retrieval: make authority proposition treatment machine-checkable

[zackees]

Context

TAXONOMY.md requires important propositions to be labeled as holding, dicta, party-allegation, requested-relief, awarded-relief, settlement-term, official-license-text, inference, or unresolved. The authority template records these distinctions only in prose and supplies no structured proposition/pinpoint field.

That gap is most dangerous for complaints, pending cases, temporary orders, defaults, and settlements. A summarizer can promote a correctly qualified sentence about requested source publication or a settlement term into a general judicial remedy because card-level posture alone does not label each proposition.

Proposal

Add a machine-checkable proposition structure containing treatment, proposition text, source pinpoint, jurisdiction, and limitation. Migrate the highest-risk authority set first, then gate all new cards. Ensure answer generation cannot cite a requested remedy or settlement term as a holding or awarded relief.

Acceptance criteria

• templates/AUTHORITY.md defines a structured proposition representation with controlled treatment values and direct pinpoints.
• All complaint, pending, TRO/preliminary-injunction, default, and settlement cards are migrated in the first pass.
• Each migrated proposition records treatment, source/page or docket pinpoint, jurisdiction, and a material limit.
• A validator rejects missing/unknown treatment labels for the high-risk set and for all newly added cards.
• Retrieval tests for can a court order source publication? cannot use requested-relief, party-allegation, or settlement-term as holding or awarded-relief.
• Card-level procedural posture remains available and does not substitute for proposition-level treatment.
• Existing prose qualifications are preserved until attorney-reviewed structured entries replace them.

Decisions

• Confirmed retrieval-control gap, P1. This issue does not assert that current prose is legally wrong.
• Migrate high-risk cards first. A staged migration controls workload while immediately reducing the largest overstatement risk.
• Pinpoints are required. A label without a source location is not independently auditable.

Open questions

• Should every proposition receive a stable local identifier for citation and supersession?
• Should all 102 cards be migrated in one pass after the high-risk validator proves useful?
• Can page/paragraph fragments be normalized across HTML, PDF, docket, and license-text sources?

Related issues

• Parent: audit: adversarially review the paralegal research corpus #19
• Corpus creation: docs: build a license-agnostic copyleft enforcement research corpus #17

[zackees]

Sol adversarial review - 2026-08-25

Verdict: confirmed but narrowed; leave open. The defect is not that the corpus has no machine-readable safeguards. All authority cards already expose card/source-level procedural_posture, precedential_weight, authority_type, authority_status, jurisdiction, topics, and source-quality fields; 19 of 20 matters also use a consistent Markdown Authority | Proposition supported | Posture and limits table, and the proposition map supplies human-readable routing. Those controls partially disprove the broad phrase ?only in prose.?

The narrower core finding is valid: there is no schema-validated, proposition-level representation tying each atomic assertion to its treatment and exact source artifact/locator. There is also no checked-in corpus validator, retriever, or answer-generation test that consumes such data. Card-level metadata cannot safely classify mixed assertions inside one source.

Reproducible evidence

At main commit 40ee7108f4b46b6b8fb1976f5c8b3cd996bdfb42:

(rg -l '^id: AUTH-' paralegal-research/authorities -g '*.md' | Measure-Object).Count
# 102

(rg -l '^\| Authority \| Proposition supported \| Posture and limits \|' paralegal-research/matters -g '*.md' | Measure-Object).Count
# 19

rg -l '^(propositions?|treatment|pinpoint|source_pinpoint|proposition_id):' paralegal-research/authorities -g '*.md'
# no matches

rg -l 'paralegal-research|procedural_posture|precedential_weight|proposition' tools tests .github -g '*.py' -g '*.yml' -g '*.yaml'
# no matches

The only executable validator/test surface is for source-license headers (tools/license_headers.py, tests/test_license_headers.py, and header-policy.schema.json). The PR #18 validation described controlled-metadata checks, but no reusable corpus validator or retrieval consumer landed. PR #18's CodeRabbit review was skipped because the PR had 140 files; there were no inline review comments.

A deterministic first-pass high-risk predicate over existing metadata?posture in complaint, temporary-restraining-order, preliminary-injunction, default-judgment, or settled; or authority_type: settlement; or an authority_status containing pending/settled?selects 13/102 cards:

AUTH-0003, AUTH-0004, AUTH-0005, AUTH-0008, AUTH-0033, AUTH-0039, AUTH-0043, AUTH-0044, AUTH-0045, AUTH-0046, AUTH-0053, AUTH-0060, and AUTH-0151.

The overlapping category counts are: complaint 2; pending-status 2; TRO 1; preliminary injunction 2; default 1; settled posture 5; settled status 3; settlement type 1. This needs a documented predicate, not a hand-maintained list: AUTH-0053's current-pending-later-history-check also shows that substring matching ?pending? conflates docket posture with a later-history freshness warning (#22).

Actual cards prove the required cardinality is one source to many atomic propositions:

• AUTH-0053 records requested publication, rejection of publication, four categories of awarded money/costs, and an infringement holding.
• AUTH-0151 awards limited trademark relief, denies preliminary GPL distribution relief, discusses cure and irreparable harm, and makes no final GPL infringement finding?all in one order. A card-level ?awarded relief? signal would produce the exact false source-publication result this issue aims to prevent.
• AUTH-0033 records requested TRO/preliminary relief and expedited discovery, denial of all of it, and no final merits decision.
• AUTH-0060 contains an interim holding plus a territorially limited EU compliance order and penalty mechanism.
• AUTH-0008 reports several negotiated terms from a secondary settlement announcement without the operative agreement.

Corrections and loopholes to address

1. Do not claim all treatment is unstructured. Card-level posture/weight and matter tables are real safeguards. State the gap as ?no deterministic proposition-level schema, validation, or consuming retrieval path.?
2. Do not model one treatment per card or source. Use a propositions[] collection of atomic assertions. A compound sentence that bundles a holding, requested relief, and an award merely moves the prose ambiguity into YAML.
3. Treatment and outcome are not obviously one axis. The current vocabulary has requested-relief and awarded-relief, but no denied-relief, reserved, dismissed, or moot. Calling the denial in AUTH-0033 a holding loses remedy outcome; calling it awarded relief is wrong. Decide whether treatment, disposition, and epistemic status are separate controlled fields.
4. A single pinpoint creates false precision. Unpaginated HTML, docket entries, secondary announcements, and researcher inferences do not share one locator form. An inference may rest on several sources/search logs and has no honest single source pinpoint. Use typed, possibly multiple source references (page, paragraph, section, docket-entry/document, stable anchor, or unavailable with reason); never manufacture a page/paragraph to satisfy CI.
5. A scalar proposition jurisdiction is ambiguous. Court/forum, governing law, precedential target, and territorial scope differ. AUTH-0060, for example, is a Dutch appellate interim decision with an EU-limited command involving a Hong Kong entity. Inherit source/card jurisdiction where appropriate and permit typed proposition-level scope overrides instead of copying a string that can drift.
6. holding is not synonymous with final, merits, or binding. Retrieval must combine proposition treatment with source posture, weight, later history, source quality, and the proposition's limitation. A preliminary order can contain holdings while remaining interlocutory.
7. Schema validity cannot prove the legal label is correct. A validator can check enums, references, and invariants; it cannot determine holding versus dicta, whether reported terms match an unseen agreement, or whether a proposition is overbroad. Preserve the attorney-review gate and record proposition review status/provenance.
8. Do not migrate known-bad prose mechanically. research: correct source and posture defects in key U.S. license cases #21 identifies source/posture defects in AUTH-0003, AUTH-0005, and related records. Verify the source artifact/order first, then encode propositions; otherwise the new structure fossilizes an existing error.
9. ?All newly added cards? needs an enforceable boundary. A new card could self-declare legacy and bypass a full-corpus validator. Check in the finite legacy-ID allowlist after the first migration and reject legacy status for any other ID; optionally require migration whenever a legacy card is substantively edited.
10. A fixture-only retrieval test is insufficient. No answer generator currently exists here. The acceptance criterion must either add an executable local consumer that emits proposition IDs/source refs, or identify and test the actual downstream consumer. A toy test that never reads the 102 cards does not close the risk.
11. Coordinate, do not merge, adjacent issues. governance: make stale-source handling operational #22, retrieval: align matter schema with doctrine, remedy, and time taxonomy #23, governance: repair matter-link and integration-state contradictions #26, and metadata: separate non-case legal effect from precedential weight #27 are not duplicates, but they all require parsing/validation. They should share one versioned schema loader and diagnostic format rather than four regex validators with divergent semantics.

Stronger acceptance criteria

• Define a versioned proposition schema and checked-in examples for a complaint, a mixed preliminary order, a default, a settlement report, a final merits order, and an official license text.
• Model source_artifacts[] separately from propositions[]. Each atomic proposition has a stable source-scoped ID, exact source reference(s), typed locator(s), treatment/outcome fields, proposition text, material limitation(s), and any proposition-specific jurisdiction/territorial override. The validator rejects dangling refs and duplicate IDs.
• Resolve whether treatment is single-valued or orthogonal to disposition/status. Add an honest representation for denied/reserved/dismissed relief and for multi-source researcher inference; do not overload holding or unresolved.
• Coordinate non-case proposition roles with metadata: separate non-case legal effect from precedential weight #27. The present treatment vocabulary has official-license-text but no equally clear statutory/rule/guidance text role, so an all-102 migration would otherwise force bad labels.
• Generate and commit the high-risk migration report from a documented predicate. Migrate the 13 current IDs above, with explicit review of metadata corrections from research: correct source and posture defects in key U.S. license cases #21 and with no silent assumption that the predicate catches a misclassified card.
• Keep the remaining 89 IDs on an explicit finite legacy allowlist. CI rejects a new ID on that list, rejects missing/unknown treatment values for migrated cards, and rejects high-risk cards that remain legacy. A substantive edit to a legacy card should either migrate it or require an explicit reviewed exception.
• Preserve prose during migration and run a parity check showing that structured entries neither omit nor broaden claims, requested/actual relief, limits, and contrary outcomes. Structured data supplements the human context; it does not imply attorney approval.
• Add a cross-platform corpus validator to the normal CI workflow and unit-test malformed arrays, compound/missing source refs, unsupported locators, invalid enums, incompatible source/treatment combinations, and the legacy-boundary bypass.
• Test an executable retrieval path against the real corpus?not only fixtures. For ?can a court order source publication??, assert the returned evidence includes proposition IDs and roles; does not present AUTH-0005 allegations, AUTH-0008 reported settlement terms, AUTH-0004 default relief, or AUTH-0033 requested/denied relief as a contested final award; treats AUTH-0060 as interim/territorially limited; records AUTH-0053's rejection of publication; and does not let AUTH-0151's trademark award become GPL source relief.
• Add mutation tests proving that changing a treatment/source ref changes or blocks retrieval, and that a holding on a preliminary source is still qualified by posture and weight.
• Document which structured changes require attorney review, who may approve treatment labels, and how proposition IDs are superseded without rewriting citation history.

Staged migration

1. Semantics first: settle the axes (treatment, outcome/disposition, unresolved/review status), source-artifact model, locator types, and counsel questions; encode adversarial fixtures from the five mixed cards above.
2. Infrastructure: build one shared parser/schema/validator for governance: make stale-source handling operational #22/retrieval: align matter schema with doctrine, remedy, and time taxonomy #23/retrieval: make authority proposition treatment machine-checkable #24/governance: repair matter-link and integration-state contradictions #26/metadata: separate non-case legal effect from precedential weight #27, produce a baseline report, and add the finite 89-card legacy allowlist.
3. High-risk pass: correct research: correct source and posture defects in key U.S. license cases #21 source identity/posture issues, then migrate the 13-card predicate set and run the real source-publication query test. Keep all prose and review status.
4. Forward gate: require the schema for every new card and every substantively changed legacy card; expose legacy/migrated state in the authority index.
5. Broader migration: prioritize authorities actually used in answer-ready matters, then migrate the remaining corpus. Remove the allowlist only after parity review and attorney approval of treatment semantics.

Questions for legal review

• Is holding limited to a court-decided legal proposition, and how should findings, reasons, interim rulings, and operative dispositions be separated?
• Are treatment and relief disposition orthogonal? What labels should represent denied, partly granted, reserved, dismissed, moot, or vacated relief?
• May a party/representative announcement support settlement-term, or must it be labeled as a reported term until the operative agreement is authenticated?
• Should researcher inference and unresolved be source treatments at all, or separate synthesis/research-status fields with multiple source/search-log references?
• What locator is sufficient for unpaginated HTML, foreign originals, docket reproductions, and inaccessible instruments, and when must a proposition remain non-answer-ready?
• Which notion of jurisdiction belongs at proposition level: deciding forum, governing law, binding/persuasive target, or geographic scope of relief?
• Are stable proposition IDs immutable citations that require explicit supersession, or may their text/treatment be edited in place after review?
• How should statutes, rules, licenses, settlements, and guidance be represented consistently with metadata: separate non-case legal effect from precedential weight #27 without pretending they share case-law ?precedential? treatment?

History / duplicate / resolution check

The relevant corpus and template history lands in PR #18 / commit 40ee710; no later default-branch commit, open/merged PR, or issue timeline event implements #24. The issue has no resolving PR and, before this review, had no comments. Searches for proposition/retrieval-treatment issues found #24 plus complementary #19/#21/#22/#23/#27, but no true duplicate. Therefore closure would be unsupported.

This is a retrieval/governance assessment for attorney review, not a legal conclusion, license decision, or change to LICENSE.