The full scoring rubric, used by the teacher panel that labels the PRUUF judge’s training data. It does not score what readers see: the judge scores every reader’s article with its own fixed prompt.
- Runs
- After the reader already has their result (labelWithTeachers, from the teacher queue), and for admin comparisons on the Compare page and A/B tests.
- Reads
- The article, the research verdicts and prior facts, plus the category weights from Remote Config through the weight tokens below.
- Produces
- Category scores in the live rubric (v3: Facts & Sources, Framing, Reasoning, each 1–4; v2: eight categories, each 0–4), the opinion share and a summary per teacher; the panel’s median, grounded, becomes one training row in
training_log. - Shows up
- Nowhere directly. It shapes the next judge that is trained on these labels.
- Model
- The teacher panel on the Models page (ensemble members and the lead teacher).
A change here changes future training data, not today’s scores. Mixing old and new labels in one training run blends two rubrics, so note the date of any material edit.
config/scoring.rubric): …. v3 takes the v3 tokens and weights from Remote Config fvi_category_weights_v3; v2 the v2 tokens and fvi_category_weights.Checks the article’s factual claims against the web before anything is scored. Everything after it, the score, the explanation and Go Deeper, leans on the verdicts it returns.
- Runs
- First, on every article that is not answered from the cache. Skipped when the facts knowledge base already holds 3 or more fresh, relevant verified claims (≤ 30 days).
- Reads
- Today’s date
{{TODAY}}and the article’s first 5,000 characters{{EXCERPT}}. - Produces
- One line per claim:
claim → Confirmed | Contradicted | Unverifiable — reason, with the search results it relied on. - Used by
- The PRUUF judge (as evidence), grounding L2 (a contradicted claim lowers Factual Accuracy), the facts knowledge base, Why this score, Go Deeper and the writer.
- Model
- Research model and provider on the Models page (default Claude Haiku), one web search, 45 s limit.
This prompt is used only when the research provider is Anthropic, where Claude searches and judges in one call. With Tavily selected, research runs on two built-in prompts instead (pick the claims, then judge them against the search results), and edits here have no effect.
Writes the parts of the reader’s panel the PRUUF judge was never trained to produce, from the judge’s finished, grounded result. It never re-scores anything: what it writes must agree with the score.
- Runs
- After grounding, on every new analysis, alongside Why this score and Go Deeper.
- Reads
- The final scorecard and verdict, the judge’s summary, the research verdicts on the article’s claims, and the article text. No tokens are substituted.
- Produces
takeaway: the panel’s headline finding, at most 12 words, usually “what it gets right. Its weakness.”topic:health,finance,legalornone, which decides whether the “Pruuf isn’t medical / financial / legal advice” note shows. Defaultnone; set only when a reader might act on the article and be harmed.article_summary: a neutral 2–4 sentence account of what the article reports.- Shows up
- THE TAKEAWAY block in the extension panel and on shared result pages, the topic note under it, and the takeaway line in text-message replies.
- Model
- Writer model and provider on the Models page (default Claude Haiku).
- Stored
…, the field of the live rubric (writerPromptV3on v3,writerPrompton v2). Empty means the built-in copy infunctions/writer-prompt-v3.jsorwriter-prompt.jsis live; edits apply within about 5 minutes.
article_summary is no longer shown anywhere (“what the article says” was removed in extension 1.55.0), but this prompt still asks for it, so every analysis pays to write it.
Writes the section worth opening: how to read an article like this one, grounded in what research found. It is a separate call so that scoring can stay clinical while this part is written to be read.
- Runs
- Last, on every new analysis, because it reads what the earlier phases produced.
- Reads
- The article, each extracted claim with its verdict, the research notes, the pages research retrieved, and the grounded scorecard. No tokens are substituted.
- Produces
- The article’s kind (one of ten, e.g. breaking news, shopping, research coverage), a gravity register that keeps wit off grave subjects, 2–3 short bullets, two “learn more” paragraphs, a commercial flag and note, and up to three sources with tiers T1–T4.
- Shows up
- GO DEEPER · <KIND> in the panel’s Analysis tab and on shared result pages, translated into the reader’s language.
- Guards
- Only URLs the research phase actually retrieved may be cited; any other URL is dropped before it reaches the reader. Dashes are normalised and the language is checked against the article’s.
- Model
- Go Deeper model and provider on the Models page (default Claude Haiku).
Single-model scoring carries roughly ±14 points of run-to-run noise — use several articles and 2–3 repeats before trusting a delta.
| Article | A score | B score | Δ score | A framing | B framing | Δ framing | Cost |
|---|
| Category (mean across articles) | A | B | Δ |
|---|
