', unsafe_allow_html=True)
# ── Input ────────────────────────────────────────────────────
EXAMPLES = {
"Factual (Water)":
"Water boils at 100 degrees Celsius at sea level.",
"Hallucinated (Titanic)":
"The RMS Titanic was a British passenger liner operated by the White Star Line. It sank in the North Atlantic Ocean on April 15, 1912, after being struck by a German submarine during its maiden voyage.",
"Mixed (Apollo 11)":
"The Apollo 11 mission was the first crewed moon landing, commanded by Neil Armstrong in 1969. They spent 12 days on the lunar surface and discovered water ice.",
}
col_inp, col_ex = st.columns([3, 1])
with col_ex:
st.markdown("**Load example:**")
for label, text in EXAMPLES.items():
if st.button(label, use_container_width=True):
st.session_state["input_text"] = text
with col_inp:
input_text = st.text_area(
"Paste LLM-generated text to verify:",
value = st.session_state.get("input_text", ""),
height = 160,
placeholder = "Enter any LLM response here...",
key = "input_text",
)
col_run, col_slider = st.columns([1, 2])
with col_slider:
max_claims = st.slider(
"Max claims to verify",
min_value=2,
max_value=10,
value=5,
step=1,
help="Higher = more thorough but slower. Use 2-3 for single sentences, 6-10 for dense multi-sentence paragraphs.",
)
st.caption("Increase for longer or denser text (e.g. historical facts, medical descriptions).")
with col_run:
run_btn = st.button("RUN DETECTION", use_container_width=True)
# ── Run pipeline ─────────────────────────────────────────────
if run_btn:
if not input_text.strip():
st.warning("Please enter some text first.")
st.stop()
# Lazy import after env is ready
from pipeline.orchestrator import run_pipeline
progress_bar = st.progress(0)
status_text = st.empty()
def progress_cb(stage: str, pct: int):
progress_bar.progress(pct / 100)
status_text.markdown(f"{stage}", unsafe_allow_html=True)
with st.spinner(""):
t0 = time.time()
report = run_pipeline(input_text.strip(), progress_cb=progress_cb, max_claims=max_claims)
elapsed = time.time() - t0
progress_bar.empty()
status_text.empty()
st.session_state["report"] = report
st.session_state["elapsed"] = elapsed
# ── Display results ──────────────────────────────────────────
if "report" in st.session_state:
report = st.session_state["report"]
elapsed = st.session_state.get("elapsed", 0)
verdict = report["verdict"]
risk = report["overall_risk"]
# ── Top verdict ──────────────────────────────────────────
st.markdown(
f'
', unsafe_allow_html=True)
for i, cr in enumerate(report["claim_results"]):
claim = cr["claim"]
meta = cr["meta"]
web = cr["web_nli"]
qa = cr["qa"]
cr_class = claim_class(meta["verdict"])
r_color = risk_color(meta["hallucination_risk"])
with st.container():
st.markdown(
f'
'
f'
Claim {i+1}: {claim}
'
f'
'
f'
Risk {meta["hallucination_risk"]:.2f}
'
f'
Verdict {meta["verdict"]}
'
f'
Conf {meta["confidence"]:.2f}
'
f'
Web snippets {web["num_snippets"]}
'
f'
QA score {qa["consistency_score"]:.2f}
'
f'
Model {meta["model_used"]}
'
f'
'
f'
',
unsafe_allow_html=True,
)
# SHAP chart
with st.expander(f"SHAP feature contributions — Claim {i+1}"):
col_shap, col_branch = st.columns([2, 1])
with col_shap:
st.plotly_chart(make_shap_chart(meta["shap_values"]), use_container_width=True)
with col_branch:
st.markdown("**Branch weights**")
bw = meta["branch_weights"]
st.markdown(f"- Web+NLI: `{bw['web']:.0%}`")
st.markdown(f"- QA: `{bw['qa']:.0%}`")
if show_raw:
st.markdown("**Raw features**")
feat_df = pd.DataFrame([meta["features"]]).T.rename(columns={0: "value"})
feat_df["value"] = feat_df["value"].round(3)
st.dataframe(feat_df, use_container_width=True)
# Web snippets
if show_web and web["snippet_scores"]:
with st.expander(f"Web evidence — Claim {i+1}"):
st.caption(f"Search query: `{web['search_query']}`")
for j, s in enumerate(web["snippet_scores"]):
ecol, ccol = st.columns([1, 1])
ecol.metric(f"Snippet {j+1} entailment", f"{s['entailment']:.2f}")
ccol.metric(f"Snippet {j+1} contradiction", f"{s['contradiction']:.2f}")
st.caption(s.get("snippet", "")[:300])
st.markdown("---")
# QA pairs
if show_qa and qa.get("qa_pairs"):
with st.expander(f"QA pairs — Claim {i+1}"):
for pair in qa["qa_pairs"]:
st.markdown(f"**Q:** {pair['question']}")
st.markdown(f"**A:** {pair['answer']}")
cols = st.columns(3)
cols[0].metric("NLI Support", f"{pair['electra_score']:.3f}")
cols[1].metric("BERTScore", f"{pair['bertscore_f1']:.3f}")
cols[2].metric("Contradiction", f"{pair['contradiction']:.3f}")
st.markdown("---")
# ── Full JSON export ─────────────────────────────────────
st.markdown('