Sovereign Legal Research and Litigation

Nishi vs the field — every Nishi cell is measured against real organ source at emit time; each gap names the watch contract that will close it.

Nishi vs Harvey and CoCounsel and Westlaw Precision and LexisNexis Protege -- the research and litigation lane; the practice, document, e-sign and client-portal stack is measured on /compare/legalpractice against DocuSign, Clio, NetDocuments and PandaDoc

Layer 1 · Executive

Where we are. The sovereign research-and-litigation lane is deterministic where the field hallucinates: cite-or-abstain verification with NO LLM in the trust path (exceed), document-level citation redline (exceed), good-law treatment signals, the composed safe-to-cite test (exceed), an allow-list production screen (exceed), a fail-closed statute-of-limitations engine, redaction completeness as a hard block (exceed), prejudgment interest on an exact day count, privilege-log adequacy, litigation-hold defensibility, the 37(e) sanction ladder and e-filing timeliness. Open: a legal research corpus the sovereign search engine can be pointed at, a citator fed by real treatment data rather than fixtures, and an authority-graded answer surface over the verified lane.

Where we need to go. A legal research answer that a Utah practice can rely on: corpus-backed, every citation verified byte-for-byte against the banked source, every authority good-law-checked, and an abstention wherever the record cannot support the sentence -- the exact inverse of the 24 percent unsupported-citation wall the 2026 tools carry.

The unit. 1 u = one measured session-leg (estate calibration: graphics R21 in one leg 2026-08-15). Local evidence: the hold, privilege-log and e-filing libs each landed gated in one leg on 2026-08-18; estimates are relative to those.

Where we are: 3 open rungs. Verification and litigation engines present and exceeding; research corpus and answer surface absent. Counts measured at emit below this line.

Cost to a research corpus: 1.5 u. LG1 point the sovereign search engine at a banked legal corpus with the citation verifier on every hit.

Cost to a real citator: 2 u. LG2 treatment signals ingested from a declared public source with provenance, replacing fixture-only status.

Cost to an answer surface: 2 u. LG3 the research answer composed from LG1 hits through cite_scan_doc and cit_safe_to_cite, abstaining on any struck citation.

Research bar. Westlaw Precision and LexisNexis Protege is measured on corpus search, KeyCite and Shepard's treatment, research answers. Theirs: the research bar. Ours: LG1 LG2 LG3 measured on this page.

Research bar. Magesh et al. 2024 (hallucination-free?) is measured on unsupported-citation rates in legal AI tools. Theirs: the honesty bar we already beat by abstaining. Ours: LG3 carries the verifier into the answer.

12 of 15 capabilities measured|5 of them measured exceeds|3 open|coverage 800/1000|adoption 8 full / 4 partial

Layer 2 · Roadmap

Do this next — computed by the ranker, never chosen by a seat

Order from nx_compare_rank (nx_dr_ocm: (deficit + cost-of-delay + option + enables) x sponsor x self-sufficiency x momentum / cost). FINISH rows are rungs whose symbol is present but whose organ is short of full adoption: the cheapest closures on this board, listed before any new work. Stamp: # asof=1787883485 domain=legal target_version=0.1 rungs=6 done=3 open=3 finish=1 ranker=nx_dr_ocm

#	Stage	Rung	Priority	Derivation
F	FINISH	Safe-to-cite composition (LR2) cit_safe_to_cite	LIB-GATE-ONLY importers=1	imported only by validation organs (gates, tests, benches): wire it into a shipping program
#1	0.1	Legal corpus semantic search (LG1) legal_search_query	6000	v=21 m=2 c=7
#2	later	Verified research answer (LG3) legal_answer_compose	1600	v=16 m=1 c=10
#3	later	Citator fed by treatment data (LG2) cit_treat_ingest	700	v=7 m=2 c=20

Critical path — contract, done-rule, executor, cost

Rung	Closes with	Definition of done (pre-declared)	Executor	Est.
Cite-or-abstain verifier (LR0)	cite_verify	A citation passes only as an exact byte-span of a real source -- LANDED	Organ	0 u
Document redline (LR1) after LR0	cite_scan_doc	Every inline citation verified, unverifiable ones struck with a reason -- LANDED	Organ	0 u
Safe-to-cite composition (LR2) after LR0	cit_safe_to_cite	Verified AND good law, or refused -- LANDED	Organ	0 u
Legal corpus semantic search (LG1) after LR0	legal_search_query	The sovereign search engine over a banked corpus of opinions and statutes, each hit carrying its banked source so cite_verify can run on it; gate proves a known holding is found at rank 1 with a verifiable span and a fabricated citation finds nothing	Organ	1.5 u
Citator fed by treatment data (LG2) after LR2	cit_treat_ingest	Treatment signals (followed, distinguished, overruled) ingested from a declared public source with provenance per row into the CID plane the citator already reads; gate proves an overruled fixture authority reads unsafe after ingest and the worst-wins rule still holds	Organ	2 u
Verified research answer (LG3) after LG1,LG2	legal_answer_compose	An answer composed from LG1 hits, redlined by LR1 and gated by LR2, that ABSTAINS on any struck citation and names why; gate proves a question whose only support is overruled yields an abstention, never a confident paragraph	Organ	2 u

Milestones

Milestone	Rungs	Cumulative
M1 · A corpus to search	LG1	1.5 u
M2 · A citator with real data	LG2	3.5 u
M3 · An answer that abstains	LG3	5.5 u

Layer 3 · Engineering

How this is scored. Every Nishi mark is measured: the generator reads the real organ source on disk and requires the implementing symbol to exist (no self-grading). A watching tag names the organ and symbol contracted to close a gap — the mark flips itself on the next compare beat when that workstream ships, and the comparewatch- plane row flips with it. The flip is necessary, not sufficient: it proves the symbol exists, never that the capability is good. The bar is the rung's pre-declared done-rule, proven by its gate — a symbol shipped without the behaviour behind it is a defect, and the flip is exactly what makes that defect visible instead of quiet. Competitor marks record documented capability presence — presence, not depth or scale. Adoption is measured too: every measured row carries where its organ stands on the estate's ladder (source → built → promoted → registered → invoked; libraries by importer reach minus validation importers; gates by the execution surfaces that run them). A row is fully adopted only at the top of its ladder; anything short is tagged partial with the exact remedy, so a build nobody promoted can no longer read as shipped. Census stamps: importers asof 1787849099, gate census asof 1787855507 (unix seconds; -1 = census absent).

Capability matrix — measured against source

◉ leads / measured exceed● present◐ partial○ absent · click any capability for its evidence

Capability	Nishi	Harvey	CoCounsel	Westlaw	LexisNexis
Cite-or-abstain citation verification Measured exceed: `cite_verify` in `runtime/nx_cite_lib.nx`, verified at emit. Deterministic verifier, NO LLM in the trust path: a citation passes only if its quote is an exact byte-span of a real source. The 2026 wall is ~24pct of legal answers citing unsupported law [magesh2024]; a subtly altered quote ABSTAINS here Adoption: `LIB-WIRED importers=3 nonval=1` — fully adopted (top of its ladder).	◉	◐	◐	●	●
Document-level citation redline Measured exceed: `cite_scan_doc` in `runtime/nx_cite_lib.nx`, verified at emit. Scans a whole drafted paragraph, verifies every inline citation, strikes the unverifiable ones with a reason; a filing is clean only at zero struck. The unterminated-citation dodge is itself a strike [cocounsel] Adoption: `LIB-WIRED importers=3 nonval=1` — fully adopted (top of its ladder).	◉	●	●	●	●
Citation treatment signals good-law check Measured: `cit_status` exists in `runtime/nx_citator_lib.nx`, verified at emit. Shepard's/KeyCite analog: worst-wins so one overruling kills reliance despite any followings. Westlaw and Lexis are the reference here [westlaw-keycite]; ours is on the tamper-evident CID plane [lexis-shepards] Adoption: `LIB-GATE-ONLY importers=1` — PARTIAL: imported only by validation organs (gates, tests, benches): wire it into a shipping program. adoption LIB-GATE-ONLY importers=1	●	○	○	◉	◉
Composed safe-to-cite verified plus good law Measured exceed: `cit_safe_to_cite` in `runtime/nx_citator_lib.nx`, verified at emit. The composition incumbents do not close: a citation is safe ONLY if the quote verifies AND the authority is still good law. A real quote from an OVERRULED case is refused [harvey] Adoption: `LIB-GATE-ONLY importers=1` — PARTIAL: imported only by validation organs (gates, tests, benches): wire it into a shipping program. adoption LIB-GATE-ONLY importers=1	◉	○	○	●	●
E-discovery production allow-list screen Measured exceed: `ed_is_producible` in `runtime/nx_ediscovery_lib.nx`, verified at emit. Production is an ALLOW-LIST not a deny-list: a doc is producible only if affirmatively reviewed+responsive, so a privileged OR unreviewed doc cannot leak by omission. Incumbents review-then-deny-list; ours cannot disclose by construction [frcp-26] Adoption: `LIB-GATE-ONLY importers=1` — PARTIAL: imported only by validation organs (gates, tests, benches): wire it into a shipping program. adoption LIB-GATE-ONLY importers=1	◉	●	◐	◐	◐
Statute-of-limitations deadline engine Measured: `sol_status` exists in `runtime/nx_sol_lib.nx`, verified at emit. Exact proleptic-Gregorian deadline math (leap-year and Feb-29->Feb-28 clamp correct), tolling, and the discovery rule. Westlaw/Lexis deadline calculators are the reference on jurisdiction rule sets; ours is FAIL-CLOSED -- it returns UNDETERMINED-FOR-REVIEW on an unestablished accrual date rather than the confident date a rule-set calculator always produces. Missing a limitations deadline is the top legal-malpractice claim. GRADED PRESENT, NOT EXCEEDS (2026-08-07): this row marks Westlaw and LexisNexis Best on this same axis, and both cannot lead it. The note above already concedes they are the reference on jurisdiction rule sets; our claim is the narrower one of fail-closed correctness, which is a different tradeoff, not a superiority Adoption: `LIB-WIRED importers=3 nonval=2` — fully adopted (top of its ladder).	●	●	●	◉	◉
Redaction completeness for production Measured exceed: `redact_production_ok` in `runtime/nx_redact_lib.nx`, verified at emit. The content half of e-discovery production (nx_ediscovery is the allow-list): a document is producible only if reviewed AND redaction-clean. Deterministic byte-exact SSN detection with precise number boundaries (no false positive on a dd-dd-dddd grouping or an SSN-shaped run embedded after digits) plus privilege-marker scan; FAIL-CLOSED as a HARD BLOCK, not a flag-for-human-review, so PII/privilege cannot leak by omission. Incumbents flag; ours refuses Adoption: `LIB-GATE-ONLY importers=1` — PARTIAL: imported only by validation organs (gates, tests, benches): wire it into a shipping program. adoption LIB-GATE-ONLY importers=1	◉	●	◐	◐	◐
Prejudgment interest / damages engine Measured: `dmg_simple_interest` exists in `runtime/nx_damages_lib.nx`, verified at emit. Statutory prejudgment interest computed on an EXACT proleptic-Gregorian day count (composes nx_sol's date engine, not a re-implementation): simple and annually-compounded, integer minor units, at a declared statutory rate. FAIL-CLOSED on a judgment date before accrual or negative inputs. Day-count and compounding errors are exactly where damages calculators go wrong. GRADED PRESENT, NOT EXCEEDS (2026-08-07): same rule as the SOL row -- Westlaw and LexisNexis are marked Best here, so an exceed flag would contradict this row own competitor grades Adoption: `LIB-WIRED importers=2 nonval=1` — fully adopted (top of its ladder).	●	●	●	◉	◉
Privilege log adequacy under FRCP 26(b)(5)(A) Measured: `pl_log_adequate_pure` exists in `runtime/nx_privlog_lib.nx`, verified at emit. Each entry must carry the elements that let the other side assess the claim (attorney-client versus work-product tier, waiver posture); an inadequate entry fails the log, never a silent pass [frcp-26]. Measured 2026-08-18 from the shipped lib; graded present, not exceeds Adoption: `LIB-WIRED importers=2 nonval=1` — fully adopted (top of its ladder).	●	◐	◐	●	●
Litigation hold defensibility -- duty trigger, custodian notice and acknowledgment, system suspension, premature release refused Measured: `hd_defensible` exists in `runtime/nx_hold_lib.nx`, verified at emit. The EDRM left side (preservation) beside the production allow-list: a hold is defensible only when the duty has attached, notice has issued, every custodian has acknowledged and no system is left unsuspended; a release before the duty lapses is named premature. Uncontested by these four research tools -- the hold reference is the e-discovery vendor class, not on this board. Measured 2026-08-18 from the shipped lib Adoption: `LIB-WIRED importers=2 nonval=1` — fully adopted (top of its ladder).	●	○	○	○	○
FRCP 37(e) spoliation sanction ladder Measured: `hd_sanction_tier` exists in `runtime/nx_hold_lib.nx`, verified at emit. Lost ESI graded none, curative or severe from reasonable steps, restorability and intent -- the rule's own structure; Westlaw and Lexis carry practice guidance on 37(e) (Part), neither computes the tier from the matter record. Measured 2026-08-18 from the shipped lib Adoption: `LIB-WIRED importers=2 nonval=1` — fully adopted (top of its ladder).	●	○	○	◐	◐
E-filing timeliness -- submission is not filing, rejection and relation-back on the deadline Measured: `efl_status_pure` exists in `runtime/nx_efile_lib.nx`, verified at emit. A brief submitted before midnight and rejected the next morning was NEVER FILED unless relation-back applies inside the cure window; status is one of not-filed, pending, filed-timely, filed-late or rejected and a late filing names its days late. Westlaw and Lexis practice guidance covers court rules (Part); the Utah connector itself is a watch on /compare/legalpractice. Measured 2026-08-18 from the shipped lib Adoption: `LIB-WIRED importers=2 nonval=1` — fully adopted (top of its ladder).	●	○	○	◐	◐
Legal research corpus semantic search Open — watching `runtime/nx_legal_search.nx` : `legal_search_query`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. Declared gap; the sovereign search engine exists, pointing it at a legal corpus is the rung. Named as a watch contract on 2026-08-18 so the hive can flip it watching legal_search_query	○	◐	◐	◉	◉
Citator fed by real treatment data (provenanced ingest) Open — watching `runtime/nx_citator_lib.nx` : `cit_treat_ingest`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. The treatment-signals row above is present on fixture data; Westlaw KeyCite and Shepard's are fed by editorial treatment pipelines [westlaw-keycite] [lexis-shepards] -- the contract is ingest from a declared public source with provenance per row into the CID plane the citator already reads, worst-wins preserved watching cit_treat_ingest	○	○	○	◉	◉
Verified research answer that abstains Open — watching `runtime/nx_legal_search.nx` : `legal_answer_compose`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. Harvey and CoCounsel compose research answers [harvey] [cocounsel]; ours must compose ONLY from corpus hits redlined by cite_scan_doc and gated by cit_safe_to_cite, abstaining and naming why on any struck citation -- the answer surface over the verified lane, not a new trust path watching legal_answer_compose	○	◉	◉	◉	◉

Risk register

Risk	Likelihood x impact	Mitigation
A corpus search that returns a hit without its banked source cannot be verified and reintroduces the hallucination wall	possible x high	LG1 done-rule: no hit without a verifiable span.
Treatment data from a single source with no provenance is a citator that lies quietly	possible x high	LG2 rows carry source and as-of; an unprovenanced row is refused at ingest.

On these two registers. Rows are declared in the domain's plan file and carry the debt id, which is the join key back to the sovereign debt plane — that plane, not this page, is the authority on state. Reconciling them automatically (the regen reading the plane and refreshing these rows) is a named, owed rung; until it lands, treat an id here as a pointer to look up, not a status to trust.

Honest verdict. The coverage above is capability presence measured against source — not depth, scale, or polish, where mature rivals may lead. Exceeds are claimed only where a mechanism backs them. Every open gap is a watch contract: it names the organ and symbol that closes it, and this page flips the cell itself when that workstream ships.

Person · product · place — not yet measured for this domain

Every compare carries this layer. Declare knowledge/compare/legal.ppp (rows surface|nishi or c1..c4|label|url|connect naming OUR live surface and each rival's front door), run nx_ppp_probe domain legal, and this section fills itself on the next beat: the same ruler on both sides — privacy and CX (third-party hosts, tracker classes, cookies, security headers), design and longevity (design hygiene, computed WCAG contrast, render-blocking resources, unsized media, script weight, theme and motion queries), findability (landmarks, skip link, on-site search, breadcrumb, headings, internal links).

References

Beyond a link list. Every reference below resolves twice — the publisher's copy and, where banked, the estate's own non-rottable library mirror with a content pin — and carries its evidence class plus the exact claim on this page it grounds. Keyed marks like [key] in the matrix notes jump here. A dash means honestly absent, never assumed.

[harvey] Harvey AI. Harvey -- professional-class AI for legal work (product surface), harvey.ai, accessed 2026-08-18. publisher · read in our library knowledge/fetched/cmp_legal_harvey.html · pin h20e6cf3d4f1c82bc89ac98b0e634e678be80f099b4ad829936acb55cc6313b33 · accessed 2026-08-18 · vendor-docGrounds: The Harvey column: an LLM assistant for research, drafting and review with no citator of its own, which is why 'Citation treatment signals good-law check' and 'Composed safe-to-cite verified plus good law' grade it 0 while the drafting rows grade it Part.
[cocounsel] Thomson Reuters. CoCounsel Legal -- AI legal assistant reasoning from Westlaw primary law and Practical Law, with agentic Deep Research and guided workflows (product page), accessed 2026-08-18. publisher · read in our library knowledge/fetched/cmp_legal_cocounsel.html · pin h43d16a4ddd7a8051449f674b6aa0c3997daf4cce513e14d534205791ec81ed31 · accessed 2026-08-18 · vendor-docGrounds: The CoCounsel column: research, drafting and document-review assistance graded Part on 'Cite-or-abstain citation verification' and Yes on the redline row; its citation checking rides Westlaw rather than a byte-exact quote verifier.
[westlaw-keycite] Thomson Reuters. KeyCite -- Westlaw citator: instantly verify whether a case, statute, regulation or administrative decision is still good law (product page), accessed 2026-08-18. publisher · read in our library knowledge/fetched/cmp_legal_westlaw-keycite.html · pin h6c1e27ee052cc1247eee6fbf8e2a8965d666449962199230e7dc124dfc36062c · accessed 2026-08-18 · vendor-docGrounds: The Westlaw column Best code on 'Citation treatment signals good-law check' and its Best codes on 'Statute-of-limitations deadline engine' and 'Prejudgment interest / damages engine': KeyCite is the treatment-signal reference the nx_citator worst-wins analog is measured against.
[lexis-shepards] LexisNexis. Shepard's Citations Service -- validate case law, statutes and regulations, find splits of authority (product page), accessed 2026-08-18. publisher · read in our library knowledge/fetched/cmp_legal_lexis-shepards.html · pin ha0ac149b160703df2831125717d98c45bf69315365600d1dbcaa18f470b592b4 · accessed 2026-08-18 · vendor-docGrounds: The LexisNexis column Best code on 'Citation treatment signals good-law check': Shepard's is the second citator reference named in that row's note (Shepard's/KeyCite analog), and the reason ours is graded present-not-best on that axis.
[magesh2024] Magesh, Surani, Dahl, Suzgun, Manning, Ho. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. arXiv:2405.20362, 2024 (Stanford RegLab / HAI): measured hallucination rates of 17 percent and over 33 percent for two leading RAG-based legal research products. publisher · read in our library knowledge/fetched/cmp_legal_magesh2024.html · pin hf724e03a034bb9f1b3a099f1782fd1219d67a3eb9c7cd721a7de8ac97f6c66f0 · accessed 2026-08-18 · published-paperGrounds: The 'Cite-or-abstain citation verification' row and its note that a large share of AI legal answers cite unsupported law: this is the published measurement of that wall, and the reason nx_cite_lib keeps NO LLM in the trust path and abstains on any quote that is not an exact byte-span of a real source.
[frcp-26] Federal Rules of Civil Procedure, Rule 26: Duty to Disclose; General Provisions Governing Discovery -- including 26(b)(5) claims of privilege and the privilege-log requirement. Legal Information Institute, Cornell Law School. publisher · read in our library knowledge/fetched/cmp_legal_frcp-26.html · pin h589a12e473a7b17945707c882551a9bf7ca65394d0b5f474cb11f2e6ec90e3bc · accessed 2026-08-20 · published-standardGrounds: The 'E-discovery production allow-list screen', 'Redaction completeness for production' and 'Privilege log adequacy under FRCP 26(b)(5)(A)' rows: the discovery-scope and privilege-assertion duties these organs enforce fail-closed (a privileged or unreviewed document cannot be produced by omission, an inadequate log entry fails the log) are this rule's.

generated by nx_swcompare_matrix (sovereign NishiLang organ) from knowledge/compare/legal.matrix · every Nishi cell verified against organ source at emit time · watch cells re-measured on every compare beat · zero JS, zero trackers