Prisme

Prisme shows each day’s front-page news from the major French outlets, and how each political leaning covers it — same story, every leaning, side by side.

How it works

Every build (once an hour, all 24 hours, triggered by an external cron-job.org job that calls the workflow's workflow_dispatch API — GitHub's own scheduler dropped most hourly slots) runs the pipeline, then commits the result to data/ — the git repo is the database, and CI publishes it as a static site. Purple-stroked nodes run a language model; dashed nodes are stores on disk — only what changed is written or re-read.

flowchart LR
    classDef store stroke-dasharray:4;
    classDef model stroke:#8250df,stroke-width:2;
    A[Collect<br>every Outlet’s RSS feed] --> B[Classify<br>Decision model — Kind + Section]:::model
    B --> C[Group<br>Grouping model — new Articles into Stories]:::model
    C --> T2[Topic<br>Membership — threads Episodes into sujets:<br>same-subject checks, label + slug]:::model
    T2 --> T[Title<br>Title pass — every Story title anchored on<br>a Jev-confirmed thread]:::model
    T --> D[Rank<br>front pages, Coverage, recency → Edition]
    D --> E[Summarize<br>Flash-Lite — Summaries + Differences]:::model
    E --> F[(Write<br>edition.json + topics.json + story files)]:::store
    F --> G[Publish<br>Astro prerender]
    B -. new ids only .-> H[(Classification cache)]:::store
    C -. changed Story files .-> I[(Story files)]:::store
    I -. episode matching .-> T2
    I --> D

Loading

Five cheap model steps run the show, three pinned models; each pinned for traceability.

Step	Model	Provider
Classify (Kind + Section), Membership and Topic checks	Jev (typesafe/jev-1.13)	OpenRouter — chosen by a hand-labeled benchmark
Group (Stories from new Articles; never titles)	Flash-Lite (gemini-3.5-flash-lite)	Gemini free tier
Topic (Sujets: same-subject checks, labels, slugs — data/topics.json)	Jev (typesafe/jev-1.13)	OpenRouter — one check per birth that matches an episode
Title (every Story title, at birth and on retitle)	Flash-Lite (gemini-3.5-flash-lite)	Gemini free tier
Summarize (Summaries + Differences)	Flash-Lite (gemini-3.5-flash-lite)	Gemini free tier

Leanings are set by hand, one per Outlet, backed by sources — never classified by a model.

Fixing a bad split: merging two Stories by hand

The Grouping model sometimes splits what a human judges should be one Story — two pages about the same development, each holding a handful of Articles. pnpm merge-stories merges them by hand:

pnpm merge-stories <base-slug> <absorbed-slug>            # review, then confirm
pnpm merge-stories <base-slug> <absorbed-slug> --dry-run  # print the plan only

The slugs are the story pages' own (/<day>/<slug>/). The base (first argument) survives whole — its URL, title, Summaries and Topic stay — and the absorbed story (second) gives it its Articles and disappears. The next hourly build re-summarizes the merged set from scratch, so the bullet points take the incoming Articles into account.

Guard rails: the command prints its plan and asks for confirmation before writing, and never commits — it prints the git command to review and paste. Two merges are refused without --force: across story days, and into a frozen Story (whose Summaries the pipeline never rewrites). A Story that once ranked into the Edition keeps its published URL alive after the merge: a redirect entry lands in data/redirects.json, and the site serves the old path as a 0-second meta-refresh redirect to the surviving page (Astro's built-in redirects — honored by the dev server and emitted as stub pages by the build).

Benchmarking the Decision model (Kind + Section)

Same pattern as the grouping benchmark (issue #4): hand labels check the models, docs/research/decision-model-benchmark.md carries the graded results and the winner rule (not_news precision first — never wrongly drop a real Article — then kind accuracy, then section accuracy).

1. Sample — pnpm benchmark:export fetches every Outlet's feeds and samples ~110 real headline+teaser pairs (seed 4; 70 general, 40 “suspects” oversampled by a rare-kind heuristic — horoscopes, live blogs, tribunes…). Writes the gitignored .benchmark/samples.json (grader input) and .benchmark/worksheet.md (what you read), plus the committed docs/research/decision-model-labels.json — ids and URLs only, teaser text never enters git (ADR-0003).
2. Label — fill kind and section for every item in decision-model-labels.json (kinds: news, opinion, live, not_news).
3. Grade — pnpm benchmark:grade replays every configured model over the labeled sample and rewrites docs/research/decision-model-benchmark.md. Keys from .env: OPENROUTER_API_KEY (OpenRouter models), JEV_API_KEY (Jev direct), CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID (Clef-flash) — a missing key just skips that model. Needs network (samples are taken from live feeds; grade replays the gitignored teaser text).

Benchmarking a new Grouping model

The pinned Grouping model can be challenged anytime — issue #41 settled the method, docs/research/grouping-benchmark.md carries the full methodology, recorded results and the decision rule. The short recipe:

1. Pick a candidate on OpenRouter: paid (skip :free listings and $0 prices — free-tier variance is what this escapes), structured outputs, 1M context. Sort by input price.

2. Run it from a checkout whose data/ matches the recorded edition (every results file carries its fixtureBuiltAt — the fixtures rebuild from data/, so new pipeline runs move the reference; re-run the baseline alongside when the edition moved):

   pnpm benchmark:grouping -- gemini/gemini-3.5-flash-lite \
                            openrouter/<vendor>/<model>
   # large-batch give-ups? try the chunked variant:
   pnpm benchmark:grouping -- openrouter/<vendor>/<model>:chunked150

3. Read .benchmark/grouping/report.md — it aggregates every recorded results file, so reruns extend the comparison.

4. Apply the decision rule: dup-seeds 0, coverage 100%, split rate ≤ baseline on both fixtures, < $0.50/build, two valid runs for stability.

Keys come from the local .env (GEMINI_API_KEY, OPENROUTER_API_KEY) — never the repo. Results land in gitignored .benchmark/grouping/; a candidate that cannot answer one build call in 5 minutes is disqualified by that fact (production aborts grouping at 60 s).