Integrate Hans reviewer workflow improvements

.streamlit/config.toml

@@ -0,0 +1,10 @@
+[theme]
+base = "light"
+primaryColor = "#176b5b"
+backgroundColor = "#f4f6f8"
+secondaryBackgroundColor = "#ffffff"
+textColor = "#17212b"
+font = "sans serif"
+
+[server]
+headless = true

app.py

@@ -15,7 +15,9 @@
 from proofstack.parser import parse_model_output
 from proofstack.providers import get_provider
 from proofstack.review_state import assessment_matches_case, reviewer_submission_errors
-from proofstack.storage import save_decision
+from proofstack.storage import load_decisions, save_decision
+
+st.set_page_config(page_title="ProofStack", page_icon="🧾", layout="wide")
 
 
 DECISION_LABELS = {
@@ -36,29 +38,54 @@ def load_manifest():
 questions = load_questions()
 manifest = load_manifest()
 q_by_id = {q["question_id"]: q for q in questions}
+try:
+    saved_decisions = load_decisions()
+    storage_read_error = ""
+except Exception as exc:
+    saved_decisions = []
+    storage_read_error = str(exc)
+latest_decision = {row.get("evidence_id"): row for row in saved_decisions}
 
-st.set_page_config(page_title="ProofStack", page_icon="🧾", layout="wide")
 st.title("ProofStack")
 st.caption(
     "Review a security claim against the evidence submitted for it. "
     "The AI gives a recommendation; the reviewer makes the decision."
 )
 
 with st.sidebar:
-    st.header("Reviewer checklist")
-    st.write("1. Read the claim.")
-    st.write("2. Inspect the evidence.")
-    st.write("3. Run the AI assessment.")
-    st.write("4. Verify the cited excerpt.")
-    st.write("5. Record your decision.")
+    st.header("Review queue")
+    st.caption("Claim → evidence → advisory → your decision")
+    done_docs = len(set(latest_decision) & {row["evidence_id"] for row in manifest})
+    st.metric("Documents decided", f"{done_docs} / {len(manifest)}")
+    status_filter = st.radio("Show claims", ["All", "Needs review", "Completed"], horizontal=True)
     st.divider()
     st.caption(
         "Synthetic evidence only. ProofStack is a review aid and does not certify compliance."
     )
 
+question_ids = [q["question_id"] for q in questions]
+if storage_read_error:
+    st.sidebar.warning(f"Could not load decision history: {storage_read_error}")
+if status_filter != "All":
+    question_ids = [qid for qid in question_ids if (lambda docs: bool(docs) and (all(doc["evidence_id"] in latest_decision for doc in docs) == (status_filter == "Completed")))([row for row in manifest if row["question_id"] == qid])]
+if not question_ids:
+    question_ids = [q["question_id"] for q in questions]
+
+with st.expander("Decision history", expanded=False):
+    st.subheader("Reviewer decision history")
+    st.caption("Each entry preserves the reviewer label, advisory context, and evidence reference at the time of review.")
+    if saved_decisions:
+        st.dataframe(
+            [{"Recorded (UTC)": r.get("timestamp_utc", ""), "Claim": r.get("question_id", ""), "Document": r.get("evidence_title", r.get("evidence_id", "")), "Reviewer decision": r.get("reviewer_label", ""), "Advisory": r.get("model_label", "—") or "—", "Reviewer": r.get("reviewer", ""), "Notes": r.get("notes", "")} for r in reversed(saved_decisions)],
+            width="stretch",
+            hide_index=True,
+        )
+    else:
+        st.info("No reviewer decisions have been recorded yet. Saved decisions will appear here.")
+
 q_id = st.selectbox(
     "Security question",
-    [q["question_id"] for q in questions],
+    question_ids,
     format_func=lambda value: (
         f'{value} — '
         f'{q_by_id[value].get("reviewer_friendly_question") or q_by_id[value]["question"]}'
@@ -78,6 +105,8 @@ def load_manifest():
     st.write(question.get("reviewer_hint", ""))
 
 mapped = [item for item in manifest if item["question_id"] == q_id]
+reviewed_for_claim = sum(item["evidence_id"] in latest_decision for item in mapped)
+st.caption(f"{reviewed_for_claim}/{len(mapped)} documents decided · review each document against the claim")
 e_id = st.selectbox(
     "Submitted evidence",
     [item["evidence_id"] for item in mapped],
@@ -89,6 +118,9 @@ def load_manifest():
 )
 
 evidence = next(item for item in mapped if item["evidence_id"] == e_id)
+previous_decision = latest_decision.get(e_id)
+if previous_decision:
+    st.info(f"Previous decision: {previous_decision.get('reviewer_label','')} by {previous_decision.get('reviewer','reviewer')} · {previous_decision.get('timestamp_utc','')}. Recording again appends a revision.")
 evidence_path = ROOT / evidence["path"]
 evidence_text = evidence_path.read_text(encoding="utf-8")
 
@@ -120,10 +152,15 @@ def load_manifest():
     st.text(evidence_text)
 
 with right:
-    st.subheader("AI recommendation")
-    st.caption("Advisory only. Verify the evidence and citation before deciding.")
+    st.subheader("Model advisory")
+    st.caption("Advisory only · reviewer makes the decision")
+    active_provider = os.getenv("PROOFSTACK_PROVIDER", "mock").strip().lower()
+    if active_provider == "mock":
+        st.info("Source: Local mock · deterministic heuristic, not an AI model")
+    else:
+        st.info(f"Source: configured AI provider · {active_provider}")
 
-    if st.button("Assess this evidence", type="primary", use_container_width=True):
+    if st.button("Assess this evidence", type="primary", width="stretch"):
         try:
             provider = get_provider()
             raw = provider.assess(question, evidence, evidence_text)
@@ -189,9 +226,9 @@ def load_manifest():
         st.caption("Run the assessment when you are ready to compare the claim with this evidence.")
 
 st.divider()
-st.subheader("Reviewer decision")
+st.subheader("Your decision for this document")
 st.caption(
-    "Your decision is authoritative. Use the model as a second opinion, not as an approval engine."
+    "Decide whether this document supports the claim. Revisions add a new record; earlier decisions remain in the log."
 )
 
 reviewer = st.text_input("Reviewer name")
@@ -211,7 +248,7 @@ def load_manifest():
     if reviewer_label != assessment["label"]:
         st.info("Your decision differs from the AI recommendation. Explain why in reviewer notes.")
 
-if st.button("Record final decision", use_container_width=True):
+if st.button("Record reviewer decision", width="stretch"):
     errors = reviewer_submission_errors(
         reviewer,
         reviewer_label,
@@ -241,9 +278,17 @@ def load_manifest():
 
         try:
             record_id = save_decision(row)
+            st.session_state["last_saved_case"] = f"{q_id}/{e_id}"
+            st.session_state["last_saved_at"] = row["timestamp_utc"]
             if record_id:
-                st.success(f"Decision recorded. Record ID: {record_id}")
+                st.success(f"Decision saved to the review log · Record {record_id}")
             else:
-                st.success("Decision recorded.")
+                st.success("Decision saved to the review log. Any later decision is recorded as a new entry.")
         except Exception as exc:
             st.error(str(exc))
+
+with st.expander("How to review this case", expanded=not st.session_state.get("review_guide_seen", False)):
+    st.markdown("**Start with the claim.** Check its scope, systems, people, and time period.")
+    st.markdown("**Inspect what the document demonstrates.** A policy describes intent; an operating record shows an event or current state.")
+    st.markdown("**Verify the citation, then decide.** A grounded excerpt can still be insufficient for the claim. Your recorded decision is authoritative.")
+    st.session_state["review_guide_seen"] = True

data/questions.json

@@ -1,4 +1,3 @@
-
 [
   {
     "question_id": "Q01",
@@ -71,5 +70,3 @@
     "reviewer_hint": "Examine active third-party administrative access records. Verify documented approval, enforced MFA, and a specific access expiration date and time for each grant. All three security requirements must be supported by evidence."
   }
 ]
-
-]

src/proofstack/storage/__init__.py

@@ -9,3 +9,14 @@ def save_decision(row: dict):
     else:
         raise ValueError(f"Unknown PROOFSTACK_STORAGE: {mode}")
     return append_decision(row)
+
+
+def load_decisions():
+    mode = os.getenv("PROOFSTACK_STORAGE", "local").lower()
+    if mode == "local":
+        from proofstack.storage.local import read_decisions
+    elif mode == "firestore":
+        from proofstack.storage.firestore import read_decisions
+    else:
+        raise ValueError(f"Unknown PROOFSTACK_STORAGE: {mode}")
+    return read_decisions()

src/proofstack/storage/firestore.py

@@ -16,3 +16,16 @@ def append_decision(row: dict):
 
     _, document_ref = client.collection(collection).add(record)
     return document_ref.id
+
+
+def read_decisions():
+    """Return reviewer records newest first for the in-app audit view."""
+    project = os.getenv("GCP_PROJECT_ID") or None
+    collection = os.getenv("FIRESTORE_COLLECTION", "proofstack_decisions").strip()
+    if not collection:
+        raise RuntimeError("FIRESTORE_COLLECTION must not be empty")
+    client = firestore.Client(project=project)
+    return [
+        {**snapshot.to_dict(), "record_id": snapshot.id}
+        for snapshot in client.collection(collection).order_by("timestamp_utc", direction=firestore.Query.DESCENDING).stream()
+    ]

src/proofstack/storage/local.py

@@ -38,3 +38,14 @@ def append_decision(row: dict, path="local_decisions.csv"):
         if new:
             writer.writeheader()
         writer.writerow({key: row.get(key, "") for key in HEADER})
+
+
+def read_decisions(path="local_decisions.csv"):
+    target = Path(path)
+    if not target.exists():
+        return []
+    with target.open("r", newline="", encoding="utf-8") as f:
+        reader = csv.DictReader(f)
+        if reader.fieldnames != HEADER:
+            raise RuntimeError("Existing local decision log uses an older schema.")
+        return list(reader)