Wave 3 · instrument family · distributions and bands

A band is the product. A point estimate is what we withhold.

Draw the interval, draw what was censored, name the denominator.

How likely is an appeal still open at day t?

Axis truncated at 1,500 days. The longest circuit p90 in the cohort is 1037 days.

Admissible region — every non-increasing curve through the two anchors lies inside it
Measured anchors — pooled median and p90
Per-circuit anchors — twelve pairs, unbound individually

Why no curve

Two quantiles do not determine a distribution. Drawing a smooth curve through them would show a shape the data does not contain, so the instrument draws what is actually known: the region a monotone survival curve must pass through. The region is the claim; the curve would be the invention.

The exceedance function is known at exactly two days per circuit. Every other day is unmeasured, and no value is imputed for it.
circuit	n	day where P(open) = 0.50	value	day where P(open) = 0.10	value
pooled (served)	73,897	398	0.50	827	0.10
5th Cir.	10,727	307	0.50	612	0.10
6th Cir.	5,357	313	0.50	564	0.10
8th Cir.	4,524	371	0.50	575	0.10
7th Cir.	3,087	374	0.50	743	0.10
10th Cir.	3,167	381	0.50	710	0.10
3d Cir.	4,175	382	0.50	763	0.10
11th Cir.	7,461	384	0.50	805	0.10
4th Cir.	6,631	427	0.50	840	0.10
9th Cir.	17,097	444	0.50	1005	0.10
D.C. Cir.	2,176	474	0.50	810	0.10
1st Cir.	2,346	485	0.50	1033	0.10
2d Cir.	7,149	532	0.50	1037	0.10

watchtwo pooled anchors served; the 24 circuit anchors are not

producer N0813-CLA-APPELLATE-TIMING-WEDGE-R1 over fjc_appellate_ap08on.parquet · frame P(still open at day t), derived only from the two published quantiles of the same cohort · gate pass for the two pooled anchors, watch for the twenty-four per-circuit anchors · litan/reports/N0813-CLA-APPELLATE-TIMING-WEDGE-R1.json

How long does a civil case actually take?

Withheld: no served_truth key binds a survival median, so this console quotes none.

the band belongs here

withheldcard_aggregate_registry is DARK

producer card 01, Kaplan–Meier over FJC civil · as of 2026-03-01 · n five NOS cohorts, 256,559 filings · litan/reports/N0904-CLA-STRATEGY-REASSESSMENT-R2/cards/01-civil-timing-survivorship-gap.md

Share of cases terminated, by month

Antitrust (NOS 410) in two filing vintages. The 2023–2024 cohort is a third still open — which is exactly why its resolved-only median is wrong by eleven months.

Knots at 12 / 24 / 36 / 48 months are the only measured points. Nothing is drawn between them.

2015–2022 filings — KM, n = 4,802
2023–2024 filings — KM, n = 1,245
Greenwood 95% interval — 2015–2022 only
Admissible region — where a monotone curve must lie between two knots
Not measured — the segment between knots

Censored mass, drawn — never dropped.

What this does not say

TDATEUSE IS NULL is PENDING (DISP = -8) and enters KM censored at as_of - FDATEUSE. It is never given a duration. A censored case is never given a duration and never counted as a fast resolution. Greenwood 95% CI (Table A only; Table B publishes no interval) — the 2023–2024 vintage publishes no interval, so its line is drawn without a band rather than with an invented one.

Erratum carried from card 13. ERRATUM 2026-09-06 (NOS 830 row; from card 13). The Table A patent row's "+1.0 mo (+17%)" gap and Table B's "+1.1 mo" compare two rulers: `FDATEUSE`/`TDATEUSE` are month-truncated (one distinct day-of-month across 32,517 rows), so the KM median was read in a `ceil(days/30.4375)` bucket (7) while the resolved-only median was continuous (6.01). On matched month units the CY2015–22 patent gap is 0.0 (KM median 6 [6–6] vs resolved-only 6; pending 1.53%). The survivorship thesis still holds where pending mass is large — CY2023–24 KM 7 vs 6, CY2024 KM 7 vs 5 — and the antitrust (NOS 410) demonstration in this card is unaffected. Card 13 carries the corrected NOS 830 table; the sealed table hash below covers the uncorrected rows.

Rows refused by the sentinel / ordering / magnitude gate: 0 in every one of the five cohorts — the pending rows are genuine pendency, not gate casualties.

Kaplan–Meier share terminated by month t. Every plotted value is here. Denominator n is the stratum's own filing cohort; as of 2026-03-01.
cohort	n	pending (censored)	pending share	12 mo	24 mo	36 mo	48 mo	KM median (mo)	resolved-only median (mo)	gap
190 contract · 2015-2022	87,293	1,121	1.28%	0.6562	0.8593	0.9320	0.9672	8	7.98	+0.0 mo
410 antitrust · 2015-2022	4,802	577	12.02%	0.5100	0.6195	0.6774	0.7381	12	7.98	+4.0 mo (+50%)
440 civil rights · 2015-2022	118,771	1,704	1.43%	0.6394	0.8368	0.9206	0.9621	8	7.95	+0.1 mo
830 patent · 2015-2022	32,517	497	1.53%	0.7257	0.8712	0.9314	0.9627	7	6.01	+1.0 mo (+17%)
850 securities · 2015-2022	13,176	439	3.33%	0.6671	0.7917	0.8709	0.9290	5	4.01	+1.0 mo (+25%)
190 contract · 2023-2024	24,672	3,433	13.91%	0.6510	0.8522	0.9327	8	6.01	+2.0 mo (+33%)
410 antitrust · 2023-2024	1,245	424	34.06%	0.4610	0.6693	0.8044	16	5.03	+11.0 mo (+218%)
440 civil rights · 2023-2024	35,457	5,038	14.21%	0.6650	0.8455	0.9279	7	5.03	+2.0 mo (+39%)
830 patent · 2023-2024	6,454	1,006	15.59%	0.6675	0.8372	0.9057	7	5.95	+1.1 mo (+18%)
850 securities · 2023-2024	2,682	615	22.93%	0.5906	0.7510	0.8580	7	3.02	+4.0 mo (+132%)

producer card 01 — Kaplan-Meier on data_local.nosync/litan.duckdb::fjc_civil (FJC Integrated Database, civil, 10,857,386 rows, read_only) · as of 2026-03-01 · gate withheld — no served_truth key binds any survival median; card_aggregate_registry.py is DARK=True and no route reads this file · litan/reports/N0904-CLA-STRATEGY-REASSESSMENT-R2/cards/01-civil-timing-survivorship-gap.md

The withhold beside the band

The estate refuses a duration estimate here — on purpose.

The venue comparison serves settlement, dismissal and judgment shares for any two districts. It will not serve a duration, and it says so in its own bytes: “The ordering depended on a duration figure we withdrew, so the ranking is withdrawn with it.” The KM band above is what an honest answer looks like; the slot below is what the product does instead of guessing.

Median days to termination, D. Mass.

Withheld: pending cases are excluded from the only duration we can compute, so the number would be biased downward.

a number belongs here

withheld

What is served instead

Settled: 48.2% of n = 9,122
Dismissed: 21.3%
Judgment: 13.8%

pass

as of 2026-03-31 · Personal-Injury & Product-Liability Torts (NOS 350-368, 385)

withheld watch pass

withheldthe rail carries a refusal

How long does a federal appeal run?

398 days, median · docketed → disposition · n = 73,897

passserved_truth-bound

p90 827 days · circuit range 307 days–532 days · key appellate-pendency-docketed-to-disposition-median-days

Which clock is the reader on?

4clocks to one terminus · 1 served, 3 inert as of 2026-08-19

stalebinding minted 2026-08-19

coverage 1.00 / 1.00 / 0.79 / 0.70 of the same cohort · litan/reports/N0813-CLA-APPELLATE-TIMING-WEDGE-R1.json

Days from docketing to disposition, by circuit

The pooled median and its p90 carry served_truth keys. The twelve rows below them do not — they are the same producer’s own output, drawn here with each circuit’s own denominator and interval.

Axis begins at zero. Circuits are ordered by median — an ordering of the axis, not a ranking of courts. Every 95% interval is 6–19 days wide, narrower than its own marker at this scale; the table carries the endpoints.

median — with its 95% interval as the bar
p90 — one in ten appeals is still open past this day
served pooled median, 398 days — the only bound figure on this chart
served circuit range, 307–532 days — bound as a range, not per circuit

Clock and floor

Clock. DKTDATE (date the appeal was docketed in the court of appeals) to JUDGDATE. This is the clock N0804-CLA-APPELLATE-PENDENCY-BUILD-R1 used.

Cohort. counseled (PROSETRM='0') merits terminations (DISP in {1,2} = decided after oral hearing or on submission), cohorted on JUDGDATE (never DKTDATE), judgment 2021–2025. n = 73,897.

n-floor. 95% order-statistic half-width <= between-circuit SD of medians; smallest passing n = 2,176; suppressed circuits: 0.

Serving posture. PARTIALLY BOUND, DURATIONS ONLY. The docketed_to_disposition clock's pooled median, pooled p90, n, and the range of its admitted per-circuit medians carry served_truth keys (appellate-pendency-docketed-to-disposition-*), minted 2026-08-19 by N0819-CLA-APPELLATE-CLOCK-BINDING-T0 and resolved on /matter for federal appeals. EVERY OTHER CLOCK IN THIS FILE REMAINS INERT for external purposes -- argued_to_disposition and lastbrief_to_disposition are conditional on a stage the reader has not reached, noa_to_disposition covers 0.79 of the cohort, and none of the three has a served_truth key or a publication. No RATE from this file or its neighbours is published in either direction. Any surface must render the clock definition and the n-floor suppression disclosure in the same bytes as the number, and must state the per-circuit range beside the national pooled figure.

Docketed → disposition, in days. Median with its 95% binomial order-statistic interval, and p90. n is each circuit's own count; the pooled reference is n = 73,897.
circuit	n	median (days)	95% interval	p90 (days)	half-width	passes n-floor
5th Cir.	10,727	307	302–310	612	4.0	yes
6th Cir.	5,357	313	308–316	564	4.0	yes
8th Cir.	4,524	371	364–376	575	6.0	yes
7th Cir.	3,087	374	365–384	743	9.5	yes
10th Cir.	3,167	381	371–390	710	9.5	yes
3d Cir.	4,175	382	376–388	763	6.0	yes
11th Cir.	7,461	384	378–391	805	6.5	yes
4th Cir.	6,631	427	418–434	840	8.0	yes
9th Cir.	17,097	444	441–448	1005	3.5	yes
D.C. Cir.	2,176	474	463–485	810	11.0	yes
1st Cir.	2,346	485	474–503	1033	14.5	yes
2d Cir.	7,149	532	529–543	1037	7.0	yes

producer N0813-CLA-APPELLATE-TIMING-WEDGE-R1 over fjc_appellate_ap08on.parquet · bound keys appellate-pendency-docketed-to-disposition-* resolved through litan.served_truth.resolve_external_display · gate watch — the pooled median, p90, n and the circuit range are served; the twelve per-circuit rows are the same producer's own output and are not individually bound · litan/reports/N0813-CLA-APPELLATE-TIMING-WEDGE-R1.json

Four clocks, one terminus

The same 73,897 appeals, measured from four different starting columns. Only the docketed clock is bound; the other three are conditional on a stage the reader may not have reached, and are drawn without a served figure.

Bands run median → p90. Coverage is printed per stage because three of the four cover less than the whole cohort.

median → p90 — the band, per clock
median
inert — no served_truth key, conditional on a stage

Four clocks to the same terminus. Only one carries a served_truth key; the other three are inert for external purposes and are drawn as such.
clock	columns	n	coverage of cohort	median (days)	p90 (days)	posture
notice of appeal -> disposition	APPDATE → JUDGDATE	58,374	79.0%	406	829	inert
docketed in the court of appeals -> disposition	DKTDATE → JUDGDATE	73,897	100.0%	398	827	served
last brief filed -> disposition	BRFILED → JUDGDATE	52,013	70.4%	220	545	inert
argued or submitted -> decision	ARGSUB → JUDGDATE	73,897	100.0%	36	210	inert

producer N0813-CLA-APPELLATE-TIMING-WEDGE-R1 over fjc_appellate_ap08on.parquet · gate stale — the one bound clock’s key was minted 2026-08-19 against a cohort cut at judgment 2025 · litan/reports/N0813-CLA-APPELLATE-TIMING-WEDGE-R1.json