UI: pin verdict at top with placeholder; auto-delete source+intermediates after audit (opt-out checkbox)

legal_doc_redteam/zerogpu_gui.py

@@ -118,6 +118,19 @@ def bind_vlm_fn(vlm_fn: Callable[[Any, str], str], *, model_id: str) -> None:
     _BOUND_VLM_MODEL_ID = model_id
 
 
+_VERDICT_PLACEHOLDER = (
+    "## Upload a document and click **Audit document** to begin.\n\n"
+    "_The verdict will appear right here as soon as the audit finishes._"
+)
+
+_CLEANUP_NOTICE = (
+    "\n\n---\n"
+    "_🔒 Source file and intermediate artefacts (rendered page images, "
+    "converted PDFs, temporary JSON) have been deleted from the server. "
+    "Only this verdict markdown and the download below remain._"
+)
+
+
 def run_full_audit(
     file_path: str | None,
     dpi: int | float,
@@ -129,11 +142,12 @@ def run_full_audit(
     reviewer_model_id: str,
     reasoning_effort: str,
     hf_token: str,
+    cleanup_after_audit: bool = True,
     progress: gr.Progress = gr.Progress(track_tqdm=False),
 ) -> tuple[str, list[list[str]], list[list[str | int | float]], dict[str, Any], str | None]:
     if not file_path:
         return (
-            "Upload a PDF, DOCX, HTML, Markdown, or text file to begin.",
+            _VERDICT_PLACEHOLDER,
             [],
             [],
             {},
@@ -223,18 +237,60 @@ def run_full_audit(
         "countermeasures": audit_report,
         "ocr_integrity": ocr_report,
     }
-    md_path = work_dir / "integrity_verdict.md"
-    md_path.write_text(render_markdown(summary), encoding="utf-8")
+    verdict_text = render_markdown(summary)
+
+    if cleanup_after_audit:
+        download_path = _persist_verdict_for_download(verdict_text)
+        _scrub_audit_artifacts(work_dir=work_dir, source_path=source)
+        verdict_text = verdict_text + _CLEANUP_NOTICE
+        verdict_path: str | None = str(download_path) if download_path else None
+    else:
+        md_path = work_dir / "integrity_verdict.md"
+        md_path.write_text(verdict_text, encoding="utf-8")
+        verdict_path = str(md_path)
 
     return (
-        render_markdown(summary),
+        verdict_text,
         controls_as_rows(audit_report),
         report_table_rows(ocr_report),
         combined_report,
-        str(md_path),
+        verdict_path,
     )
 
 
+def _downloads_dir() -> Path:
+    base = _runs_base_dir() / "downloads"
+    base.mkdir(parents=True, exist_ok=True)
+    return base
+
+
+def _persist_verdict_for_download(verdict_text: str) -> Path | None:
+    """Write the verdict markdown to a separate downloads dir so the work
+    dir can be wiped while the Download File component still resolves."""
+
+    try:
+        out_path = _downloads_dir() / f"verdict_{uuid.uuid4().hex[:10]}.md"
+        out_path.write_text(verdict_text, encoding="utf-8")
+        return out_path
+    except OSError:
+        return None
+
+
+def _scrub_audit_artifacts(*, work_dir: Path, source_path: Path) -> None:
+    """Best-effort delete of the per-run work dir and the Gradio upload."""
+
+    try:
+        if work_dir.exists():
+            shutil.rmtree(work_dir, ignore_errors=True)
+    except Exception:
+        pass
+    try:
+        if source_path.exists():
+            Path(source_path).unlink(missing_ok=True)
+    except Exception:
+        pass
+
+
 def _runs_base_dir() -> Path:
     override = os.environ.get("LEGAL_DOC_REDTEAM_RUNS_DIR")
     if override:
@@ -256,9 +312,9 @@ def _allocate_work_dir() -> Path:
 
 
 def cleanup_old_runs(retention_hours: int = DEFAULT_RUN_RETENTION_HOURS) -> int:
-    """Prune per-run audit directories older than ``retention_hours``.
+    """Prune per-run audit directories and stale verdict downloads.
 
-    Returns the number of directories removed. Intended for startup, so a
+    Returns the number of items removed. Intended for startup, so a
     long-running public Space does not accrete data forever.
     """
 
@@ -268,14 +324,22 @@ def cleanup_old_runs(retention_hours: int = DEFAULT_RUN_RETENTION_HOURS) -> int:
     cutoff = time.time() - max(0, retention_hours) * 3600
     removed = 0
     for entry in base.iterdir():
-        if not entry.is_dir() or not entry.name.startswith("audit_"):
-            continue
         try:
-            if entry.stat().st_mtime < cutoff:
-                shutil.rmtree(entry, ignore_errors=True)
-                removed += 1
+            if entry.is_dir() and entry.name.startswith("audit_"):
+                if entry.stat().st_mtime < cutoff:
+                    shutil.rmtree(entry, ignore_errors=True)
+                    removed += 1
         except OSError:
             continue
+    downloads = base / "downloads"
+    if downloads.exists():
+        for entry in downloads.iterdir():
+            try:
+                if entry.is_file() and entry.stat().st_mtime < cutoff:
+                    entry.unlink(missing_ok=True)
+                    removed += 1
+            except OSError:
+                continue
     return removed
 
 
@@ -336,6 +400,15 @@ def build_app(
             f"_Public Space safeguards: uploads are capped at {DEFAULT_MAX_UPLOAD_MB} MB, "
             f"per-run audit data is pruned after {DEFAULT_RUN_RETENTION_HOURS} h._"
         )
+
+        # ----------------------------------------------------------------
+        # B1: verdict pinned at the top, gets populated in place after Audit
+        # ----------------------------------------------------------------
+        verdict_md = gr.Markdown(value=_VERDICT_PLACEHOLDER)
+
+        # ----------------------------------------------------------------
+        # Inputs row (always visible)
+        # ----------------------------------------------------------------
         with gr.Row():
             with gr.Column(scale=2):
                 file_in = gr.File(
@@ -354,12 +427,19 @@ def build_app(
                         ".text",
                     ],
                 )
-                run_btn = gr.Button("Audit document", variant="primary")
+                run_btn = gr.Button("Audit document", variant="primary", size="lg")
+                cleanup_after = gr.Checkbox(
+                    value=True,
+                    label="Delete uploaded file and intermediate artefacts after the report is generated (recommended)",
+                )
             with gr.Column(scale=1):
                 dpi = gr.Number(value=180, precision=0, label="Render DPI")
                 max_pages = gr.Number(value=8, precision=0, label="Max pages")
 
-        with gr.Accordion("CPU OCR engines", open=True):
+        # ----------------------------------------------------------------
+        # Configuration accordions — collapsed by default per B1
+        # ----------------------------------------------------------------
+        with gr.Accordion("CPU OCR engines", open=False):
             cpu_engines = gr.CheckboxGroup(
                 choices=CPU_OCR_ENGINES,
                 value=cpu_defaults,
@@ -370,7 +450,7 @@ def build_app(
                 "`tesseract` system binary; the Space includes it via `packages.txt`._"
             )
 
-        with gr.Accordion("Vision LLM OCR (GPU)", open=True):
+        with gr.Accordion("Vision LLM OCR (GPU)", open=False):
             vlm_backend = gr.Radio(
                 choices=["none", "local_transformers", "hf_inference"],
                 value=default_vlm_backend,
@@ -388,7 +468,7 @@ def build_app(
                 "handles hard PDFs better. `PaddlePaddle/PaddleOCR-VL` is the most compact._"
             )
 
-        with gr.Accordion("Reasoning verdict", open=True):
+        with gr.Accordion("Reasoning verdict", open=False):
             reviewer = gr.Radio(
                 choices=["deterministic", "local_transformers", "hf_inference"],
                 value=default_reviewer_backend,
@@ -404,14 +484,17 @@ def build_app(
                 label="Reasoning effort",
             )
 
-        hf_token = gr.Textbox(
-            value="",
-            label="HF token (optional, used for hf_inference backends)",
-            type="password",
-            visible=expose_hf_token,
-        )
+        with gr.Accordion("Optional HF token (only for hf_inference)", open=False):
+            hf_token = gr.Textbox(
+                value="",
+                label="HF token (leave empty to use the Space's own HF_TOKEN secret)",
+                type="password",
+                visible=expose_hf_token,
+            )
 
-        verdict_md = gr.Markdown()
+        # ----------------------------------------------------------------
+        # Detail accordions — collapsed by default
+        # ----------------------------------------------------------------
         with gr.Accordion("Countermeasures detector matrix", open=False):
             audit_table = gr.Dataframe(
                 headers=CONTROL_HEADERS,
@@ -426,7 +509,7 @@ def build_app(
             )
         with gr.Accordion("Full JSON report", open=False):
             json_out = gr.JSON()
-        verdict_file = gr.File(label="Verdict markdown", interactive=False)
+        verdict_file = gr.File(label="Verdict markdown (download)", interactive=False)
 
         run_btn.click(
             run_full_audit,
@@ -441,6 +524,7 @@ def build_app(
                 reasoning_model,
                 reasoning_effort,
                 hf_token,
+                cleanup_after,
             ],
             outputs=[verdict_md, audit_table, ocr_table, json_out, verdict_file],
         )