lingo stands on a lot of shoulders. This ledger records every library we learned from, what we took, and its license — kept current as we borrow, so we can give proper props in the README and release notes. Ideas and API shapes carry no license obligation; we do not copy code without noting it here explicitly.
| library | license | what we took |
|---|---|---|
| libphonenumber-js | MIT | The existence proof: parse/format/validate for one messy human domain behind a tiny API. Metadata-driven size tiers; AsYouType's feed-text→rich-result incremental shape; its restraint (ignore junk, don't punish). |
| date-fns | MIT | Pure functions, tree-shakeable modules, locale packs as additive data (our i18n plan), formatDistance threshold discipline. |
| Humanizer (C#) | MIT | The two-way ethos: humanize ⇄ dehumanize as paired APIs; number-to-words as core competency. |
| inclusive-dates | MIT | Accessible natural-language date input as a product pattern: type "next friday", get an accessible confirmation; aria-live parse feedback; a11y-first component design. Our DOM layer is its spiritual generalization beyond dates. |
| library | license | what we took |
|---|---|---|
| convert-units | MIT | toBest() best-fit unit selection concept; measure/system organization of unit data. |
| js-quantities | MIT | Temperature absolute-vs-delta separation (tempC vs degC) — our convert/convertDelta split. |
| unitmath | Apache-2.0 | Formatting options design; custom unit definition ergonomics. |
| mathjs unit system | Apache-2.0 | Prefix handling and parsing grammar cautionary study (powerful but heavyweight — we deliberately stay non-algebraic). |
| mathjs expression trees + expression syntax | Apache-2.0 | Typed node AST + parse/evaluate split studied for plan 032 (input calculations); we'd take the tiny node union and canonical toString, reject scopes/functions/CAS. Deeper pass 2026-07-29 added two things: its percentage operators (100 + 3% → 103 as a dedicated operator distinct from modulus) as the model for our phase-1 percent-of family, and its security page as the cautionary framing — an expression evaluator needs one only because it evaluates arbitrary code, so our closed node union's inability to express a call or a side effect is a guarantee to advertise at the LLM tool boundary, not just an internal constraint. Shipped 2026-08-22 as @pascal-app/lingo/calc (D73). |
| Pint (Python) | BSD-3-Clause | delta_degC / delta_degF as prior art for making affine-delta semantics explicit rather than warning users off the operation (mathjs's own advice is "avoid calculations using celsius and fahrenheit"). Motivated plan 032's AFFINE_DELTA_ASSUMED warning: lingo's compound path already delta-converts correctly, it just never said so. |
| ECMA-402 / Intl.NumberFormat unit style | spec | Sanctioned unit identifier list; free locale-aware formatting we lean on instead of shipping CLDR. |
| library | license | what we took |
|---|---|---|
| chrono-node | MIT | Known-vs-implied component certainty model; reference-date (now) discipline; parser+refiner pipeline concept (we hand-roll smaller). |
| moment / dayjs relativeTime | MIT | The classic relative-time threshold table (ss:44, m:45, h:22, d:26, M:11) our reversible humanizer descends from. |
| date-fns formatDistanceStrict | MIT | Qualifier-free strict humanize as our default model; roundingMethod; DST-normalized minutes before day bucketing. |
| Luxon | MIT | toRelative vs toRelativeCalendar split (elapsed-time vs calendar-compare units) — how "yesterday" stays correct across DST. |
| timeago.js | MIT | ≤10 s "just now" special case; two-sided past/future phrasing. |
| humanize-duration | Unlicense | Duration decomposition + largest units option; humanizer() factory over global config. |
| ms | MIT | Short-form duration lexicon (2.5 hrs, 1y); cautionary: silent undefined, month-less model. |
| tinyduration | MIT | ISO-8601 duration parse/serialize with a tiny surface, plain objects, typed error, no normalization. |
| Intl.RelativeTimeFormat | spec | Baseline phrasing conventions for relative time; numeric: 'auto' calendar words. |
| library | license | what we took |
|---|---|---|
| Maskito (@maskito/core) | Apache-2.0 | The size bar (~4 kB gz, zero deps); new X(el, opts) + single destroy(); pure transform function for programmatic values; adapters as ref-returning hooks. |
| imask | MIT | updateOptions live reconfig; value/unmaskedValue/typedValue triad → our raw/value/quantity; accept-vs-complete event split → parse-vs-commit. |
| cleave-zen | MIT | DOM-decoupled pure formatter core with optional DOM layer (their own deprecation lesson from cleave.js). |
| react-aria NumberField + @internationalized/number | Apache-2.0 | isValidPartialNumber → our isValidPartialQuantity tri-state; commit-on-blur model; validationBehavior: 'native'|'aria'; OS-aware inputmode; deriving locale symbols from formatToParts instead of shipping data. |
| react-number-format | MIT | Three-representation callback payload + isAllowed veto hook. |
| Base UI Field | MIT | Field-state data-attribute vocabulary (data-invalid/touched/dirty/filled), validationMode + validationDebounceTime options, ValidityState-keyed error matching. |
| Radix Primitives / Headless UI | MIT | Zero shipped CSS; one data-* attribute per state (not combined strings); namespaced CSS custom properties. |
| downshift | MIT | Prop-getter pattern for the React hook adapter. |
| flatpickr | MIT | altInput pattern: pretty text visible, hidden canonical input submits — our name option. |
| Excel / Google Sheets / Notion / Airtable formula entry (studied 2026-07-29) | convention | The leading-= mode switch as the answer to plan 032's hardest open question. One text box serves two grammars: bare input keeps its literal meaning, = opts into arithmetic. Lets lingo add expressions where - currently means range without changing a single existing reading — the feature becomes purely additive instead of a corpus-breaking reinterpretation. |
| Soulver · Numi · Raycast/Spotlight calculators (studied 2026-07-29) | commercial, studied not copied | The "calculator in a text field" UX bar for plan 032: unit-aware natural-language lines ($50 with 20% off, 5 hours in minutes) resolved inline with the result shown beside the input rather than replacing it. Confirms our choice to surface evaluated expressions through ./complete as a ranked completion instead of silently committing a computed value into the field. |
| WAI / W3C forms tutorials & WAI-ARIA APG | W3C | Polite-while-typing vs alert-on-commit announcement split; debounce guidance; error summary + focus management. |
MDN Constraint Validation, :user-invalid |
CC / spec | Post-interaction error timing semantics; setCustomValidity interop. |
| reference | what we took |
|---|---|
| Velvet UI (24h.studio) | Named motion scale + ease tokens; tinted demo stages with meta below; error choreography (decaying shake, 12%-alpha ring, reserved message space); aphoristic captions; achromatic OKLCH discipline. |
| jakub.kr — "Details that make interfaces feel better" | The pre-ship detail checklist: tabular-nums on live outputs, concentric radii, blur-morph icon swaps, interruptible transitions, staggered enters/blur exits, 3-layer shadow recipe with dark-ring collapse. (Also cited by Pascal's own design language.) |
| muload.dev (Murat Alpay) | Rail TOC with overlapping active border; tiny-h2 + hairline + air section rhythm; command-bar hero; ring-shadow code containers; variant walls captioned by prop diff; ghost-width tabs. |
| BE UI | Demo-before-prose page anatomy; collapsing code bodies; input state choreography vocabulary (gliding caret, shake, check draw). Tabs (components/motion/tabs.tsx): one shared layoutId per instance driving a spring pill/underline indicator (stiffness 170, damping 24, mass 1.2), inactive panels kept mounted as hidden (no AnimatePresence, enter-only fade+4px rise), named easing/spring token file (lib/ease.ts) pattern. Range Slider (components/motion/range-slider.tsx): plain pointer events (no Framer drag) with instant-commit value decoupled from a spring-smoothed visual position (SPRING_GLIDE 700/50/0.5 for glide, SPRING_BOUNCY 500/14/0.7 for an underdamped grab-scale bounce), single vertical-bar thumb with self-offsetting x transform to stay inside the track at both ends. |
| shadcn/ui docs code blocks (dissected 2026-07-03) | ComponentSource/CodeCollapsibleWrapper/CodeBlockCommand anatomy: theme-follows (not forced-dark) code surface one step off page bg; radius-xl panel with no border; npm↔pnpm/yarn/bun command-transform lookup table; PM-tab overflow layout. Confirms our own dual-theme codeToHtml + defaultColor:false approach is already sound — no rehype-pretty-code or jotai needed. |
| shadcn/ui ComponentPreviewTabs (inspected 2026-07-06) | Preview-first component docs rhythm: keep the live demo as the default surface and move code/source or raw payloads behind a secondary tab/toggle. Adapted to lingo's local Tabs and DemoFrame instead of importing upstream dependencies. |
| shadcn/ui docs app implementation (inspected 2026-07-04) | Actual docs shell anatomy: Fumadocs page tree feeding a grouped DocsSidebar, H1 row with split DocsCopyPage action and previous/next buttons, right-side DocsTableOfContents, and a command dialog that combines static page groups with indexed search results. We adapted the patterns route-stably: local docs catalog, grouped anchor nav, //Cmd+K search, split copy/view-markdown controls, and query-scoped markdown routes. |
| shadcn/ui v4 navigation and theme source (inspected 2026-07-05) | site-header, main-nav, mobile-nav, mode-switcher, and theme-provider patterns: sticky header, desktop ghost-button nav, mobile nav organized into primary/resources/sections, header search always reachable, System/Light/Dark theme choices, and system color-scheme propagation. We adapted those directly while removing our tab/docs-nav Motion layout transitions for a snappier docs feel. |
| shadcn/ui Base Dialog + Base Table (inspected 2026-07-05) | Base UI-backed Dialog anatomy (Portal/Overlay/Content/Title/Description/Close) and the semantic table slot contract. We adapted them manually so the site could keep its local Button styling, Tailwind v4 data-open/data-closed state classes, and scroll-contained docs tables. |
| shadcn/ui Base Field, Textarea, and Separator (inspected 2026-07-05) | Field composition (Field/FieldLabel/FieldDescription/FieldError/FieldGroup) plus shared Textarea/Separator primitives. We adapted them manually to avoid overwriting local Input/Button styling, then used them to normalize AI JSON and escalation controls. |
| shadcn/ui Base Dropdown Menu + Base UI Menu (inspected 2026-07-05) | Base Menu composition (Root/Trigger/Portal/Positioner/Popup) plus semantic Item/LinkItem roles and trigger expanded state. We adapted it manually with local Button styling, lucide icons, and Tailwind v4 state classes, then used it to replace the docs actions' bespoke absolute menu. |
| shadcn/ui Base Tabs + Base UI Tabs (inspected 2026-07-05) | Tablist/tab/tabpanel semantics for small option sets. We reused the site's already-owned Tabs primitive to replace the install command block's bespoke package-manager buttons while preserving local motion timing and package-manager persistence. |
| shadcn/ui Base Button (inspected 2026-07-05) | Base UI Button link guidance: anchors should use buttonVariants on plain links instead of Button render={<a />} because Base UI applies button semantics. We adopted that pattern for header and hero link CTAs while keeping the local button visual system. |
| Simple Icons (inspected 2026-07-06) | Integration picker brand SVG paths and default HEX colors for JavaScript, webcomponents.org, React, React Hook Form, TanStack, shadcn/ui, Next.js, Vue.js, Svelte, Node.js, and Anthropic. Kept inline to avoid a runtime dependency or runtime CDN request. |
| WAI-ARIA APG Combobox Pattern (inspected 2026-07-05) | Editable combobox + listbox popup focus contract: keep DOM focus on the input, point aria-activedescendant at the active option, connect the popup with aria-controls, and mark the visually active option with aria-selected. Applied to the docs search dialog while keeping local Dialog/Input/Button primitives. |
| coss.com/ui | Dense component-gallery pacing: compact section nav, practical demos before prose, and fewer ornamental status chips. |
| Motion Primitives Border Trail | Subtle edge motion as an affordance around interactive demo surfaces; adapted as a dependency-free neutral top-edge trail that disables under reduced motion. |
| Motion Primitives (ibelick, MIT — dissected 2026-07-04) | text-effect.tsx: word/char/line splitting with a non-animated per-word wrapper for char mode, preset variant shapes (blur/fade-in-blur/scale/fade/slide), and the sr-only-plus-aria-hidden pairing as our house pattern for any split-text animation. text-scramble.tsx: cautionary reference (linear reveal-front sweep, not per-glyph settle) and an explicit gap to fix, not copy — it has zero accessibility accommodation for its live-updating text node. text-shimmer.tsx: the two-layer background-clip:text gradient + animated backgroundPosition shimmer technique, spread-scales-with-content-length idea (credited by them to JohnPhamous), and its AI-loading demo framing as precedent for our own async/parsing-in-progress affordance. All three re-timed onto our own --motion-*/--ease-out tokens and given reduced-motion fallbacks upstream doesn't have; no motion/clsx/tailwind-merge deps pulled in. |
| llms.txt proposal | One-file agent orientation: H1 package identity, short summary, and a concise Markdown link section for deeper docs. |
| Svelte llms tiers | Size-tiered agent docs: index → small/medium/full bundles plus per-package files. Adapted as /llms.txt (index), /llms-small.txt (npm reference), /llms-full.txt (complete narrative). |
Mintlify append-.md |
Per-page markdown endpoints and Visibility for human-only vs agent-only content. Adapted as /docs/<section>.md slices with self-contained context headers. |
AI structured-output & Standard Schema domain (research pass 2026-07-04, wiki/research/ai-structured-output.md)
| source | license/type | what we took |
|---|---|---|
| Standard Schema / @standard-schema/spec | MIT, types-only spec | The ~standard.validate Input≠Output transform contract as the correct shape for a lingoField() — coercion belongs inside validate, not bolted on; confirmed the package is pure types (zero runtime bytes), so depending on it costs nothing against our zero-dep budget. |
Vercel AI SDK (ai@7.0.14) |
Apache-2.0 | experimental_repairText's local-repair-first design (return null to fall through, no forced LLM round-trip) as the model for a deterministic lingo repair function; the FlexibleSchema/asSchema pattern (accept Zod ∪ Standard Schema ∪ plain JSON Schema behind one union) as prior art for how lingo could expose fields that compose into someone else's schema; experimental_refineToolInput's post-validate/pre-execute placement as a second, narrower interception point. |
| instructor-js | MIT | Cautionary contrast, not borrowed: its retry-on-validation-failure loop always costs an LLM round-trip. Sharpened our positioning — deterministic local canonicalization (parse "2 lbs" without asking the model again) is the gap it leaves open. |
| Anthropic computer-use tool docs | spec/docs | The type action's documented failure mode ("Claude sometimes assumes outcomes... take a screenshot and evaluate") as the concrete justification for lingo's DOM layer's untrusted-event handling and blur-time canonicalization — a problem statement, not code. |
LLM/agent formatting-failure evidence domain (research pass 2026-07-04, wiki/research/llm-formatting-failures.md)
| source | license/type | what we took |
|---|---|---|
| OpenAI — Structured Outputs docs & launch post | docs/spec | The primary-source admission, in the vendor's own words, that anchors the whole positioning: Structured Outputs guarantees schema shape but "the model may still make mistakes within the values of the JSON object" — the exact syntax/semantics line lingo sits on. |
| Berkeley Function-Calling Leaderboard / Gorilla and Gorilla paper | research/blog | The "5% → 5, not 0.05" implicit-conversion example — the crispest, primary-source illustration of a value-format error a schema can't catch. |
| ComplexFuncBench | arXiv paper | Its value_error (78.8% for Qwen2.5-72B) as the dominant category in a 5-way error taxonomy — quantified evidence that value correctness, not structural correctness, is function-calling's real bottleneck. |
| The Constraint Tax (Jaideep Ray) | arXiv preprint | The clearest quantified validity/correctness tradeoff: schema validity 61.5%→100%, answer accuracy 19.7%→11%, wrong-valid-schema outputs up to 88.9% under hard constraints (sub-3B models) — used to argue constraints alone can increase silently-wrong output. |
| Outlines issue #1178 | GitHub issue (Apache-2.0 project) | A reproducible counterexample: a grammar-constrained "float format" decoder still emits tokens that fail float() conversion — proof that even working constrained decoding doesn't close the semantic gap. |
| JSONSchemaBench | arXiv paper | Compliance-rate numbers (13% unconstrained → 41% best constrained-decoding framework on hard schemas) as evidence that failure rates are schema-complexity-dependent, framing why a canonicalization layer's value doesn't show up on toy schemas. |
| Tool-Reflection-Bench | arXiv paper | Repair@1 ≤ 9.6% baseline figure — the retry-economics argument that naive re-ask-on-failure is a low-yield fix, sharpening the case for fixing formatting before validation instead of after. |
| WebVoyager, OSWorld, FormFactory, UI-CUBE | arXiv papers | Confirmed that published browser/computer-use agent failure taxonomies are dominated by grounding/navigation, and that FormFactory (the benchmark closest to testing form-filling formatting) explicitly admits its own metric can't yet detect wrong-format field entry — the "instrumentation gap, not an absence of the problem" framing for lingo's computer-use pitch. |
| apple-health-data-bridge issues #14 / PR #13 | GitHub project | Concrete, if small-scale, real-world incidents of LLM unit hallucination (mg/dL vs mmol/L, both valid-looking) and calendar-invalid date extraction — used as illustrative, explicitly anecdotal-tier examples, not load-bearing statistics. |
| source | license/type | what we took |
|---|---|---|
Standard JSON Schema / @standard-schema/spec@1.1.0 |
MIT, types-only | The ratified StandardJSONSchemaV1 sibling interface (~standard.jsonSchema.input/output({target}), published 2025-12-15) — lingo implements it verbatim, so fields drop into AI SDK v6/v7 and LangChain with zero adapter. Now vendored as a types-only devDep behind a conformance test so drift from upstream fails CI. |
Vercel AI SDK @ai-sdk/provider-utils asSchema() |
Apache-2.0 | The vendor-dispatch contract we conform to (vendor!=='zod' → the SDK reads ~standard.jsonSchema.input({target:'draft-07'})); its experimental_repairToolCall shape as the model for repairToolCallWith (since experimental_repairText/generateObject are deprecated in AI SDK 6). |
| Zod v4 | MIT | .meta()→JSON-Schema passthrough and the toJSONSchema(schema,{target}) helper name (adopted as our toJSONSchema(field)); .overwrite() as the exact shape of our type-preserving "benign forgiveness"; Zod Mini's function-call tree-shaking (which our option-object fields already get for free). z.coerce.number()→NaN stays the cautionary wedge. Skipped deliberately: a fluent chain builder and .catch()-style silent fallback (contradicts "LLM tools safer"). |
Valibot + @valibot/to-json-schema · ArkType |
MIT | Two contrasting StandardJSONSchemaV1 shapes: Valibot's separate-package modularity (mirrors our separate ./ai entry) and ArkType's implements-both-halves-from-construction (mirrors our LingoField). |
Effect Schema transform({decode,encode}) |
MIT | Independent confirmation of our two-way guarantee (hard rule 4): its canonical teaching example is height↔string via paired decode/encode — the same law, arrived at separately. |
@hookform/resolvers standardSchemaResolver |
MIT | The "one resolver adapter per ecosystem" pattern and the transform-on-submit contract (returns the parsed result.value unless {raw:true}) that lets a lingoObject canonicalize a whole react-hook-form on submit. |
TanStack Form · shadcn/ui Field FieldError |
MIT | Native vendor-agnostic Standard-Schema validators and Standard-Schema-agnostic error rendering — lingo fields plug into both unmodified; TanStack's documented "onSubmit gets raw input" limitation is what our useLingoInput bridge closes. |
| Braintrust autoevals · promptfoo | MIT | The {pass,score,reason}/Score grader shape our quantityMatch/dateMatch duck-type into; autoevals' unit-blind NumericDiff (5000 g vs 5 kg ≈ 0) and string-vs-number JSONDiff Levenshtein fallback are the cautionary bug that motivates "canonicalize both sides before you diff." |
| Baymard Institute · GoodUI | research | The form-UX evidence base: fewer fields + single-input-entity integrity (Baymard); natural-language forms win for one value+unit but lose replacing whole forms (GoodUI's mixed A/B data, incl. a −22% test) — which scopes our claim honestly. |
| Google Home TemperatureSetting trait · Stripe minor units | docs | Prior art for canonical-unit storage with presentational display (Celsius-canonical; integer cents) — exactly our display:'canonical' vs 'echo' split, validated on trusted platforms. |
@openai/agents tool guardrails · LangGraph interrupt() · Stagehand |
MIT/Apache-2.0 | Interception points for agent workflows: defineToolInputGuardrail and interrupt() as homes for lingo's did-you-mean candidate; Stagehand's Zod-only extract() as the cautionary case motivating the extract-then-canonicalizeValues recipe. |
| MDN ElementInternals / web.dev — form-associated custom elements | docs/spec | formAssociated + setFormValue/setValidity as the way <lingo-input> becomes a real form control in one node (collapsing the visible+hidden two-node trick) for framework-agnostic use. |
| source | license/type | what we took |
|---|---|---|
| Pascal home-services & private-editor repos (same owner, internal) | internal | The agent-first knowledge layer this repo now runs on: AGENTS.md-as-canonical with CLAUDE.md = @AGENTS.md; the plans↔wiki lifecycle ("plans = where it's going, wiki = how it works today; ship → promote the durable explanation, keep the plan as history") with explicit read-before/update-when triggers; goal + success_criteria -> proof artifact plan frontmatter with [MET]/[SKIPPED] annotations; the decisions offer-gate (hard to reverse / surprising / real trade-off — all three); CONTEXT.md glossary with Avoid synonym lists; plans/backlog.md parking lot ("append, don't act"); behavioral-guardrails block; doc-rules compiled into deterministic gate scripts (our check-zero-deps.mjs); playbooks ending in numbered PR checklists; "don't invent nits" review closing rule. |
| Pascal home-services resource/API design playbooks (same owner, internal) | internal | The resource-style output contract behind describeResult(): stable object names, grouped related fields, outsider-readable primitives, composition over flattened blobs, and the decision not to copy persistence fields such as id/timestamps onto deterministic parser values. |
| shadcn/ui | MIT | The public-library monorepo shape: packages/* (published) + apps/* (docs site) with turbo at the root (bun as the package manager here), CONTRIBUTING at root, the site as a first-class app beside the library. |
| Matt Pocock's skills & ADR-FORMAT (via the repos above, which credit him) | — | The one-paragraph-ADR minimalism and the offer-gate framing that our decisions.md rubric descends from. |
| agents.md convention | open convention | AGENTS.md as the single cross-tool agent entrypoint. |
- AutoNumeric (MIT) — feature completeness vs 43 kB gz: the monolith we refuse to be.
- cleave.js (Apache-2.0, deprecated) — hard DOM binding as original sin.
- Luxon (MIT) — API elegance reference; too heavy for our budget.
- Quill CJK breakage (chrome 65 / quill#2009) — why we never mutate during IME composition.
- react-scan (Aiden Bai, MIT) — dev-only render highlighting + FPS toolbar wired
into the site via
src/components/dev/dev-tools.tsx; the NODE_ENV-gated dynamic-import pattern keeps it at zero prod bytes. - react-grab (Aiden Bai, MIT) — ⌘C-over-element copies component stack + source locations for agents; dev-only alongside react-scan. Pairs with our own agent workflow on this repo.
- react-doctor (Million, MIT) — static React audit (state/effects, perf, a11y,
security); adopted as an advisory PR workflow (
.github/workflows/react-doctor.yml). - taste-skill (Leonxlnx, tasteskill.dev) — anti-slop frontend skill; its
redesign protocol (audit-first, targeted-evolution levers, IA preservation
rules) governs site Pass 10; installed via
npx skills add.
| source | license/type | what we took |
|---|---|---|
| Base UI Combobox (v1.6.0) | MIT | The ARIA combobox contract (role="combobox" + aria-autocomplete="list" + aria-activedescendant pointing at highlighted option) as the accessibility pattern for lingo's DOM completions popup — keeps DOM focus on the input while screen readers navigate the list. Applied to the planned completions ARIA wiring in packages/lingo/src/dom/controller.ts. |
| FFmpeg FATE + plan 010 | LGPL-2.1 (the practice pattern, not code) | Seeded property round-trip testing: encode-then-decode oracle (FATE's enc_dec_pcm pattern) as the direct model for lingo's two-way guarantee property test — generate random magnitudes, format(), parse(), assert base equality within epsilon. Validates the plan 010 layer-2 spec from independent prior art. |
| es-toolkit | MIT | Per-entry size/benchmark publishing as a documentation practice: each function measured individually with esbuild min+gzip, results surfaced on the site with head-to-head comparisons. Lingo already has the measurement infrastructure (scripts/size.mjs, scripts/bench.mjs); the idea is to pipe those numbers into the published docs site as a future docs-site item. |
| source | license/type | what we took |
|---|---|---|
| CLDR RBNF (Rule-Based Number Format XML) | Unicode License (ICU), MIT-compatible | Canonical number-word spellout rules for 100+ locales as the authoritative source for generating exhaustive numberWords.ones/tens/scales entries per locale pack. Specifically: French vigesimal composition rules (soixante-dix through quatre-vingt-dix-neuf), CJK scale multiplication patterns, Spanish gender-inflected compound hundreds (doscientos/doscientas). Adopted as the single source of truth for number-word table generation rather than manual curation. |
| CLDR dateFields JSON | Unicode License, MIT-compatible | Structured per-locale relative-date vocabulary (dayOffset -2 to +2, relative time pattern templates, calendar period labels) directly mappable to lingo's dayOffsets, calendarPeriodPhrases, and relative pack fields. Fills missing entries like avant-hier (fr), pasado manana (es), 一昨日/明後日 (ja), 前天/后天 (zh) from a machine-parseable canonical source instead of manual inventory. |
| facebook/duckling Corpus.hs per-language corpus files | BSD-3-Clause (permissive, MIT-compatible for derived test data with copyright notice) | Gold-standard (input_text, expected_value) test pairs across 49 languages x 15 dimensions as validation sets for lingo's per-locale corpus gates. Priority targets: ZH/JA Numeral (validates scale composition), FR Time (validates vigesimal + spoken clock), EN/ZH Quantity. A script will mine Corpus.hs into TypeScript test-row format. Not embedding Duckling code — only the test expectations as validation oracles. |
| Microsoft/Recognizers-Text AmbiguityFiltersDict pattern | MIT | The false-positive exclusion mechanism (a per-locale dictionary of strings where numeral characters appear in non-numeric words: '十足', '大陆', '放肆' in Chinese) as the model for a planned excludePatterns pack field. Prevents CJK numerals from firing on idiomatic compounds that happen to contain numeral kanji. Adopted as a design pattern, not as embedded data. |