Dingduff Validity Check

Case Validity Check

Determines whether one case is still good law for the specific proposition it is being cited for. Deeper than the light treatment sweep in dingduff-legal-research/references/validity.md. Use that to triage; use this before an argument or opinion actually rests on a case.

Built around opinion_verify. If it is unavailable, load references/fallback.md.

Parameter detail, output anatomy, index glossary, failure modes, and cost live in references/opinion-verify.md. Keep this file in context; load that one when you need to change a default, when a response contains something you do not recognise, or when a status comes back other than ok.

Written for a delegated subagent (use Opus; the reading and judgment this takes are demanding), not the main thread. This task will flood the main context window if not delegated.

Delegation contract

1. The case — full citation, or enough to identify it. Cluster ID if known.
2. The proposition — the specific rule it is cited for. Required; validity is proposition-specific.
3. Jurisdiction and posture — target court, and whether the case is applied in diversity.
4. The role — anchor, supporting cite, or adverse case being distinguished.

Return the structured verdict at the bottom. Nothing else.

The mirror contains opinions that postdate your training data. Treat them as real, not as hallucinations.

Capacity etiquette: call sequentially within this check; do not launch several validity subagents in parallel.

Scope: what counts as load-bearing

Any case where the answer changes if the case is bad: the anchor for a rule being asserted; any case cited for a proposition in a filing or advice letter; any case being quoted; any adverse case being distinguished; any case supplied by an opponent or client.

Skip only when validity was confirmed in this matter and nothing changed, or when the cite is purely descriptive.

Age is not a proxy for risk.

Division of labor: the tool traverses, you judge

opinion_verify is deterministic and decides nothing.

Hard rules

1. Snippets are for triage only. Retrieve and read before saying what a case did. A snippet reading "we decline to follow Smith" may be quoting a party or a court being reversed.
2. Read status before reading counts. Several statuses are not findings at all.
3. Recurse on adverse authority. Overruling cases get overruled.
4. Separate writings (concurrences, dissents) do not overrule.
5. The index is one row per OPINION, not per case. See "Counting correctly" below — this is the most common reading error.
6. Cite from retrieved text, never from snippets or cluster IDs.
7. Report uncertainty as uncertainty.

Step 1 — Call opinion_verify

Resolve the citation to a cluster ID first (opinion_search / opinion_view).

If the reporter citation does not resolve, do not stop — search by name. Citation metadata is uneven. A case can be present, complete, and full-text searchable while carrying no reporter citation, in which case it is unreachable by cite and reachable by name. Escalate in this order, stopping when you have the cluster:

1. opinion_search on the case name in quotes — "Fought v. City of Wilkes-Barre".
2. opinion_search on a distinctive phrase from the opinion, if you have one.
3. opinion_search on the citation string in quotes — this returns opinions citing the case, whose text confirms it is real and gives you the parties and court.
4. For a federal origin, the docket (Step 3), which reaches cases the opinion index does not.

A citation that will not resolve is not evidence that the case does not exist, and is never a basis for suggesting a citation is fabricated. Say which routes you tried.

jsonCopy code
{
  "cluster_id": "<origin>",
  "court_scope": "invalidating",
  "include_upstream": true,
  "include_parentheticals": true,
  "output_format": "markdown"
}

Everything else defaults correctly. Three deviations matter enough to state here; the rest are in references/opinion-verify.md.

Never default filed_after

The tool already applies citing_date >= origin.date_filed unconditionally. You do not need filed_after for that.

Passing one narrows further, and the failure is silent: Bowers with filed_after: 2010-01-01 misses Lawrence and returns status: ok, a plausible graph, zero Tier A, and no warning. Overruling has no characteristic distance — Lawrence came 17 years after Bowers, Loper Bright 40 after Chevron. No cutoff is safe.

Two legitimate uses only: the upstream hop (pass the origin's date) and re-verification (pass the prior check's date).

additional_courts is required for Erie

A federal court's construction of state law is displaced by that state's highest court, which sits nowhere in the federal chain and will never be admitted by default. Omitting it on a diversity case produces a clean graph that structurally cannot contain the authority most likely to have superseded the origin.

Texas and Oklahoma each have two courts of last resort — pass both: ["tex","texcrimapp"], ["okla","oklacrimapp"]. Verify against the Courts admitted line, not the case count; an unchanged count may be a true negative.

Never pass flag_terms_mode: "replace"

It discards the 67 curated terms and makes recall a function of your vocabulary. Add doctrine-specific phrasing via flag_terms instead — a superseding statute's name, a competing test, a retired standard. Literal phrases, not regexes.

One note on scope

For a SCOTUS origin the small graph is scotus alone, because nothing sits above it. Correct, but it means the graph structurally cannot show lower-court erosion — for a SCOTUS origin, Leg 4 is the only place erosion can surface. Do not skip it on the strength of a clean graph.

Step 2 — Read status first

Never collapse any of these into "no adverse treatment."

Step 3 — Check the docket (federal origins)

Direct history — was this decision appealed, reversed, vacated, amended, or superseded on reconsideration — is the most common way a lower-court ruling stops being law. The citing graph reaches it only indirectly, and opinion_verify reads only sparse cluster fields and party-name overlap. The docket records it directly, in the clerk's own entries.

Do this whenever the origin is a federal district, circuit, or bankruptcy decision. PACER covers federal courts only — skip it for state origins and say so in the coverage line.

Sequence, cheapest first:

1. Resolve the docket. search_for_pacer_docket with court_id and the shortest distinctive party name; add docket_number from the opinion's metadata when you have it. case_name matches as a phrase, not as ANDed tokens, so a full caption returns zero: "Dorley" resolves, "Dorley South Fayette" returns nothing. A zero result here is far more often a query that was too long than a docket that is absent. Retry with a single party name before concluding the case is not in RECAP.

2. Confirm and orient. pacer_docket_view with include_entries: false — metadata only, fast. Check the parties, judge, and terminated date against the opinion so you know you have the right case.

3. List everything filed after the decision. search_within_pacer_docket with filed_after set to the origin's decision date and no document_type. No text query is needed — filed_after works on its own. This is the highest-value call in the sequence: it enumerates everything the court did after the opinion issued.

   Do not narrow this call with document_type: "orders". An appellate mandate is classified appeals, not orders, so the orders filter drops the single entry most likely to decide validity — MANDATE of USCA — Affirming in part, Reversing in part, Remanding. Take the unfiltered list and read it. If the docket is large enough that you must narrow, run orders and appeals as separate calls.

   (The caution against defaulting filed_after under Step 1 governs opinion_verify only. On search_within_pacer_docket, filed_after is the correct and intended filter.)

4. Target the dispositive entries. title_query for mandate, notice of appeal, amended judgment, reconsideration, vacate, remand, stay, amended opinion, substituted opinion, withdrawn, superseding, errata, clarify, alter or amend, Rule 59, Rule 60, relief from judgment. Use this to supplement item 3, never to replace it — some dockets return entries with an empty description, and those are invisible to any text query.

5. Read the (Attachments: ...) parenthetical in every entry description. An entry listing an Amended Opinion, Substituted Opinion, Corrected Opinion, or Errata is notice that the published opinion was changed — and the parenthetical is frequently the only evidence, because those attachments are often not archived as separate retrievable documents. Report a known, unretrieved amendment. Never record it as an absence.

6. Follow the appeal. For an appellate origin, or to learn how an appeal came out, search the reviewing court's own docket — search_for_pacer_docket with court_id set to the circuit plus the party names.

7. Before retrieving any PDF, batch the document IDs through check_pacer_availability (up to 300 at once). It reports which documents are actually archived and saves a failed pacer_document_view.

What the docket settles that the citing graph cannot: whether an appeal was taken at all; whether the mandate issued and what it said; whether the judgment was amended or vacated on reconsideration without any published opinion to cite; whether the case was consolidated or stayed.

Limits — state these whenever you rely on the docket.

• The opinion index and the docket are separate sources. A later opinion in the same case is often absent from the opinion database while sitting on the docket as a retrievable PDF. Failing to find a second opinion by citation or keyword search is not a finding. Check the docket before reporting a coverage gap — and check is_available before reporting a document as unretrievable.
• Coverage is the RECAP archive, which is crowd-sourced and incomplete. A missing entry is not evidence that nothing happened. Older dockets are thin.
• Docket text is the clerk's language, not a holding. "VACATED and REMANDED" gives you the disposition, not its scope — retrieve the opinion before characterizing it.
• A notice of appeal tells you an appeal was filed, not how it came out. Look for the mandate.
• PACER does not cover state courts.

If the docket shows the origin was reversed or vacated, the verdict is settled — retrieve the reviewing court's opinion to confirm and name it, then go to Step 5.

Step 4 — Triage

Direct history in the tool output

opinion_verify reads only cluster text fields and the docket linkage, and measured on this mirror generation those are almost never populated — history on ~0.5% of clusters, date_cert_granted on effectively none. Its silence is not evidence. Step 3 is the real check. What the tool does find reliably is same_litigation_candidate by party-name overlap — treat those as leads to confirm, not findings.

Direct history means same-litigation appellate treatment only — was this opinion reversed, vacated, or modified on appeal. A later unrelated case overruling the origin is adverse treatment, not direct history. Do not put it in the DIRECT HISTORY field.

Before calling something direct history, confirm the later court had authority to review or modify the origin. A coordinate court reaching the same litigation — after transfer, for instance — can disagree with the origin and reach the opposite result, but it cannot review it. That disagreement is worth reporting; report it as a conflict rather than as direct history, and name both decisions.

Finding citing opinions — run both routes

The citation graph and full-text search each miss cases the other finds. Run both on every load-bearing case. Neither alone is complete coverage.

1. The graph — opinion_verify, and show_citing_opinions where you need the list.
2. opinion_search on the case name in quotes — "Fought v. City of Wilkes-Barre". Catches citers the graph holds no edge for, including courts citing by short form, docket number, or Westlaw number.
3. opinion_search on the citation string in quotes — "466 F. Supp. 3d 477".

Deduplicate by case name and triage the combined set exactly as you would Tier A output — the snippets give you the citing court, the date, and the surrounding language.

Running court_scope at more than one setting is not a cross-check: every setting reads the same edge table, so a missing edge is missing at all of them. Agreement between scopes says nothing about completeness.

A zero or a thin graph is not a finding until the searches are run. State in the coverage line which routes produced the citing results.

Counting correctly

Three kinds of repetition appear in the index, and conflating them inflates your adverse-authority count: same cluster with different sub-opinions (Lawrence is three rows in Bowers's index, all cluster 130160); genuinely duplicate clusters, which the tool cannot merge (Georgia v. Public.Resource.Org under three IDs, Loper Bright under two); and collapsed sub-opinion records, which the tool does handle and reports.

Deduplicate by case name before counting or reporting adverse authorities. Bowers's 35-row index is roughly 24 distinct decisions. Column-by-column detail is in references/opinion-verify.md.

Reading order

Integrity warning → provenance and content_hash → graph size, admitted courts, unaudited courts, truncation → index → signal summary → analytics (the hub is the likely replacement rule; the co-cited list is the doctrinal line Leg 3 needs).

Tiers

Tier A is mostly noise, and you should expect that. Measured on Bowers: 6 Tier A, of which only Lawrence is actual treatment. Dobbs is Tier A because it lists Lawrence-overruling-Bowers in a string cite about overruled cases; Casey because the word "overruling" appears near a Bowers reference in a passage about overruling Roe; McDonald because Scalia discusses stare decisis. A 1-in-6 signal rate is normal. Zero Tier A does not mean valid — that is Leg 4's job.

Do not widen cooccurrence_window to catch more; it lowers precision far faster than it raises recall.

Matching treatment to the proposition — read for it

Validity is proposition-specific: a case gutted on one holding may be untouched on another. Nothing in the tool output tells you which holding a citing case engaged, and page numbers will not tell you either — CourtListener frequently lacks reporter pagination, so pages are absent from most snippets and unreliable where present. Do not try to filter by page.

Determine it by reading. For each Tier A candidate, read the passage around the origin reference and answer three questions:

1. Is this actually a reference to the origin? Snippet matching can fire on a party name shared with an unrelated case. Confirm the citation is to your case before going further.
2. Which of the origin's holdings is the court engaging? A case abrogating the origin's standing analysis says nothing about its merits holding.
3. Is the court doing something to the origin, or merely mentioning it? Inventorying it in a string cite, quoting a party's argument, and reciting history all read like treatment in a snippet and are not.

Never report treatment without having identified which holding it reached.

Retrieving full text without drowning

Full SCOTUS opinions exceed single-response limits. Do not try to read one straight through. fetch_opinion_file, save to disk, then read targeted regions — use the snippet start/end offsets and the origin's party name to locate the passage that matters. For a shortlist of ambiguous cases, submit_batch_screen (≤20 clusters) is cheaper than reading each.

Step 5 — Early exit

If Step 3 or Leg 2 produces a clear, express reversal, vacatur, or overruling that you have retrieved and read, and it (a) addresses the proposition at issue and (b) comes from a court that binds the target court — the finding is dispositive. Legs 3 and 4 are moot. Say so in the coverage line and stop.

Still apply the recursion rule: verify the overruling case is itself good law before naming it as the replacement.

Otherwise continue.

Step 6 — Leg 3, the upstream hop

The forward sweep only sees edges that exist. It cannot see the origin relying on a parent that was later killed without anyone telling the origin's citing history.

1. Identify the 3–8 authorities the origin leans on for the proposition — from the tool's upstream list, ranked by depth.
2. Call opinion_verify on each with filed_after = the origin's decision date. ~150–305 ms each.
3. A parent overruled or limited after the origin relied on it, on the same point, makes the origin AT RISK even with a spotless citing history.
4. Sibling check from the co-cited list — those cases are the doctrinal line.

Step 7 — Leg 4, the non-citation leg

A graph cannot detect a court that changed the law without citing anyone in the line. The tool does none of this.

1. State the current rule independently. Search the proposition in current doctrinal language, controlling jurisdiction, -dateFiled, last ~10 years. Read the recent authoritative statements, then compare to the origin. A new element, a shifted burden, a differently-phrased standard, a vanished exception — any is a silent-shift flag.
2. show_related_opinions for subject-matter neighbors that never cite the origin.
3. Statutory supersession — codes_search, check effective dates. A statutory amendment can abrogate a line while citing nothing.
4. Doctrinal resets — intervening en banc, state high court, or constitutional decision, named or not.

Step 8 — Verdict

VALID · VALID BUT NARROWED · AT RISK · QUESTIONED · INVALID FOR THIS PROPOSITION · INVALID · INDETERMINATE

Always qualified by the proposition. Name the replacement whenever the case is unusable.

Unaudited courts. 3,330 courts in the hierarchy, 205 human-audited; the state intermediate appellate layer is largely unaudited. The tool names them — pass that through.

Return to the caller

Roughly 250 words. Structured, quoting the operative language, no traversal narrative.

Copy code
CASE: <full citation>
PROPOSITION CHECKED: <the rule>
JURISDICTION: <target court>

STATUS: <one of the seven> — confidence <high/moderate/low>

DIRECT HISTORY (same litigation only): <reversal/vacatur/amendment chain. Say whether
the docket was checked and what it showed, or why it could not be (state origin, no
RECAP coverage).>

ADVERSE TREATMENT: <each: citing case, full cite, the operative language quoted, and
which of the origin's holdings it reached. DEDUPLICATED BY CASE NAME. Or "none located.">

AT-RISK FINDINGS: <upstream authority undermined after this case relied on it, or
divergence between this case and the current rule. Or "none.">

RECENT APPLICATION: <most recent case applying it for this proposition — or the silence>

IF UNUSABLE — WHAT REPLACES IT: <case or statute now stating the rule, with cite and
operative language. The graph hub is usually the candidate.>

COVERAGE: <small graph size and distinct-case count; courts admitted; mirror generation
and watermark; API supplement status; truncation/unscanned; unaudited courts; docket
check status; which routes produced the citing results; content_hash. Say which legs ran
and which were moot under Step 5.>

Most of the coverage line is copied from provenance rather than composed. A verdict without a stated boundary invites more reliance than the method can bear.

Worked example

Smith v. Acme, 900 F.3d 100 (5th Cir. 2018), cited for presumptive enforceability of a forum-selection clause in an employment contract under Texas law. Diversity; brief headed for N.D. Tex.

1. Resolve to a cluster ID. Call with include_upstream: true and, because this is Erie, additional_courts: ["tex","texcrimapp"]. Confirm all four courts in Courts admitted.
2. status: ok; no integrity warning; 41 rows, no truncation; 5 Tier A.
3. Federal origin, so check the docket: resolve it in ca5, then search_within_pacer_docket with filed_after the opinion date and no document_type. Panel rehearing denied, no en banc, mandate issued. No direct history.
4. Deduplicate: 5 rows are 3 distinct decisions (one appears as combined + dissent).
5. One is a dissent — note as pressure, not treatment. Two remain. Retrieve both and read the region around the snippet offsets.
6. The first engages Smith only on personal jurisdiction — a different holding. Set aside. The second, a 2023 ca5 panel (depth 7, out-degree 5, term "to the extent that"), engages the enforceability holding directly and narrows it to at-will employment, reserving fixed-term contracts.
7. Not an express overruling — no early exit. Continue.
8. Upstream: a 2011 Texas Supreme Court case, filed_after = Smith's date → clean. Top sibling → clean.
9. Leg 4: recent Texas authority is consistent with Smith as narrowed.
10. VALID BUT NARROWED — confined to at-will employment — confidence high.

Limitations to keep in mind

These bound what the evidence can support. Raise one in the verdict only when it actually bit on this case.

1. A zero result is never proof of validity.
2. Silent overruling — a court changing the rule without citing the origin or its line — is outside the reach of any citation-graph method. Leg 4 is the only route to it, and it is not exhaustive.
3. Flag-term recall in the tool is bounded by its curated list; the word searches in Step 4 are what widen it.
4. Snippets and tiers do not distinguish holding from dictum, argument, or quotation. Reading does — see Hard Rule 1 and Step 4.
5. Inherits every CourtListener gap: thin state appellate coverage, unpublished dispositions, eyecite failures, unreported orders, incomplete citation metadata.
6. CourtListener frequently lacks reporter pagination, so a pinpoint page often cannot be produced from tool output; take pincites from retrieved text.
7. opinion_verify's own direct-history detection is near-blind on this mirror generation; the docket check in Step 3 is the real one.
8. Docket coverage is the RECAP archive — crowd-sourced, incomplete, federal only. A missing entry is not evidence.

Related skills

Called by dingduff-legal-research, dingduff-legal-analysis, dingduff-legal-writing, and dingduff-citation-check. Note the capacity ceiling: a 20-citation cite-check is ~160 tool calls and ~2 minutes occupying both admission slots — do not run two at once. Citation form is dingduff-legal-citation-format; this skill checks whether a case is good law, not whether the cite looks right.