incitez

incitez ("insights on citations, Zig") is a legal-citation extraction engine in Zig: a drop-in-grade alternative to eyecite (Free Law Project, BSD-2-Clause) — the tool that powers CourtListener and Harvard's Caselaw Access Project — built to meet and surpass it.

It extracts full case citations, short forms, supra, id/ibid, journal, law, section, and reference citations, with party names, pin cites, courts, years, parentheticals, and short→full resolution. Pure in-memory Zig core (no I/O) behind a C FFI, with a C CLI and a browser WebAssembly build.

▶ Try it live: incitez-web.pages.dev — load a sample brief (or paste your own) and watch citations extract entirely in your browser via WebAssembly. Nothing leaves your machine.

How it compares to eyecite

The whole design is honest about the bar: equivalent where eyecite is the authority, faster everywhere, and more correct where eyecite has real-world weaknesses.

Parity — the trust anchor (verified, not claimed)

incitez agrees with the pinned eyecite oracle 215/215, 0 divergences across eyecite's own ported test corpus, re-checked on every push by a live differential gate (scripts/differential, CI checks.differential). On eyecite's corpus eyecite is the ground truth, so this proves incitez is not worse — the credibility floor that makes the claims below believable. It is not by itself a superiority claim.

Speed — measured superior

Same work, warm, engine-processing-only (no Python interpreter-import artifact), recomputed live — never hardcoded:

input	eyecite (Python)	incitez (Zig)	faster
real 90-page appellate brief (121 KB text)	~970 ms	~31 ms	~31×
benchmark corpus, steady-state (116 KB)	2561 ms	47 ms	54.8×
cold-start single invocation	0.432 s	0.0069 s	62.7×

(Apple M4 Max; see BENCHMARKS.md, hardware-keyed + two-sided regression-gated. Run it yourself: ./demo or ./bm.) Cite counts match or slightly exceed eyecite on real input (e.g. 2160 vs 2159 — incitez recovers one eyecite's finditer drops; 324 = 324 on the brief).

Accuracy — verified-superior in specific, real-world classes

Where eyecite is demonstrably wrong (confirmed against the live oracle), incitez is right. These are objectively checkable case classes, additive to the 215/215 parity (the marker-free corpus never triggers them). Full catalog + methodology in docs/exceeds_eyecite.md:

• En-dash / em-dash pin ranges — eyecite drops 241–242 (non-ASCII dash); incitez recovers it.
• Structural-newline boundaries — a section heading on the line above a citation bleeds into eyecite's party name; incitez treats the boundary as a hard stop (TABLE OF AUTHORITIES\nAlbritton v. Gandy → Albritton, not TABLE OF AUTHORITIES … Albritton).
• Leading Bluebook signals — eyecite leaves Compare/Accord/Contra/ Consider glued to the plaintiff; incitez strips them, leading-position-only, so a signal word embedded in a real name survives.

These are documented case classes, each independently verifiable. A systematic precision/recall study over a labeled real-document corpus (the rigorous, quantified next step) is tracked, not yet complete — we don't claim a single accuracy multiple we can't yet defend.

Architecture

C CLI (cli/main.c, all I/O) ──► C FFI (include/incitez.h) ──► Zig core (src/, pure, no I/O)
                                                              └► WASM (src/wasm.zig) for the browser

• The C CLI dogfoods the FFI; sibling-Zig consumers may import the core directly.
• A VM-only WASM artifact (packages.wasm) powers the privilege-preserving browser demo (incitez_web) — citations extracted client-side, nothing leaves the browser. Consumer contract: docs/wasm_abi.md.
• Pipeline split: docscan owns structure-aware PDF→text (and emits structural boundary markers); incitez owns citation logic and the span contract; incitez_web composes them client-side in the browser.

Usage

./build                      # build via nix (sandboxed); → zig-out/bin/
./test                       # full suite: unit + acceptance + mutation + CLI + WASM smoke
./bm                         # benchmarks (per-function + vs-eyecite), regression-gated
./demo [brief.pdf]           # the incitez-vs-eyecite head-to-head (parity / speed / surpass)

zig-out/bin/incitez extract <file|-|@stdin> [--json]

Input is UTF-8 plain text; for real PDFs, extract via docscan (structure-aware) or run the eyecite recipe (incitez_clean: collapse \s+, strip __). See docs/wasm_abi.md §5.

Verification (MFIC-shaped)

1. eyecite's ported corpus = the acceptance suite.
2. Live differential gate vs pinned eyecite (CI).
3. Mutation suite — perturbed citations must degrade predictably.
4. Per-function microbenchmarks + a machine-independent scaling-ratio gate (tests/scaling) that fails any phase growing super-linearly.
5. std.testing.allocator proves leak-freedom.

License

Business Source License 1.1 — © 2026 Peter Marreck (d/b/a Mecha LLC). Source-available: non-production use plus limited production use (per the Additional Use Grant — direct CLI/internal use by orgs under 100 staff and $1M revenue) are permitted; everything else needs a commercial license. Converts to the MIT License on 2030-06-16. Incorporates BSD-2-Clause data/grammar from Free Law Project (reporters-db, courts-db, eyecite); those notices are retained in THIRD_PARTY_LICENSES.