Add factual-accuracy and sourcing guardrail to content-generation prompts

New pipeline/factual_accuracy.py holds a strict no-fabrication guardrail
(no invented statistics/citations, no merged/guessed sources, no false
authority, prefer omission over confabulation, real captions/alt text only,
flag uncertainty, stay internally consistent). Injected into the two prompts
that actually produce published content: writer._build_system() (the blog
body) and captions.py's new system message (image captions/alt text).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

pipeline/captions.py

@@ -9,6 +9,11 @@ from typing import List
 from huggingface_hub import InferenceClient
 
 from . import config
+from .factual_accuracy import FACTUAL_ACCURACY_GUIDELINES
+
+_CAPTION_SYSTEM = (
+    "You are writing image captions for a published blog post.\n\n" + FACTUAL_ACCURACY_GUIDELINES
+)
 
 _CAPTION_PROMPT = (
     "Write a concise, engaging one-sentence caption for this blog illustration. "
@@ -34,13 +39,14 @@ def caption_images(client: InferenceClient, images: List[dict]) -> List[dict]:
             resp = client.chat.completions.create(
                 model=config.MODEL_VISION,
                 messages=[
+                    {"role": "system", "content": _CAPTION_SYSTEM},
                     {
                         "role": "user",
                         "content": [
                             {"type": "text", "text": _CAPTION_PROMPT},
                             {"type": "image_url", "image_url": {"url": _data_uri(Path(path))}},
                         ],
-                    }
+                    },
                 ],
                 max_tokens=80,
                 temperature=0.5,

pipeline/factual_accuracy.py

@@ -0,0 +1,59 @@
+"""Factual-accuracy and sourcing guardrail injected into every content-generation prompt
+(the blog writer and the image captioner) — anywhere an LLM produces text that ends up in
+the published post.
+"""
+
+FACTUAL_ACCURACY_GUIDELINES = """\
+## Factual accuracy and sourcing
+
+You are writing published content that readers will trust. Fabricating facts or
+citations is the most serious error you can make — worse than being incomplete,
+vague, or shorter than requested. Follow these rules without exception.
+
+- Never invent specifics. Do not state statistics, figures, dates, word counts,
+  percentages, technical techniques, or named phenomena unless they are real and
+  you are confident they are accurate. A plausible-sounding number or technique is
+  not acceptable if you are not sure it is true. When you lack a specific figure,
+  describe the point qualitatively ("save files can grow large") rather than
+  inventing a precise-sounding claim ("saves exceeded 1MB due to flag tracking").
+
+- Cite only what you can verify. Attach a claim to a named source (author, article
+  title, book, study, tool, video) ONLY when you are certain both that the source
+  exists AND that it actually makes that specific claim. If unsure of either, do
+  not name a source: make the point without attribution, or drop it. Never guess
+  an author or a title.
+
+- Do not merge or confuse sources. Never combine details from two half-remembered
+  sources into one citation, attach one source's idea to another's title or author,
+  or cite the same work under different names. Keep distinct ideas attributed to
+  distinct sources, and only name each if you are certain of it.
+
+- No false authority. Do not cite anonymous forum threads, random social-media
+  posts, or unnamed videos as authoritative references. If a claim's only basis is
+  an unverifiable or low-quality source, omit it or state it plainly as a general
+  observation — never dress it up as a cited fact.
+
+- Separate established concepts from specific assertions. You may explain
+  well-established general concepts in your own words without citation. But any
+  specific assertion — a statistic, a claim about what a particular game/product/
+  person did, a study's finding, a coined term — must be genuinely true and, if
+  attributed, correctly attributed.
+
+- Prefer omission over confabulation. If you cannot support a point with something
+  real, leave it out. An accurate, slightly less detailed piece beats a detailed
+  one containing invented facts. Never fill a gap with a fabricated example, quote,
+  statistic, or reference.
+
+- Flag genuine uncertainty. When a claim is uncertain or contested, hedge honestly
+  ("estimates vary," "widely repeated but poorly documented") instead of asserting
+  it with false confidence.
+
+- Stay internally consistent. Names, dates, numbers, and terminology must agree
+  throughout and not contradict each other (e.g., a "published" date must not come
+  after its "last updated" date).
+
+- Real captions and alt text only. Every caption, alt text, and label must be
+  meaningful, correct text. Never output placeholder, garbled, or nonsense strings.
+
+When in doubt, leave it out.
+"""

pipeline/writer.py

@@ -13,6 +13,7 @@ from huggingface_hub import InferenceClient
 
 from . import config, llm
 from .aeo_guidelines import AEO_GUIDELINES
+from .factual_accuracy import FACTUAL_ACCURACY_GUIDELINES
 
 # Content-goal directives shape the post's stance and tone. One is selected in the UI.
 GOAL_GUIDANCE = {
@@ -66,6 +67,7 @@ def _build_system(target_wordcount: int, content_goal: str) -> str:
         "- End with a short conclusion.\n\n"
         "AI-citation (AEO) requirements:\n"
         f"{AEO_GUIDELINES}\n"
+        f"{FACTUAL_ACCURACY_GUIDELINES}\n"
         "Return Markdown only — no preamble, no code fences. Use Markdown pipe tables for "
         "any tables."
     )