ladder's claim was then stated as four falsifiable propositions before the numbers were read. One is refuted outright, two fail to find any trend at all, and the one that passes does so on a test too weak to carry it.

E2 and E4 are reported as not supported rather than refuted, and the distinction is not pedantry. A non-significant rank correlation is a failure to find a monotone trend. It is not evidence that no trend exists, and an earlier draft of this chapter labelled both REFUTED, which claims a great deal more than ρ = 0.11 at p = 0.29 can support. Getting it right matters here rather than in the abstract, because §VII's quartile table shows something a whole-sequence rank correlation is poorly designed to detect: mean top rung steps from 2.13 to 3.00 between the first and second quartiles and then stops — 3.00, 2.96, 2.75. That is a step change, not a trend, and Spearman over 93 ordinals will read it as noise, which is precisely what it did. So the honest version is that E2 survives in a weaker, non-monotone form — something changed once, early in 2025, and then held — and that §VII's replacement finding is that story told properly.

The choice E1 turns on

E1 fails on a single observation, and that observation rests on a definitional decision, so the decision has to be tested rather than assumed.

The protocol's four original perturbation axes do not test it. They swap the Platform and Regulator ordering, drop the three mass-defendant actions, and substitute the corpus population for the filer population, and under all four every verdict above is unchanged. What none of them touches is rung 6 itself. C099 reaches the top rung because the American Arbitration Association and the arbitrator it appointed are coded as adjudicators alongside ten judges of a public court. Whether a private arbitral institution belongs on the same rung as a sitting county judge is a definitional call, and the refutation of E1 turns on it entirely. Two further axes now split it:

E1's refutation is not robust to that choice, and this chapter reports it rather than quoting the axis that suits it. If a private arbitral body is a kind of adjudicator — the reading this project coded — then the top rung opened at the sixth action and there is no progression to find. If instead it is a service provider the counterparty chose, nearer to a platform than to a judge, then C099 reaches only rung 3, the first public adjudicator is Judge Mehalchick in October 2025, the rungs do open in date order, and E1 holds.

Three things follow, in order of how much they cost the chapter.

It costs E1 and nothing else. E2, E3 and E4 return the same verdict under all six axes, and §VII's replacement finding — that the ladder is additive rather than substitutive — is untouched, because it rests on the rung-1 share and the distribution table rather than on where the AAA sits.

The reading that rescues E1 is the weaker one on the facts of this action. C099 names the arbitrator personally, by name, as a defendant, and the claims summary records the AAA as the institution that administered the arbitration the suit attacks. A platform is sued for what it hosts; an adjudicator is sued for how it decided. That is why the coding went the way it did, and it is an argument rather than a stipulation — which is the point of publishing it.

And the dating observation in §VI survives both readings. Under the service-provider reading C099 still puts three distinct rungs into one complaint at the sixth action, and the filer still named his own former attorney as a party in January 2023. What the split changes is whether that counts as reaching the top — not whether the record shows a filer working patiently upward.

VI. The sixth action

E1 fails on a single observation, and it is the most important thing in this chapter.

Here is when each rung first opened, over the 93 actions this plaintiff commenced:

C099 is Tufano v. American Arbitration Association et al, filed in the Southern District of New York on 2023-01-11 — the sixth action he commenced. It names exactly four defendants:

• Adam Starowicz, the opposing party in the underlying business dispute — rung 1;
• Jeffrey K. Davis, Esq., his own former attorney — rung 2;
• Sarah Thompson, recorded in the case's claims summary as the arbitrator;
• the American Arbitration Association, the institution that administered the arbitration — rung 6.

One complaint, four defendants, spanning the bottom, the second rung and the top of the ladder. It was filed two and a half years before he first sued a judge (Karoline Mehalchick, C059, 2025-10-17), and before platforms, regulators or government bodies appeared in the corpus at all. The companion state filing, C091, reused the identical four-defendant lineup three weeks later.

There is no progression to find. Suing the adjudicator is not the end of an escalation; it is a move that was available, and taken, at the second dispute. Whatever explains the 2018–2026 record, it is not a filer who worked his way up a ladder.

The proposal's specific ordering is wrong in two further places, which matters less but should be corrected rather than quietly dropped. Platforms opened in May 2024, seven months before the first licensing or professional board — the reverse of the outline's sequence. And rung 5 was opened by a federal agency, the USDA, in February 2025, not by the county: the Lackawanna County Assistance Office follows in May and the county itself in July.

VII. Additive, not substitutive

If the ladder is not a progression, what does change over the sequence?

The first quartile spans nearly seven years; the last spans five months. That alone is the chapter's most robust observation, and it belongs to chapter 1's story rather than this one.

Within it: the mean top rung jumps between the first and second quartiles and then stops moving — 2.13, then 3.00, 2.96, 2.75. The counterparty share falls from 0.57 to 0.26 and then climbs back to 0.46. E3's pre-registered test compares only the endpoints and therefore passes, but the shape is a dip and a rebound, not a decline, and the endpoint result should not be reported as though it were a trend.

What the record supports is narrower and, I think, more interesting: the ladder is additive rather than substitutive. Rungs open and never close. By 2026 he is suing adjudicators and still suing counterparties — fifteen of the thirty-four actions filed in 2026 top out at rung 1. The distribution shows the same thing:

Seventy-four of the ninety-three actions still name a counterparty. The base never empties. What the later period adds is not height but volume — and volume, not escalation, is what chapter 1's two footnotes were counting.

VIII. What the distribution is not

Two readings this chapter does not support.

It is not evidence of a strategy. Nothing in the corpus establishes that actions were distributed across forums in order to avoid aggregation. The clusters cross forum boundaries; ten of them do so because a defendant removed the case, which is the defendant's choice and not the plaintiff's. Chapter 5 records that he narrates his own filings constantly and in public, which is close to the opposite of concealment. Distribution is a property of the record. Intent is not in it.

It is not a finding about any of these entities. A rung label says what kind of party was named, under this project's published rule. It is not a holding that naming them was improper. The one court that has ruled on the top-rung defendants dismissed the claims against them on immunity grounds — which is a determination about the defendants' status, not about the plaintiff's filing conduct, and certainly not a vexatious-litigant determination. No court in this record has made one of those.

IX. What this chapter cannot see

Stated plainly, because the exhibit's gates cannot catch a gap that is uniform.

1. Twenty-six of 97 actions carry no cluster assignment. Every cluster figure describes the other seventy-one.
2. cases.DisputeID is single-valued, so overlapping disputes are recorded as a partition. C014/C053 is the known case; there may be others that were never noticed because nothing forces them to disagree with themselves the way that one did.
3. The appellate layer is now first-class, and the axis it opens is measured — this limitation is retired. When this chapter was drafted, appeals lived in Status prose rather than as rows and the second de-duplication axis went unmeasured: an appeal is another docket, in another court, arising from the same originating action. The appeals table landed later the same day, and thesis/exhibits/appellate_layer.py now measures it. The answer is that the axis multiplies records without widening reach — 30 appellate dockets from 22 trial-court records, and not one crossing a sovereign boundary. Counting both layers, the corpus's 97 actions have produced 137 docket records: 107 at trial and 30 on appeal. Verification followed on 2026-09-10: 19 of the 30 are now checked against the reviewing court's own docket, up from 7, and the pass added three appeals that had not been recorded at all — which is itself an instance of this chapter's argument, since two of them were found only by sweeping a court's own party index rather than by following any case the tracker already held.
4. The rung classification had one coder. What stands in place of a second is an evidence-anchoring audit — 49 of the 69 elevated entities (71.0%) rest on a table, a court finding or a claims summary rather than on a name — and the four-axis perturbation run. Neither is inter-rater reliability and neither is offered as such.
5. Both canonicalisation defects that sat inside the measurement have been repaired, and the repair found five more. Jeffrey K. Davis, Esq. was unaliased, so under the name-matching scheme used elsewhere in this project the Counsel rung opened 762 days late; this chapter keyed on DefendantID to avoid it. And the same two people were recorded as four entities across the two halves of the C030/C064 removal. Repairing the second exposed the shape of both — a party listed surname-first on the state docket and first-name-first on the federal one — and a gate rewritten to recompute that pattern rather than remember the known instances found five more, in C036/C068 and C015/C057. The canonical-entity count fell from 335 to 327; the rung distribution and every pre-registered verdict are unchanged, and the anchoring audit moved from 50-of-69 to 49-of-68 — and to 49-of-69 again on 2026-09-10, when C106's sole defendant, the City of Carbondale, was adjudicated onto the Government rung on the strength of the caption alone, RECAP holding no complaint to read. The residual limitation is not a defect but a method: the detection rule matches name-token permutations exactly, so it would not have caught a misspelling, and McCarty Judith C against Judith C McCarthy was found by reading two rosters rather than by any rule.
6. ProSe is blank on 45 of 107 records. Nothing in this chapter turns on it. Chapter 6 was drafted the same day and carries the same gap as a stated limit.