IWAC Sentiment Analysis

A SvelteKit dashboard for exploring how large language models annotated the same corpus — 12,348 francophone West African press articles from 57 newspapers in Benin, Burkina Faso, Côte d'Ivoire, Niger and Togo, published 1961–2025 — on three dimensions: polarity, subjectivity, and the centrality of Islam and Muslims to the article. The current panel is six models; the archived one is three.

The corpus is the Islam West Africa Collection (IWAC). The dashboard reports each model's distributions, their agreement, where they diverge, and blind third-party arbiter verdicts on the sharpest disagreements.

Live site: https://iwac.frederickmadore.com/sentiment-analysis/

Sharing analyses: Citable URLs, supported state and citation limits. Article, comparison and arbiter CSVs append release, model, selection and URL provenance. Downloads capture the selection at click time and load only its required justification shards.

Two top-level areas: ma-visualisation-sentiments/ (the SvelteKit app) and data-preprocess/ (the Python that builds its JSON payloads).

---

Two analysis generations

The corpus has been annotated twice, under two different prompts. Both are published side by side. Generation 2 is what the dashboard shows by default; generation 1 stays online, unchanged, so figures and URLs published from it keep resolving.

	v2 — current	v1 — archived
Models	GPT-5.6 Luna, Mistral Small 4 (2603), DeepSeek v4 Flash, Gemma 4 31B, Qwen3.8 27B, Claude Haiku 5.5	GPT-5 mini, Gemini 3 Flash preview, Ministral 14B 2512
Repo ids	luna, mistral-small, deepseek, gemma, qwen, haiku	chatgpt, gemini, mistral
Pairs	15	3
Articles annotated	12,298 of 12,348 — except Qwen3.8 27B, at 12,098, and Claude Haiku 5.5, at 12,251	12,279 of 12,348
Subjectivity	ordinal label upstream, mapped to rank 1–5; declined → null	integer 1–5
Arbiter	Claude Opus 5, whole panel (one verdict per article); Claude Haiku 5.5 recused	Gemini 3 Pro, pairwise (one verdict per model pair)
Contract	src/lib/data/sentiment-v2.json	src/lib/data/sentiment-v1.json

The generation is derived from the id, never stored. There is no generation URL parameter: generationOf() in src/lib/domain/sentimentContract.ts reads it off the dataset or pair id, which is why the two generations' ids must never collide (an import-time invariant asserts this). ?dataset=chatgpt and ?pair=chatgpt-gemini therefore keep working exactly as before and land you in the archive.

Reaching the archive: the model pickers offer only the current generation. A quiet footnote link in the methodology card switches to v1, after which a banner makes the state unmistakable with one click back.

Reading a v1 ↔ v2 difference

The v2 prompt is not the v1 prompt. Subjectivity is answered as a label (Très objectif … Très subjectif) rather than an integer, the self-checklist instruction is gone, and boundary rules were added for Muslim actors in secular stories, Arab-state cooperation, and armed groups. A difference between the two generations confounds the model change with the prompt change — say so when reporting one. The app never mixes them: agreement statistics, prefetching and comparison pairs are all scoped to the active generation, because a six-rater kappa would measure the prompt rewrite as if it were disagreement between models.

Four further caveats specific to v2:

• The 50 articles that are neither French nor English are skipped by design. The prompt is French, and a French-prompted model returns confident but unusable output for them. They ship as all-null rows.
• Qwen3.8 27B is 200 articles short of the rest of the panel, and permanently so: it declined them across a full pass and three further rounds of retries, after which they were retired. The gap is not missing at random — it falls disproportionately on articles where Islam is peripheral (6.2% of the articles the panel calls Marginal, against 0.7–2.6% of every other centrality band). Anything computed on complete cases — Fleiss' κ, the consensus charts, the arbiter's selection frame — therefore leans towards material where Islam is central. Do not present the gap as repairable.
• Claude Haiku 5.5 is 47 articles short of the other API members. It joined the panel on 2026-10-22, without a pilot, and its corpus pass on 2026-10-23 logged 47 French articles as having no content in Omeka, although the rest of the panel had read them in August. Unlike Qwen's, this gap has a mechanical cause and may be repairable upstream; until it is, those 47 are complete cases for five models and not for six.
• Every model declines subjectivity more often than the other two dimensions, so any statistic involving it rests on a smaller sample than the polarity and centrality figures above. Declined over and above the rows the model skipped entirely: DeepSeek v4 Flash 489 (11,809 usable), Claude Haiku 5.5 406 (11,845), GPT-5.6 Luna 290 (12,008), Qwen3.8 27B 287 (11,811), Mistral Small 4 254 (12,044), Gemma 4 31B 243 (12,055).

chatgpt/gemini/mistral are this repo's ids, not the dataset's column prefixes — the Hugging Face dataset renamed its sentiment columns from vendor slots to the model that actually produced each annotation. iwac_preprocess.contract holds the mapping and sentiment_column() is the only place a column name is assembled.

---

What you can explore

Thirteen views, all sharing one filter rail (country → newspaper → polarity → subjectivity → centrality) except the three self-contained ones. Filter state round-trips through the URL, so any view is shareable.

View	What it answers
Charts	Polarity and subjectivity distributions per newspaper — bars or global pie
Trends	Cross-model historical shares with common or available cohorts; single-model counts/shares
Distribution	Polarity × subjectivity cross-tabulation, with Spearman's ρ, p and n
Volume	Newspaper-by-period collection timeline; publication volume per country
Heatmap	Centrality by country and year, with sample counts and a minimum-count threshold
Seasonality	Coverage across the twelve Hijri months — polar cycle or bars
Newspapers	Titles ranked by mean polarity, subjectivity or centrality, with 95% CIs
Map	Bubble map of geocoded places cited by the corpus, coloured by any dimension
Table	The article list, sortable and paginated; detail modal carries the model's justifications
Comparison★	Two models side by side, with per-dimension discrepancies and where they diverge
Agreement★	Pair overview, full label patterns with article drill-down, κ, confusion matrices and calibration
Extremes	Keywords characterising the most extreme cases per category and model
Arbiter★	Blind third-party verdicts on articles the models disagreed about most

★ self-contained: full-width, own internal filters, no shared rail.

Notes worth reading before interpreting

• Seasonality exists because the Gregorian axis hides it. The Hijri year drifts ~11 days a year, so lunar patterns are invisible in any year- or month-based view. Coverage nearly doubles during Ramadan and the hajj months. Stored Umm al-Qura dates are canonical; Gregorian-only records use a tabular (arithmetic) fallback. Neither establishes local lunar observance, as stated on the chart. Missing scores leave gaps; they never become zero.
• The map counts mentions, not aboutness. A bubble counts articles that mention a place (dcterms:spatial is item-level tagging, ~3.8 places per article), never articles about it. Non applicable is excluded from the mean but the article is still counted — it means "no stance expressed", so averaging it in would drag heavily-covered places toward the negative pole.
• Agreement offers only corpus-scope facets (country, newspaper). Sentiment filters are deliberately absent: selecting by the label under comparison would make the statistics circular. For v2 polarity, weighted κ and within-one-step agreement exclude Non applicable; the categorical matrix and unweighted agreement retain it, with their own denominators. The archived v1 calculation remains unchanged. A higher weighted κ alone does not establish a directional offset. Panel-scope statistics count only articles every model rated, so on v2 they inherit Qwen's coverage gap.
• "Who breaks ranks" means something different at six models than at three. The dissent profile names a model only when it stands alone against all the others; a 4–2 or 3–3 split has no lone dissenter and lands in "divided several ways". The ternary triangle is offered for the three-model archive only — a simplex over six models has no honest 2-D projection.
• Newspaper ranking omits titles under 30 rated articles, and states how many it omitted. Its confidence intervals describe within-sample variation in model ratings, not uncertainty from archival selection, missing articles or annotation error.
• Historical model shares default to a common applicable cohort for the chosen dimension. Available-case mode exposes each model’s own denominator. The pale band spans model minima and maxima; it is descriptive, not a confidence interval. Source-only country/newspaper filters apply, and a minimum count suppresses sparse points.
• The newspaper timeline counts collected articles, independent of model annotations. Empty cells indicate no articles in the selection, not that a newspaper published nothing. All titles are available through pagination and keyboard-operable filter buttons.
• Full label patterns preserve the complete three- or six-model combination and open the matching articles. The Sankey shows adjacent transitions only; merged ribbons do not retain full article paths.

Example URLs

# Charts, filtered to one country
?view=charts&countries=Togo

# Trends in English, positive polarity only
?view=trends&lang=en&polarities=Positif

# Comparison mode on a generation-2 pair, wide discrepancies only
# (the v2 panel has fifteen pairs; `pair` takes any of them)
?view=comparison&compare=true&pair=deepseek-qwen&diffMin=2&diffMax=11

# Heatmap restricted to high centrality
?view=heatmap&centralities=Central&centralities=Très%20central

# The archived generation-1 analysis
?dataset=chatgpt

Core parameters: view, lang, dataset, compare, pair, countries, journals, polarities, subjectivities, centralities, diffMin, diffMax. Each view adds its own analytical and presentation state (comparison dimensions/excludeNA, agreement scope/dimension/declined, detail ids, sort and pagination, chartState); the full per-view list is in Citable URLs. Every name the app reads or writes is declared in src/lib/stores/url/constants.ts (URL_PARAMS) or, for presentation options, as a key of VIEW_OPTIONS.

---

Data

Everything ships as static JSON under ma-visualisation-sentiments/static/data/. The split is for load time, not tidiness.

File	Contents
iwac_articles_base.json	Shared article metadata, stored once for both generations (3.8 MiB, 598 KiB gzip)
iwac_sentiment_{model}.json	One model's scores keyed by article id (~58 KiB gzip)
iwac_justifications_{model}_{00..31}.json	The ~90% prose portion, deterministically sharded by article id
iwac_extreme_analysis_{model}.json	Keyword analysis of extreme cases, normalised (index + per-category id lists)
iwac_arbiter_evaluations_{pair}.json	v1 pairwise arbiter verdicts (Gemini 3 Pro)
iwac_arbiter_evaluations_v2.json	v2 panel arbiter verdicts (Claude Opus 5) — see below
iwac_places.json	Geocoded Lieux authority records plus an article → place edge list
world-110m.geojson	Natural Earth 1:110m basemap (public domain), minified on purpose
iwac_data_manifest.json / _v2.json	Per-generation contract version, immutable HF revision, byte sizes and SHA-256 for every file

The frontend validates each JSON boundary at runtime, joins base metadata to the selected model's score map by article id, and exposes a retryable error state on failure. Only the selected model's scores load at startup; prose loads when an article detail opens (one shard) or a CSV is exported (all shards, in bounded batches); other models load when comparison mode activates or the panel arbiter view opens; the map, arbiter and extremes data load with their views.

Places ship as edges rather than pre-computed averages so the map answers to the same filter rail as every other view; aggregation happens in the browser.

The generation-2 panel arbiter

iwac_arbiter_evaluations_v2.json is the published --rule valence run of data-preprocess/arbiter-evaluation-v2.py: every article where one model reads the coverage as positive and another as negative — 301 articles, judged in 281 paid calls for US$17.61 at effort medium, as recorded in the file's metadata.usage. The arbiter reads the full article text and the five analyses under blind labels, rates each dimension itself, names the analysis it prefers per dimension and overall, and explains why. The arbiter view under v2 lists the articles with their verdicts; opening one shows the article itself (IIIF viewer where a manifest exists), why it was selected, and every model's rating against the arbiter's, with the models' own reasoning on demand. A build that omits the file shows an empty state pointing here. The v1 pairwise arbiter data is complete and unaffected.

Two things to keep in mind when reading the verdicts:

• The frame is the panel's dissenter's frame. Mistral Small 4 is the only model across the valence line in 189 of the 301 articles (162 as the sole negative reader); the strict four-against-one exact-label split is 56. The arbiter's polarity coincides with Mistral's in 37 articles, against 190–205 for each of the other four. Per-model tallies therefore measure how often the dissenter was upheld on the articles where it dissented, not the models' general quality.
• Claude Haiku 5.5 is recused, because the judge is also Anthropic's. Haiku joined the panel after this run, and a judge reading a same-developer analysis is not a neutral judge: blind labels hide a model's name but not its house style, and LLM judges measurably favour their own output and that of related models (Panickssery et al., 2024; Wataoka et al., 2024; Goel et al., 2025). The contract's arbiter.recusal therefore keeps Haiku out of the arbiter's panel altogether — selection, prompt, blind labels and fingerprints run over the other five, which is exactly the panel the published run judged, so no verdict was invalidated and nothing was re-bought. Haiku is compared with the verdicts afterwards: the verdict panel shows its rating beside the arbiter's, unlettered and marked Recused. On the 301 articles its polarity coincides with the arbiter's in 192, inside the 190–205 of the four members other than Mistral Small 4, so there is no sign of a reading the two Anthropic models share. What recusal cannot do is hear Haiku's own dissents: with six models the panel has 346 valence flips, and the 45 that exist only because of Haiku (38 as the sole negative reader, 7 as the sole positive one) are outside the frame. Settling those needs a judge from another developer — and since every frontier developer now has a model on the panel (OpenAI, Google, Anthropic, Mistral, DeepSeek, Alibaba), that judge must recuse its own developer's member in turn.
• Recusal is derived from the judge's developer, never listed. Every model in sentiment-v2.json names its developer, the arbiter names its own, and arbiter.recusal.rule is same-developer; SentimentContract.recused_for() / recusedFor() compute the recused set in both languages. An unregistered developer is an error rather than a judge that recuses nobody. The published fingerprints are unchanged (a test pins one).
• A second judge is built but has not run. data-preprocess/arbiter-evaluation-v2-second-judge.py runs a judge from another developer — Gemini 4 Argon by default, recusing Gemma — over the six-model valence-flip frame, with the same prompt, labels, presentation order and schema as the published run, and its own output file (iwac_arbiter_evaluations_v2_<judge>.json), durable cache and lock. --frame recused judges only Haiku's 45; judges all 346, so the 301 shared with the published run give a judge-agreement figure and a difference-in-differences estimate of each judge's affinity for its own developer's model. The dashboard does not read that file yet.

How the v2 arbiter differs from v1, and why:

• Selection is the spread across the whole panel (max minus min across every model, per dimension) rather than a pairwise gap. On the arbiter's five-model panel the contract's spread rule (any dimension ≥ 3) selects 2,102 articles, but 1,762 of those are triggered by subjectivity alone — the dimension the models argue about most and where "who is right" is least well defined. Polarity triggers 103 and centrality 239. Widening the panel widened the frame: three models selected 1,449. A polarity spread of 3 already implies a sign reversal, so selecting on spread finds only the 103 widest reversals and structurally cannot reach the two-rank Positif/Négatif ones; the contract's second rule, valenceFlip (--rule valence), is what selects all 301. --rule, --dimensions and --threshold narrow the frame; they can only ever tighten it, because the validator recomputes eligibility from the contract.
• The arbiter reads unmasked article text. This fixes a real v1 flaw: the public Hugging Face projection masks OCR per row, so a large share of v1-arbitrated articles were judged on an empty string. v2 reads the private mirror (needs HF_TOKEN) and records both revisions — public scores and private text — in the metadata and the cache fingerprint. Only verdicts and justifications are published; no OCR is ever serialised.
• Blind assignment is one global permutation of the arbiter panel's model ids onto labels A–E, persisted and reused by every incremental run, so cached and new rows always mean the same thing. parseArbiterV2EvaluationData rejects any file whose permutation is not a bijection over the contract's models, which is also what keeps the label list and the panel the same size.

Regenerating the data

Python 3.12+, from the repo root:

# Development and CI, including all runtime dependencies
pip install --require-hashes -r data-preprocess/requirements-dev.lock

# Runtime-only environment for generating data
pip install --require-hashes -r data-preprocess/requirements.lock

The .lock files pin every dependency and its hashes. The .txt files are the human-maintained version constraints; after changing them, regenerate both locks with the commands documented in data-preprocess/requirements.txt. The development lock is constrained to the runtime lock so generation and CI use the same runtime versions. CI audits that exact runtime set with pip-audit -r data-preprocess/requirements.lock --disable-pip --require-hashes.

Newly generated data manifests also record the runtime lock's filename and SHA-256 under environment.requirements_lock, identifying the dependency specification available to that generation. This optional field is absent when no lock is available; it does not verify the packages installed in the running interpreter. Existing published manifests are unchanged and make no claim to have been generated with the new lock.

Source reads resolve a single immutable Hugging Face commit per repository before loading. Set IWAC_HF_REVISION for the public dataset and IWAC_HF_FULL_REVISION for the private OCR mirror to reproduce a prior run; both the parquet and datasets fallback paths use the resolved commit.

Every export — scores and prose, extreme analysis, places, basemap and the panel arbiter — is staged and validated before publication, then promoted under one writer lock, so two exports can never interleave a mixed set. A recovery journal restores the previous complete set after an interrupted promotion. The lock, temporary export stages and retained panel-arbiter cache live in ma-visualisation-sentiments/.data-staging/ (gitignored), never under static/, whose contents are published with the site. Run the same export command again to recover. Builds refuse a pending journal. Production builds bind the frontend to content-addressed data/releases/<release>/ URLs; the service worker never substitutes files from a different release. The source static/data/ layout and frozen v1 files remain unchanged.

--generation is required and deliberately has no default. An unflagged re-run would rewrite the frozen v1 files from whatever revision is current. A v2 run does not write iwac_articles_base.json at all — it fails loudly if an article in the base was deleted upstream.

The base is the corpus, and it never grows by itself. Every export is scoped to the base's article ids: articles ingested upstream since the panel ran are excluded and logged, because only part of the panel has read them — in October 2026 the collection gained 4,367 Burkina Faso articles that carry Claude Haiku 5.5's ratings and no other model's. A --generation v1 run drops articles deleted upstream and admits new ones only with --admit-new-articles, which gives every model that has not annotated them null rows. Run v1 first, then v2, then the extremes and places exports, as the 2026-10-27 refresh did when article 78587 was deleted.

# Scores, prose shards and the per-generation manifest
python data-preprocess/data-fetch.py --generation v2

# Keyword analysis of extreme cases
python data-preprocess/extreme-analysis.py --generation v2

# Map payload (generation-independent)
python data-preprocess/places-export.py

# Basemap — only when the Natural Earth source itself changes
python data-preprocess/basemap-export.py

Arbiter runs cost money. Use a dry run first; a paid run requires confirmation or --yes:

# v2, whole panel: counts and a cost estimate, no API call
python data-preprocess/arbiter-evaluation-v2.py --dry-run

# ...then, to actually spend (needs ANTHROPIC_API_KEY and HF_TOKEN)
python data-preprocess/arbiter-evaluation-v2.py --rule valence --yes

# Republish a narrower v2 selection from cache, without paid calls
python data-preprocess/arbiter-evaluation-v2.py --rule valence --limit 20 --prune-cache-only

# v2 second judge: frame counts and a cost estimate for both frames and every
# registered judge, no API call
python data-preprocess/arbiter-evaluation-v2-second-judge.py --dry-run

# ...then a pilot and the full frame (needs GOOGLE_API_KEY and HF_TOKEN)
python data-preprocess/arbiter-evaluation-v2-second-judge.py --frame full --limit 20 --yes
python data-preprocess/arbiter-evaluation-v2-second-judge.py --frame full --yes

# v1, pairwise (needs GOOGLE_API_KEY); frozen — reconcile only
python data-preprocess/arbiter-evaluation.py --prune-cache-only

The v2 arbiter bootstraps its durable cache from the existing published file and retains it at ma-visualisation-sentiments/.data-staging/arbiter-v2/evaluations.json. Each input fingerprint retains its verdict and source/model/prompt/effort provenance, including evaluations omitted by a later selection or belonging to an earlier source revision. Narrowing --limit or changing the selection rule only changes the public projection; expanding it again reuses matching paid work. --prune-cache-only makes no paid calls and never prunes this durable history. Recorded usage remains cumulative, rather than becoming the cost of the narrower selection.

A populated cache must retain its original valid blind-label mapping and recorded effort. A missing/invalid mapping or a different --effort stops the run before either cache is rewritten; restore the mapping or resume at the recorded effort. Run a different-effort experiment in a separate checkout with its own output/cache instead of relabelling an existing run. --limit must be a positive integer and is checked before source loading. A dedicated lock covers the whole panel-arbiter session, and paid verdicts are checkpointed before the public export, so a failed publication can resume without repeating those calls. A dry run does not write the durable cache.

Back up the arbiter-v2/ directory before deleting ignored files or changing checkouts. Git does not retain it, and the selected public file cannot recover paid evaluations that were never published or were omitted from the current selection. Restore the directory alongside the matching checkout to continue that history. Neither cache nor publication stores article OCR.

data-preprocess/significant-differences-export.py is a side helper rather than part of the pipeline: it writes ad-hoc CSVs of the sharpest pairwise disagreements into exports/, nothing the dashboard reads. It resolves its pairs through shared.py, so it still covers generation 1 only.

Validate before committing generated data:

python -m pytest data-preprocess -q && python data-preprocess/validate_generated_data.py

An offline test suite, plus a validator that checks category domains, article-id coverage, prose shard placement, arbiter eligibility and fingerprints, and manifest hashes — entirely offline, for both generations. Both Python and TypeScript read the same checked-in contracts, and shared fixtures assert identical discrepancy, label-mapping and prose-shard behaviour across the two languages.

Environment variables live in a root .env — copy .env.example: ANTHROPIC_API_KEY (v2 arbiter), HF_TOKEN (private mirror), GOOGLE_API_KEY (v1 arbiter). Only the arbiter scripts load that file.

---

Development

Node 24.15+ (24.x LTS) or 26+. .nvmrc and CI use Node 24 LTS. The Vitest 5 component suite uses jsdom 30, which requires this minimum patch level.

cd ma-visualisation-sentiments && npm install && npm run dev

Script	What it does
npm run dev	Dev server at localhost:5173
npm run build	Production build, then nesting, service-worker stamping and artifact checks
npm run preview	Serve the final Pages artifact (including postbuild releases)
npm run check	svelte-check — must report 0 errors, 0 warnings
npm run lint	Prettier, ESLint, store-cycle detection, design-token validation
npm run test:run	Vitest 5 unit and integration tests
npm run test:e2e	Playwright deep-link, failure-state and axe accessibility smoke tests

Run the checks in .claude/skills/verify/SKILL.md before committing anything that ships through CI — the ordering and pass criteria are not guessable. CSS and component rules are in DESIGN.md, and npm run lint enforces most of them.

Two repo-level helpers run against the root .venv:

• python scripts/social-preview.py regenerates static/social-preview.png from the shipped data. Re-run after a data refresh.
• python scripts/oklch-to-hex.py prints the sRGB renderings of the OKLCH design tokens, for the chart palette and theme-color meta tags.

Architecture

• src/lib/components/ — grouped by role: common/, layout/, filters/, viz/ (ECharts charts and the MapLibre map), ui/, data-display/.
• src/lib/stores/ — Svelte 5 runes accessor objects, one per domain (filters, articles, datasets, comparison, arbiter, arbiterV2, extreme-analysis, ui), plus url/ for filter-state ↔ URL synchronisation. Leaf stores import from no other store; modules inside stores/ must never import the barrel.
• src/lib/utils/ — pure helpers, including the statistics modules: agreement.ts (Cohen's/Fleiss' κ), correlation.ts (Spearman's ρ), newspaperRanking.ts, hijri.ts, placeAggregation.ts.
• src/lib/domain/sentimentContract.ts — the dual-generation registry, with import-time invariants for id collisions, pair membership, shared scales and shard counts.
• data-preprocess/iwac_preprocess/ — importable package split into the per-generation contract, revision-pinned source loading, discrepancy rules, recoverable generation publication, panel candidate selection and prompt construction, atomic serialisation, v1 cache reconciliation and the durable v2 cache lifecycle (arbiter_lifecycle.py). shared.py remains a compatibility facade for the command-line scripts.

Built with Svelte 5 (runes), SvelteKit 3 + adapter-static, TypeScript in strict mode, Tailwind CSS v4, ECharts 6, MapLibre GL v6, and Vite 8 (Rolldown). The design system is repository-owned: CSS tokens plus reusable Svelte controls, with automated token and class checks.

Performance

HTML-referenced JavaScript is measured at build time against a 300 KiB gzip budget — currently 139 KiB. The production browser tests additionally count actual transitive and dynamic imports: default charts, direct table and map links have budgets of 450, 180 and 600 KiB gzip, respectively. Data transfer is reported separately, including all six background score prefetches. The tests verify chart/map isolation using the generated chunk manifest. These measurements are build-size budgets, not claims about latency on every device.

Views load their modules on demand and offer a localized reload action when a chunk fails. Data URLs include a content-derived release identifier; unchanged data keeps its cache across code-only deployments, while a changed dataset gets a new namespace. Offline requests never substitute another release's files. The build intentionally emits no .gz/.br siblings — GitHub Pages negotiates transfer compression. npm run preview serves the final static artifact so browser tests exercise the same release paths and stamped service worker that Pages receives.

---

Deployment

Every push to main runs .github/workflows/deploy.yml. Frontend and Python jobs run in parallel: type checks, lint, unit tests, the measured production build, Playwright/axe, Python tests, Ruff, npm audit, pip-audit and pull-request dependency review all gate deployment. Actions are pinned to immutable commit SHAs, and only the deploy job receives Pages permissions.

The site is served from a custom subdomain at a sub-path, so the build output is nested: the adapter writes the app into build/sentiment-analysis/, and scripts/nest-build.mjs populates the build root with the three files GitHub Pages only reads from there (CNAME, a hoisted 404.html, and a redirect index.html). This is necessary because a GitHub project page supplies the /repo-name/ prefix itself, whereas a custom domain serves the artifact at the subdomain root.

ma-visualisation-sentiments/deploy.config.js is the single source of truth for the path and the domain — changing where the dashboard is served means editing that file and nothing else. DEPLOY_PATH = '' serves it at the subdomain root instead.

Citation

Each release is archived on Zenodo. Machine-readable metadata lives in CITATION.cff; GitHub's Cite this repository button renders APA and BibTeX from it.

Cite the concept DOI — 10.5281/zenodo.21806223 — unless you need to pin a specific version; it always resolves to the most recent release.

Madore, F. (2026). IWAC Sentiment Analysis Visualization (Version 5.0.0) [Computer software]. University of Bayreuth. https://doi.org/10.5281/zenodo.21806223

@software{madore_iwac_sentiment_analysis,
  author    = {Madore, Frédérick},
  title     = {IWAC Sentiment Analysis Visualization},
  version   = {5.0.0},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21806223},
  url       = {https://github.com/fmadore/IWAC-sentiment-analysis},
  license   = {MIT}
}

Cite the underlying corpus separately — the Islam West Africa Collection (islam.zmo.de) and its machine-readable export (doi:10.57967/hf/9857).

Acknowledgements

Eight of the nine panel members across the two generations ran at a vendor's API. Qwen3.8 27B did not: it was self-hosted with vLLM on festus, the cluster of the Bayreuth Centre for High Performance Computing, which is why its column in the panel carries no invoice. The centre states that its funding acknowledgement is mandatory rather than a courtesy, so any publication reporting generation-2 results carries it verbatim:

Calculations were performed using the festus-cluster of the Bayreuth Centre for High Performance Computing (https://www.bzhpc.uni-bayreuth.de), funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 523317330.

The dashboard carries the same sentence in its generation-2 run configuration. Wording checked against the festus system guide on 15 September 2026; CITATION.cff has no field for a funding acknowledgement under CFF 1.2.0, so it is not recorded there.

License

The application code is released under the MIT License. The IWAC corpus it visualizes is distributed under CC BY-NC-SA 4.0 and is not covered by that licence — reuse of the data follows the collection's own terms.