jet52
diff --git a/‎skills/jetredline/SKILL.md‎
Lines changed: 17 additions & 17 deletions b/‎skills/jetredline/SKILL.md‎
Lines changed: 17 additions & 17 deletions
@@ -344,7 +344,7 @@ $VENV_PYTHON $SKILL_DIR/extract_text.py --input <draft.docx> --output <TMPDIR>/<
 6b. **Delegate Pass 7** (dissent/concurrence cross-check) to a subagent if a dissent or concurrence was identified in Step 0 and `DOC_TYPE == opinion` — see Pass 7 below
 7. **Pass 2 routing:** If the opinion has **more than 30 paragraphs**, delegate Pass 2 to a subagent — see "Delegated Pass 2" below. Otherwise, perform Pass 2 in main context. **Pass 5** (analytical rigor) is always performed in main context. Pass 2 (when not delegated) and Pass 5 can proceed in parallel with subagents.
 8. Collect subagent results from Passes 1, 3, 4, 6, 7, and (if delegated) Pass 2 — **use the `TaskOutput` tool**, not Bash `tail`
-8a. **Caption mismatches → edits (closed-loop rule).** Check the Pass 3B results table for Caption Check mismatches. Generate a correcting tracked-change edit **only** when the *same* citation resolved to an official caption and the draft's name is a near-match typo of it (high `name_similarity`) — e.g., a comment "Official caption per 2023 ND 219: 'Tracey v. Tracey'". **Never** harmonize names across two *different* citations (different cite = presumptively different case), and do not auto-correct low-similarity or unresolved names — add a comment for human review instead. Include qualifying edits in the edits JSON alongside Pass 2/3A/5 edits. Likewise apply the **parallel-cite corrections** flagged by Pass 3B's parallel-cite check — add a missing N.W.2d/N.W.3d parallel to a *full* cite, or fix a wrong one, using the `formatted` value from ndcourts-mcp.
+8a. **Caption mismatches → edits (closed-loop rule).** Check the Pass 3B results table for Caption Check mismatches. Generate a correcting tracked-change edit **only** when the *same* citation resolved to an official caption and the draft's name is a near-match typo of it (high `name_similarity`) — e.g., a comment "Official caption per 2023 ND 219: 'Tracey v. Tracey'". **Never** harmonize names across two *different* citations (different cite = presumptively different case), and do not auto-correct low-similarity or unresolved names — add a comment for human review instead. Include qualifying edits in the edits JSON alongside Pass 2/3A/5 edits. Likewise apply the **parallel-cite corrections** flagged by Pass 3B's parallel-cite check — add a missing N.W.2d/N.W.3d parallel to a *full* cite, or fix a wrong one, using the `formatted` value from ndlaw.
 8b. **Treatment flags → comments.** If the Pass 3C overruling scan flagged any cited case for possible negative treatment, add a *comment* (not an edit) on each occurrence, quoting the citing context, for human review. Never auto-edit on a treatment signal.
 9. **If user requested tracked-changes .docx** (both or tracked-changes only): Produce tracked-changes .docx output using the batch edit workflow (see Step 9 details below)
 10. **If user requested analysis document** (both or analysis only): Produce the companion analysis document (incorporating all subagent results). If also producing .docx, create both outputs in the same response
@@ -356,22 +356,22 @@ python3 ~/.claude/skills/jetredline/provenance.py --file <stem>-ANALYSIS.md \
 Footer, e.g.: *Report generated by Claude Opus 4.8 (claude-opus-4-8) using jetredline v4.4.0 on 2026-06-05. AI-generated first draft for internal use; verify all citations and findings before relying.* Version and date are sourced deterministically by the script (version from frontmatter/`VERSION`, date from the system clock); you supply only `--model` from your runtime context — the script never guesses the model. The stamp is idempotent. **Stamp only the analysis document — never the tracked-changes `.docx`, which is the user's own opinion text and must not be altered with a footer.** In **audit mode** (no files written) and **web mode** (analysis returned inline, no filesystem), skip the script; in web mode you may append the same footer line manually as the report's last line.
 11. **Generate citation review HTML** (CLI and Cowork): After Pass 3 completes, generate an interactive citation review page for human verification.
 
-    **11a. Refresh ND authority text + direct URLs from ndlaw (zero token cost).** Run the export script first — it pulls each cited ND opinion's authoritative text and the court's direct opinion URL (`https://www.ndcourts.gov/supreme-court/opinions/<id>`) from the ndlaw corpus into `~/refs`, plus a metadata map for the review page. It auto-selects a backend: a local `opinions.db` (`NDLAW_DB` env or the default dev path), else a deployed ndcourts-mcp instance over Streamable HTTP (`NDLAW_URL` + `NDLAW_AUTH` env, or `--url`/`--auth`; for Claude Code users the URL and Basic-Auth header are in their MCP config — `claude mcp get ndlaw`). The script speaks to the server directly, so no opinion text passes through model context.
+    **11a. Refresh ND authority text + direct URLs from ndlaw (zero token cost).** Run the export script first — it pulls each cited ND opinion's authoritative text and the court's direct opinion URL (`https://www.ndcourts.gov/supreme-court/opinions/<id>`) from the ndlaw corpus into `~/refs`, plus a metadata map for the review page. It auto-selects a backend: a local `opinions.db` (`NDLAW_DB` env or the default dev path), else a deployed ndlaw instance over Streamable HTTP (`NDLAW_URL` + `NDLAW_AUTH` env, or `--url`/`--auth`; for Claude Code users the URL and Basic-Auth header are in their MCP config — `claude mcp get ndlaw`). The script speaks to the server directly, so no opinion text passes through model context.
 ```bash
 $VENV_PYTHON ~/.claude/skills/jetredline/ndlaw_export.py \
   --opinion <opinion_md_path> \
   --refs-dir ~/refs \
   --meta-out <TMPDIR>/sources.json
 ```
-    Exit 2 means no backend was reachable (typical in Cowork, where neither the corpus DB nor `NDLAW_URL` exists). **Fall back to a scribe subagent riding the in-context MCP connection** — if ndcourts-mcp tools are available:
+    Exit 2 means no backend was reachable (typical in Cowork, where neither the corpus DB nor `NDLAW_URL` exists). **Fall back to a scribe subagent riding the in-context MCP connection** — if ndlaw tools are available:
 
     1. From the cite JSON, list the corpus-eligible authorities that still lack refs text: case entries (not `pin_cite`, not `is_repeat`) whose `normalized` is an ND neutral cite (`YYYY ND N`) or N.W.-family reporter cite, with `local_exists` false.
     2. Launch **one** subagent (Task tool, `model: haiku`, subagent_type `general-purpose`) with that list and these instructions: for each citation, call `lookup_opinion(<cite>)` and record `case_name`, `url`, `url_source`, `date_filed`, and `citations`; then page through `get_opinion_text(<cite>, offset=..., limit=50000)` until `has_more` is false and **Write the concatenated text verbatim** (do not summarize, reformat, or strip the frontmatter block) to the refs path: `~/refs/opin/ND/<year>/<year>ND<n>.md` for `YYYY ND N`; `~/refs/opin/NW2d/<vol>/<page>.md` for `V N.W.2d P` (analogously `NW3d`, `NW`). Finally Write `<TMPDIR>/sources.json` mapping **every** citation form from `citations` to `{"case_name", "url", "url_source", "date_filed", "via": "ndlaw"}`, and return only a tally ("N exported, M not in corpus") — never opinion text.
     3. The token cost stays inside the subagent's isolated context (and within the subscription); main context must never page opinion text through its own tool results just to embed it.
 
     If MCP tools are also unavailable, skip `--sources-meta` below and fall through silently (never stall on this step) — or, at most, build a minimal `sources.json` from one cheap `lookup_opinion` call per unique ND case (URL + case name only) if a stray connection allows it.
 
-    **11b. Generate the page.** Write a `via.json` mapping each citation (as written, or its normalized form) to the tier from the Pass 3B **Via** column — e.g. `{"2024 ND 156": "ndcourts-mcp", "445 U.S. 684": "CourtListener", "N.D.C.C. § 14-05-24": "local"}`:
+    **11b. Generate the page.** Write a `via.json` mapping each citation (as written, or its normalized form) to the tier from the Pass 3B **Via** column — e.g. `{"2024 ND 156": "ndlaw", "445 U.S. 684": "CourtListener", "N.D.C.C. § 14-05-24": "local"}`:
 ```bash
 $VENV_PYTHON ~/.claude/skills/jetredline/cite_review.py \
   --opinion <opinion_md_path> \
@@ -433,7 +433,7 @@ The script operates directly on the .docx ZIP archive — no unpack/pack pipelin
 
 **9c. Check the output.** Parse the JSON summary from apply_edits.py. If any edits failed, report them.
 
-**9d. Citation hyperlinks (optional).** When the user asks for clickable / linked citations, add `hyperlink` edits to the same `edits.json` — one per citation, anchored to the citation text, using the verified URL (prefer ndcourts-mcp `absolute_url`, else the CourtListener URL, else the `url` from `cite_check.py`):
+**9d. Citation hyperlinks (optional).** When the user asks for clickable / linked citations, add `hyperlink` edits to the same `edits.json` — one per citation, anchored to the citation text, using the verified URL (prefer ndlaw `absolute_url`, else the CourtListener URL, else the `url` from `cite_check.py`):
 ```json
 { "type": "hyperlink", "para": 12, "anchor": "2024 ND 45", "url": "https://www.ndcourts.gov/..." }
 ```
@@ -459,25 +459,25 @@ For each delegated pass performed inline (1, 3B, 4, 6, 7 — and 2's entry forma
 
 **Context limits:** Without subagents, very long documents (50+ paragraphs) may approach context limits. If this occurs, prioritize: Pass 2 → Pass 5 → Pass 3A → Pass 1 → Pass 3B → Pass 4 → Pass 6 → Pass 7. Inform the user which passes were completed.
 
-## Legal-Research MCP Servers (ndcourts-mcp, CourtListener)
+## Legal-Research MCP Servers (ndlaw, CourtListener)
 
 Two optional MCP servers improve citation verification when connected. **They are augmentation, not replacement** — every check below degrades gracefully to the existing pipeline (`cite_check.py` + `~/refs/` + `WebFetch` + web search). Never fail or stall a task because an MCP server is absent or returns no data.
 
-**Availability:** Before relying on a server, check whether its tools are present in your tool set (e.g. an `ndcourts-mcp` tool such as `verify_citation`, or a CourtListener tool such as `verify_citations`). If a tool is not available, skip silently to the next tier.
+**Availability:** Before relying on a server, check whether its tools are present in your tool set (e.g. an `ndlaw` tool such as `verify_citation`, or a CourtListener tool such as `verify_citations`). If a tool is not available, skip silently to the next tier.
 
 **Source precedence — apply at every case-citation check; fall through on a miss:**
-1. **ndcourts-mcp** (primary, North Dakota cases) — local ND opinion corpus. Deterministic, no network.
-2. **CourtListener MCP** (secondary) — case data ndcourts-mcp lacks: federal and out-of-state authorities, and ND opinions missing from the ND corpus.
+1. **ndlaw** (primary, North Dakota cases) — local ND opinion corpus. Deterministic, no network.
+2. **CourtListener MCP** (secondary) — case data ndlaw lacks: federal and out-of-state authorities, and ND opinions missing from the ND corpus.
 3. **Existing pipeline** (fallback) — `cite_check.py` / `~/refs/` / `WebFetch` / web search. Always available in CLI mode; the only path when no MCP server is connected.
 
 **Out of scope for both servers:** authoritative text of statutes, court rules, the constitution, and NDAC — these always resolve through the existing pipeline.
 
-**ndcourts-mcp tools and the fields used below:**
+**ndlaw tools and the fields used below:**
 - `verify_citation(query, expected_case_name=...)` → `found`, `canonical_case_name`, `formatted` (full Redbook cite), `cites_redbook`, `absolute_url`; when `expected_case_name` is supplied, also `name_matches` (bool) and `name_similarity` (0–1). Catches wrong volume/page/year and name drift.
 - `verify_quotation(citation, quote)` → `verbatim` (bool), `paragraph` (pinpoint ¶, generally 1997+), `differences` (word-level diff), `closest_text`. Typography-tolerant (curly/straight quotes, dashes, whitespace).
 - `detect_overruled_in_draft(draft_text)` → `flagged` (cited cases with possible negative/distinguished treatment, each with `treatment_entries` + citing context), `clear`, `unresolved`. **Advisory only.**
 
-**Cardinal caution:** ndcourts-mcp is a research aid, not an authoritative text; treatment signals and the `antecedent_name` heuristic are best-effort. Use them to *flag for human review*, never to auto-edit on the signal alone.
+**Cardinal caution:** ndlaw is a research aid, not an authoritative text; treatment signals and the `antecedent_name` heuristic are best-effort. Use them to *flag for human review*, never to auto-edit on the signal alone.
 
 ## Editing Instructions
 
@@ -530,7 +530,7 @@ When the opinion exceeds 30 paragraphs, delegate Pass 2 to a Task subagent (suba
 
 ### Pass 3: Citation Check (Delegated to Subagent)
 
-Pass 3 has three parts: (A) Bluebook format checking, in main context; (B) substantive citation verification, delegated to a subagent; and (C) a negative-treatment / overruling scan, in main context. Parts B and C use the ndcourts-mcp / CourtListener servers as the primary source when available — see "Legal-Research MCP Servers" above — and fall back to local files plus web verification otherwise.
+Pass 3 has three parts: (A) Bluebook format checking, in main context; (B) substantive citation verification, delegated to a subagent; and (C) a negative-treatment / overruling scan, in main context. Parts B and C use the ndlaw / CourtListener servers as the primary source when available — see "Legal-Research MCP Servers" above — and fall back to local files plus web verification otherwise.
 
 #### Part A: Format Check (Main Context)
 
@@ -556,10 +556,10 @@ Pass 3B verifies ALL North Dakota citations — cases, statutes, constitution, c
 
 > Read `~/.claude/skills/jetredline/references/pass-instructions/pass3b-citations.md` and follow it. The opinion file is at `[opinion_path]`; write the passages ledger to `<TMPDIR>/passages.json`. Here is the extracted citation list: [numbered list — ¶, citation text, proposition, quoted text if any, signal].
 
-**Returns:** the results table (`¶ | Citation | Type | Caption Check | Quote Check | Supports? | Via | Source Link | Notes`), a summary line with counts by type, the lookup-methods tally for case citations (`ndcourts-mcp / CourtListener / local / web / not found`), and an ND web-fallback note. The subagent also writes `<TMPDIR>/passages.json` (the passages ledger embedded in the citation-review HTML). Carry the tally and note into the analysis document alongside the results table. The instructions enforce the closed-loop case-name rule (correct same-cite typos only; never harmonize different citations) — Step 8a depends on the resulting Caption Check column and `name_similarity` values.
+**Returns:** the results table (`¶ | Citation | Type | Caption Check | Quote Check | Supports? | Via | Source Link | Notes`), a summary line with counts by type, the lookup-methods tally for case citations (`ndlaw / CourtListener / local / web / not found`), and an ND web-fallback note. The subagent also writes `<TMPDIR>/passages.json` (the passages ledger embedded in the citation-review HTML). Carry the tally and note into the analysis document alongside the results table. The instructions enforce the closed-loop case-name rule (correct same-cite typos only; never harmonize different citations) — Step 8a depends on the resulting Caption Check column and `name_similarity` values.
 
 **Web mode:** No shell, Python, or `~/refs/` — but the MCP servers, if connected, are the primary way to verify case citations here.
-1. **If ndcourts-mcp / CourtListener tools are available, use them first**, following Step 1.5 of `references/pass-instructions/pass3b-citations.md` (read it from project knowledge — verify_citation for caption/existence/name drift, verify_quotation for quotes); run Part C's `detect_overruled_in_draft` too. This closes the gap where case citations otherwise can't be verified without a filesystem.
+1. **If ndlaw / CourtListener tools are available, use them first**, following Step 1.5 of `references/pass-instructions/pass3b-citations.md` (read it from project knowledge — verify_citation for caption/existence/name drift, verify_quotation for quotes); run Part C's `detect_overruled_in_draft` too. This closes the gap where case citations otherwise can't be verified without a filesystem.
 2. For ND citations the MCP servers don't cover, use WebFetch on the official URL (build it from the citation, or run `cite_check.py` if any shell is available).
 3. For ND case citations, use web search to locate the opinion on ndcourts.gov or Google Scholar.
 4. Verify quotes and substantive support as described above.
@@ -570,15 +570,15 @@ Pass 3B verifies ALL North Dakota citations — cases, statutes, constitution, c
 
 A proofreading pass that flags cited cases later opinions may have overruled, superseded, abrogated, or distinguished. jetredline has no other "still good law?" check, so this is purely additive.
 
-**If the `detect_overruled_in_draft` tool is available** (ndcourts-mcp), call it once on the full draft text:
+**If the `detect_overruled_in_draft` tool is available** (ndlaw), call it once on the full draft text:
 - Pass the draft opinion/memo text as `draft_text`.
 - From the result, take `flagged` (each entry has the cited case, a treatment signal, and `treatment_entries` with the citing opinion and sentence-local context). `clear` and `unresolved` are informational.
 - For each flagged case, queue a **comment** (handled in Step 8b) on each occurrence in the draft, quoting the citing context and the citing opinion. **Never** convert a treatment signal into a tracked-change edit.
 - In the analysis document, add a short "Negative-treatment check" subsection: list flagged cases with their signal and citing opinion, and note that `unresolved` cites (federal, out-of-state, or not in the ND corpus) were **not** checked.
 
 **Caution:** Signals are sentence-local heuristics, not a verdict — a citing sentence may use a treatment word about a *different* case. Always present them as "possible negative treatment — verify," and read the cited opinion before relying on it. An absent flag is not assurance a case is good law.
 
-**If the tool is unavailable:** skip this part and note in the analysis document: "Automated negative-treatment check not run (ndcourts-mcp unavailable); citations were not screened for subsequent history." Do not attempt to substitute web search for a citator here.
+**If the tool is unavailable:** skip this part and note in the analysis document: "Automated negative-treatment check not run (ndlaw unavailable); citations were not screened for subsequent history." Do not attempt to substitute web search for a citator here.
 
 ### Pass 4: Fact Check (Delegated to Subagent)
 
@@ -671,7 +671,7 @@ For memos: Does the memo correctly identify the standard *and* apply it consiste
 
 In multi-issue documents where each issue has a different standard, verify each standard is applied to the correct issue.
 
-#### Statutory-Construction Cross-Check (when ndcourts-mcp is available)
+#### Statutory-Construction Cross-Check (when ndlaw is available)
 
 When the draft construes an N.D.C.C. section or a court rule as a point of decision, call `find_opinions_construing(<authority>)` and scan the returned `results[].opinions` for ND opinions construing the same provision. Use it as an advisory aid: if a recent or obviously on-point construction is missing from the draft's analysis, note it for the author's consideration (do not assert it is controlling). Skip silently if the tool is unavailable. For a key *cited* precedent whose holding the draft leans on, `case_summary(<cite>)` (disposition + `syllabus_points`) can confirm what the case actually held — this aids the substantive-support assessment; it does **not** feed the draft's own Case Highlight, since the draft is not in the corpus.