Replace 3-gram substring match with:
- _last_sentence(): only the last sentence before a citation is checked
(prevents a hallucinated paragraph grounding via a phrase at its end)
- _content_words(): stopwords removed from both claim and chunk before
n-gram comparison (prevents bypass via "the discrete conformal map")
- Content-5-gram: require 5 consecutive non-stopword tokens from last
sentence to appear in the chunk's content-word stream
- Claims with <5 content words in last sentence → ungrounded by default
Add adversarial tests: stopword-3-gram bypass → grounded=False;
legitimate content-5-gram → grounded=True.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add _detect_conflicts() using adversative keyword heuristic (but,
however, in contrast, …) between chunks of different bibkeys; results
populate CompileReport.conflicts and appear in wiki/log.md.
Also add CompileReport.quarantined field (used by Fix 2).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
_inject_cross_refs temporarily replaces backtick spans with null-byte
placeholders before injecting [[slug]] cross-refs, then restores them.
This prevents concept titles inside `code` from being rewritten.
Add tests for inline-code protection and full-list cross-ref injection.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
compile_concept now accepts chunks as a positional parameter; compile_all
retrieves chunks once and passes them directly, avoiding double-retrieve.
Also separates all_concepts (full list) from compile_concepts (filtered),
so cross-ref injection always uses the complete concept list regardless of
--concept filter (Fix 6 + Fix 8).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
_parse_claims now skips matches where bibkey contains spaces or starts
with "http", preventing Markdown hyperlinks like [text](https://...) from
being misidentified as citations. Add two tests.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wrap llm.generate() in compile_concept with a broad try/except;
log a warning and return an empty ConceptPage (no crash) when the
LLM endpoint is unreachable. Add logging module import.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
