Repository navigation
fix: render lists, bold and links in articles, and stop blocking bank-verified claims - #163
Conversation
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
|
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 89cc305bcb
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
Codex Review: Didn't find any major issues. Already looking forward to the next diff. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
…ticle bodies One shared Markdown<->Lexical module (cms/src/lib/richText.ts) replaces the duplicated converters. The writer's lists, bold, italic, code, quotes and http(s)/site-relative links now become real Lexical nodes, the public page renders them with an escaped URL allow-list, and the readers emit one line per list item and line break so QA text no longer runs words together. The writer prompt names the supported subset. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Also fixes the rich-text round-trip test setup and a stale lint directive. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A line of unclosed `*` or `_` openers rescanned the rest of the line once per opener: quadratic, reachable from a template save or the writer's output (120 KB took 18 s). Remember where a closer search already failed. The CMS `lexicalToPlainText` returns plain words again (reading time and the banned-phrase scan read it); the template textareas use the new `lexicalToMarkdownText`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A first_party_measurement claim that firstPartyOutcome verified against a bank entry the draft cited now carries qualitySource 'evidence-bank', and decidePolicy no longer emits FIRST_PARTY_MEASUREMENT_PRESENT for it. Unbacked first-party claims are still blocked. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A verifier rubric guess of primary or official_docs is now suggested as secondary, so accepting a Source review suggestion unchanged can no longer create a floor-clearing rule from a model guess. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…as them
evidenceRefsIn and stripEvidenceRefs take an optional { bank }. With it,
[E<n>] is always a marker (a hallucinated [E9] still reaches QA as
unknown), while [F<n>] and [R<n>] are markers only when that ref exists in
the bank; otherwise ordinary prose like "[F5]" is left untouched. Without
the option the behaviour is unchanged until generate.ts passes the bank.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ility The judge prompt requests the field in every call, so a missing or non-numeric value is refused like duplicateProbability instead of being read as 0, which hid internal duplicates from review. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ers as one amount 5-10% now reads as two percents and $5-10 as two currency amounts, so they match 5% to 10% and $5 to $10. The re-kinded bound is flagged alsoBareNumber and still matches whatever the bare number matched, and anchors are chosen with the bare reading, so matching only gains. Tokenisation is unchanged; only two shapes are relabelled, and years, signed numbers, figures glued to letters, suffixed amounts, to-ranges and spaced em dashes read as before. twenty-five and twenty five read as 25 rather than 5 (or as a 20-5 range). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The passage still quoted an `EVIDENCE_RULES` constant that said Datum has no first-party data. Describe `evidenceRules` with and without a bank, the evidence-bank section, and when the rules appear. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The bank-aware marker rule had no effect in real runs: the generate stage still called the readers without the bank, so "[F5]" and "[F12]" in prose were stripped and recorded. Pass the workspace's bank to every reader and stripper, label the new evidence-bank quality source on the review page, regenerate the policy field's description, and document the rule. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Scoring matches a citation's excerpt against the body read back as plain text, so a cited sentence carrying a link or bold never matched and a bank-backed measurement stayed blocked. Read each excerpt through the shared converter before storing it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ape test Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…aved again Neighbouring text nodes with different marks were written as `**Bold***italic*`, which the parser reads as literal asterisks, so a second save of a template corrupted it. Join neighbours that share marks, and write the italic side with `_` where two `*` runs would touch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
24c39b4 to
37099c7
Compare
The style guide asks the writer for lists, bold text and links, but the converter only understood headings, so readers saw raw
- item,**term**and[text](url)on published pages. Some scoring rules also blocked articles that were correctly sourced. This PR addresses the open audit findings for content and scoring.Rich text. One shared, dependency-free converter (
cms/src/lib/richText.ts) now handles headings, lists, quotes, bold, italic, inline code and links (http(s) and site-relative only). The public page renders it escaped. QA, reading time and phrase matching read it back as plain text, and template fields read it back as markdown. The generate prompt names exactly this subset.Scoring.
FIRST_PARTY_MEASUREMENT_PRESENTblock.secondary.[F<n>]/[R<n>]count as evidence markers only when the bank has that ref. The generate stage now passes the bank in, and citation excerpts are stored as plain text.internalDuplicateProbabilityis refused.5-10%,$5-10) take the range's kind and still match everythingmainmatched. Compound word numbers (twenty-five) read as one amount.Proof: typecheck and lint clean; pipeline 920/920; CMS integration 862/862 after rebasing on
main. Each fix has a test that failed before it.Reviewed with sol-loop: 2 rounds with GPT-6.1 Sol. Sol verdict: approved
Created with Claude Opus 5.5 (implementer subagents Claude Opus 5.5 and Claude Sonnet 5.5) in Claude Code.
🤖 Generated with Claude Code