Decision log

Objections to the build prompt, corrections to seed data, and dependency justifications. Each entry states the reasoning.

Objections to the prompt / spec

"-th" as a productive English state suffix. The prompt's seed list includes -th (warmth, truth). It is dead (productivity: "dead"); we include it but mark it dead and exclude it from productive-affix explanations. [ESTABLISHED]
Telugu -ఐన listed as an affix. అయిన/-ఐన is the adjectival participle of అగు "become" used as a copular linker, arguably syntax rather than derivational morphology. We keep it as kind: "suffix", function ADJ, but tag it ENGINEERING with a note; treating it as a suffix is an implementation convenience for alignment with English -ic/-al and Hindi -ईय.
Hindi -ई polysemy. -ई forms abstract nouns (गरम → गरमी), feminines, and adjectives (देस → देसी). One affix row per function would fragment the atlas; instead one affix carries multiple functions and the entry names which function applies. This is why LexicalEntry.morphology.affixes[] pairs affix_id with function_id.
eq:ADV including Hindi "से". से is a postposition, not an affix. We list only -पूर्वक as the Hindi member and note in the class that the commonest adverbial strategy in Hindi is the phrase X से, which is outside affix morphology. [ESTABLISHED]
Hindi -आना as causative. The productive Hindi causative is stem-internal -आ-/-वा- (verbal inflection-like derivation: बनना → बनाना → बनवाना). We model -आ- as the causative affix (register N), noting that pairing it with English -ize/-ify is functional, not distributional. The prompt's "-आना" spelling conflates the causative increment with the infinitive -ना; we store the affix as -आ- with allomorph -वा-.
Telugu -వాడు/-ది, -గాడు as derivational affixes. These are pronominalized formatives (human masc/non-masc). Kept, tagged ENGINEERING, because they do the functional work of agentive/possessor nominalization that the Atlas aligns; a syntactician would call them bound pronouns.
"dog" contested chain. Hindi कुत्ता / Telugu కుక్క resemblance: CDIAL treats the Indo-Aryan word as of uncertain, possibly expressive or non-Aryan origin; DEDR lists a Dravidian kukka set. We cite both works without entry numbers (per the no-invented-numbers rule; numbers to be added at review time from the physical works), store the Telugu–Hindi relation as status: "contested", confidence 0.4, resolving nothing. [ESTABLISHED that it is contested]
English "king" and Sanskrit "rājan". Unrelated (kuningaz vs. h₃rḗǵs); note that English royal/regal (via Latin/French) IS cognate with rājan — we include that as a bonus COGNATE edge, properly sourced.
"name" PIE form. h₁nómn̥ (also cited h₁néh₃mn̥) — reconstruction details are debated; we cite the mainstream form with confidence 0.8, RECONSTRUCTED.
Atom cap tension. 120–160 concepts strain 50 atoms. Resolved by allowing concepts to reuse coarse structures (e.g. LONELINESS = STATE_OF(FEEL(BAD, because NOT NEAR PEOPLE)) approximated as STATE_OF + FEEL + BAD + ALONE-via-NEG(NEAR)); some concepts carry HYPOTHESIS structures where decomposition is genuinely coarse.
French/Tamil enum values. In v0.1 the Lang literal contained only en/te/hi; adding unused enum members invites accidental data. Superseded by 13, which added both languages together with their data.

Corrections to seed lists

Telugu -త్వం and -వంతుడు marked register S (Sanskritic -tva, -vant); -తనం, -రికం, -డం/-ట, -గా, -గల, -లేని, -వాడు marked N. -కారుడు is Sanskritic base (kāra) with native inflection → mixed.
Hindi -ख़ाना, -दार, -गर, बे- marked P (Persian); -ता, -त्व, निर्-, -पूर्ण, -मय, -वान, -नीय, -ईय, -इक, -शाला, -पूर्वक marked S; -पन, -ई, -आई, -आहट, -आवट, -आपा, -वाला, -सा marked N.
English -hood, -ship, -dom, -th, -er, -ful, -less, un-, -ly, -ish, -en, -ing, -let, -ling native (N-analog: Germanic); -ity, -tion, -ment, -ic, -al, -ous, -able, -ess, -ism, -ist, -ize, -ify, -ology borrowed (Latin/Greek/French → register E is meaningless for English itself, so English entries use N for Germanic strata and mixed/S-analog is not used; we use N vs E where E marks the learned Latinate/Greek stratum — an ENGINEERING reuse of the register enum, documented in DATASET.md).

Dependency justifications

Python (core): pydantic (schema + validation, the backbone). API: fastapi, uvicorn. Dev: pytest. Nothing else — no networkx (traversals are trivial), no ORM (no DB), no requests (no outbound calls).

Web: next, react, react-dom, typescript, tailwindcss (+postcss/autoprefixer), openapi-typescript (dev, generates API types), No component library, no state manager, no graph library. d3-hierarchy was planned for the etymology view but removed: the lineage graph is a DAG, not a tree, and a ~30-line longest-path layered layout fit better than forcing hierarchy tooling.

Hindi and Telugu web fonts: subsetting was measured and rejected. The two Indic faces (Anek Devanagari, Anek Telugu) are the heaviest assets on the site (about 390 KB), and pages whose content is largely Indic text score noticeably lower on throttled-mobile Lighthouse than Latin-only pages. We built a glyph-subsetting pipeline (fontTools, weight axis limited to 400–700, OpenType layout features kept) and measured it: the subset keeps 804 of 967 Devanagari glyphs, because the conjunct ligatures that real Hindi and Telugu text needs account for nearly all of them, and the variable-weight data alone is ~260 KB. The best case was a ~30% reduction, in exchange for a custom build step, committed binary fonts and a glyph-coverage check to maintain. We reverted to the simple setup — self-hosted via next/font, not preloaded, loaded only on pages that contain Indic text (the browser applies per-script unicode-range). Note that the Lighthouse penalty is partly a localhost artefact: fonts arrive before first paint locally, so the simulation counts them; on a real network font-display: swap shows fallback text first. Revisit with real-user data after the first production deploy.

v0.2 adds French and Tamil: the T.H.E.F.T. framework. Telugu, Hindi, English, French, Tamil. The two additions were chosen for contrast, not coverage: Tamil is a second Dravidian language with a markedly different stance toward Sanskrit than Telugu (native coinages such as நூலகம் ‘library’ and மொழியியல் ‘linguistics’ where Telugu and Hindi use Sanskrit), and French is both a second Indo-European branch and the main donor of English learned vocabulary. Consequences:

Register L. v0.1 reused E for the English learned Latin/Greek stratum. With French in the graph that overloading became wrong: French needs the same learned stratum, and E is needed for genuine English loans in Tamil and French (சோசலிசம், leadership). L is now the learned stratum in both languages; every English E became L. Mapping L to Sanskrit's S (mots savants ≈ tatsama, mots populaires ≈ tadbhava) is an ENGINEERING analogy, not a claim from a reference work.
Entry limit 300–850 (was 300–500), about 110–115 entries per language.
Cognates within Dravidian. Telugu–Tamil cognates (amma, peyar/pēru, pal/pallu, nīr/nīru, kaṇ/kannu, ūr/ūru, mūṉṟu/mūḍu, iraṇṭu/reṇḍu) cite DEDR or Krishnamurti without entry numbers, capped at 0.7 as before.
French etymologies cite the TLFi (inheritance from Latin) and Watkins (PIE cognacy); borrowings into English cite the OED.
Tamil -தனம் and Telugu -తనం are formally parallel; a shared Dravidian origin is plausible but is recorded only as HYPOTHESIS in the affix note, not as an edge.
The v0.1 report URL (/research/suffeffix-v0-1/) is kept as a superseded notice linking to v0.2, so existing citations resolve.