"""Run with 'streamlit run app.py'.""" from __future__ import annotations import base64 import json from pathlib import Path import streamlit as st from frontierlab.demo_data import QUESTION, QUERY from frontierlab.quantitative import inverse_variance_heterogeneity from frontierlab.openweight import OpenWeightClient from frontierlab.pipeline import make_plan, run_research, continue_after_review from frontierlab.planning_readiness import build_planning_readiness, readiness_summary from frontierlab.review import review_packet, save_review from frontierlab.reporting import export_json, summary from frontierlab.process_metrics import process_metrics from frontierlab.operating_status import research_operating_status from frontierlab.trajectory_graph import build_trajectory_graph, summarize as trajectory_summary, to_json as trajectory_json from frontierlab.interventions import record_human_event from frontierlab.claim_benchmark import benchmark from frontierlab.claim_verification import audit_predictions, conflict_transparency from frontierlab.integrity_trials import suite_report from frontierlab.evaluation_ui import render_evaluation from frontierlab.evaluation import require_evaluation from frontierlab.evidence_graph import (approve_plan_step, refresh_evidence_graph, step_gate, summarize as graph_summary, to_json as graph_json, to_svg as graph_svg) from frontierlab.literature_pilot import load_dataset, reviewer_packet from frontierlab.recorded_gate_ablation import run_recorded_pilot from frontierlab.pilot_review import canonicalize_review_upload, labels_from_review_packet from frontierlab.research_backend import NvidiaResearchClient, research_enabled, remaining_calls from frontierlab import hf_providers from frontierlab.hf_providers import HFProvidersUnavailable from frontierlab.recorded_replay import build_recorded_replay from frontierlab.evidence import europepmc_error_message from frontierlab.version import __version__ ASSETS = Path(__file__).resolve().parent / "docs" / "assets" _icon_path = ASSETS / "flatland-sphere-icon.png" st.set_page_config(page_title="FrontierLab | Research Intelligence", page_icon=str(_icon_path) if _icon_path.exists() else "◯", layout="wide") # Presentation only: black-and-white "Flatland book plate" styling (see docs/assets/). SERIF = "Georgia, 'Times New Roman', Times, serif" st.markdown(f""" """, unsafe_allow_html=True) _ICON = 'width="22" height="22" viewBox="0 0 24 24" fill="none" stroke="#F2F2F2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"' BAND_ICONS = { "point": f'', "sphere": f'', "cube": f'', } def band(label: str, icon: str) -> None: """Small-caps section band, as on the README architecture plate.""" st.markdown(f'
{BAND_ICONS[icon]}{label}
', unsafe_allow_html=True) def _pct(value) -> str: return "N/A" if value is None else f"{value * 100:.0f}%" def render_operating_status(roi: dict, *, key: str, download_label: str, file_name: str, synthetic_note: str | None = None) -> None: """Research Operating Interface body; shared by live runs and the recorded replay.""" trace = roi["traceability"] source_cov = trace["source_trace_coverage"] strict_cov = trace["strict_extracted_coverage"] o1, o2, o3, o4 = st.columns(4) o1.metric("Next required stage", roi["current_activity"]["next_required_stage"].title()) o2.metric("Source trace coverage", _pct(source_cov)) o3.metric("Strict extracted coverage", _pct(strict_cov)) o4.metric("Auto repair", "OFF") if synthetic_note: st.caption(synthetic_note) st.caption( "Trace Coverage is measured from the current Evidence Graph. " "It measures provenance coverage, not scientific accuracy or semantic truth." ) if source_cov is not None and strict_cov is not None and strict_cov < source_cov: st.info( "Some artifacts can be traced to a source but do not have an all-EXTRACTED path. " "FrontierLab preserves those provenance disagreements instead of promoting them to strict evidence." ) blocked_rows = [] for artifact in trace["artifacts"]: if artifact["strict_extracted"]: continue if artifact["blocking_edges"]: for edge in artifact["blocking_edges"]: blocked_rows.append({ "Artifact": artifact["artifact_id"], "Type": artifact["artifact_type"], "Blocking provenance": edge["provenance"], "Reason": edge["reason"], }) elif artifact["unresolved_evidence_ids"] or artifact["dead_ends"]: blocked_rows.append({ "Artifact": artifact["artifact_id"], "Type": artifact["artifact_type"], "Blocking provenance": "UNRESOLVED", "Reason": ( f"unknown sources={artifact['unresolved_evidence_ids']} · " f"dead ends={artifact['dead_ends']}" ), }) if blocked_rows: st.dataframe(blocked_rows, hide_index=True, width="stretch") a1, a2, a3, a4 = st.columns(4) audit_state = roi["source_audit"] a1.metric("Active papers", audit_state["papers"]) a2.metric("Anchored passed claims", audit_state["anchored_passed_claims"]) a3.metric( "Rejected claims", audit_state["rejected_claims_attached"] + audit_state["rejected_claims_unattached"], ) a4.metric("Human events", roi["governance"]["human_events_recorded"]) st.dataframe( [ { "Control area": "Provenance", "Status": "MEASURED", "Detail": ( f"{trace['source_traceable_artifacts']} / {trace['material_artifacts']} " "in-scope artifacts reach a known paper" ), }, { "Control area": "Governance", "Status": "HUMAN-GATED", "Detail": ( f"Claim review: {roi['governance']['claim_review_status']} · " f"Method evaluation: {'saved' if roi['governance']['method_evaluation_saved'] else 'not saved'} · " f"Planning authorization recorded: " f"{'yes' if roi['governance']['planning_authorized_recorded'] else 'no'}" ), }, { "Control area": "Maintainability", "Status": roi["maintainability"]["status"], "Detail": "Scheduled dependency canary and advisory repair proposals are active; proposals cannot edit code/dependencies, commit, merge or deploy. Autonomous repair and environment rebuild are not implemented.", }, ], hide_index=True, width="stretch", ) if roi["active_sources"]: st.caption( "Active sources: " + ", ".join(source["paper_id"] for source in roi["active_sources"]) ) st.download_button( download_label, data=json.dumps(roi, ensure_ascii=False, indent=2, sort_keys=True), file_name=file_name, mime="application/json", key=f"{key}_roi_download", ) def render_planning_readiness(readiness: dict, ready_stats: dict, *, key: str, download_label: str, file_name: str) -> None: """Planning-readiness body; shared by live runs and the recorded replay.""" st.warning( "READY ≠ APPROVED. This packet validates the current reviewed context and " "selected method, but it does not authorize or call the Planning Agent." ) p1, p2, p3, p4 = st.columns(4) p1.metric("Selected method", ready_stats["selected_candidate_id"]) p2.metric("Open reproducibility checks", len(ready_stats["open_reproducibility_checks"])) p3.metric("Required planning obligations", ready_stats["required_obligations"]) p4.metric("Planning approved", "YES" if ready_stats["planning_approved"] else "NO") st.caption( "Open checks: " + ", ".join(ready_stats["open_reproducibility_checks"]) + ". The future plan must preserve evidence/citation gates and address all obligations." ) st.dataframe( [{ "Obligation": item["obligation_id"], "Category": item["category"], "Status": item["status"], "Requirement": item["requirement"], } for item in readiness["planning_obligations"]], hide_index=True, width="stretch", ) st.download_button( download_label, data=json.dumps(readiness, ensure_ascii=False, indent=2, sort_keys=True), file_name=file_name, mime="application/json", key=f"{key}_readiness_download", ) def render_graph_overview(graph: dict, g_stats: dict) -> None: """Evidence-graph legend, edge counts, drawing and surfaced links (read-only).""" st.caption("EXTRACTED: passage located in the cited abstract and source IDs checked (not semantic entailment). " "INFERRED: model stance without an anchored quote. AMBIGUOUS: conflicting labels. " "Only all-EXTRACTED paths to a paper count as evidence.") e_count = g_stats["edges_by_provenance"] x, y, z = st.columns(3) x.metric("EXTRACTED edges", e_count["EXTRACTED"]) y.metric("INFERRED edges (not evidence)", e_count["INFERRED"]) z.metric("AMBIGUOUS edges (not evidence)", e_count["AMBIGUOUS"]) drawing = graph_svg(graph) if drawing: st.markdown('Evidence provenance graph: solid lines EXTRACTED, dashed INFERRED, dotted AMBIGUOUS', unsafe_allow_html=True) else: st.caption("Graph too large to draw legibly; use the tables and graph.json below.") if g_stats["surfaced_for_review"]: st.warning("Surfaced for researcher review. These links are NOT counted as evidence.") st.dataframe([{"Provenance": e["provenance"], "From": e["source"], "Relation": e["relation"], "To": e["target"], "Why": e["reason"]} for e in g_stats["surfaced_for_review"]], hide_index=True, width="stretch") def render_trajectory(trajectory: dict, t_stats: dict, *, key: str, model_calls_label: str, download_label: str, file_name: str) -> None: """Trajectory audit graph body; shared by live runs and the recorded replay.""" st.caption( "Audit-only process graph. It records model calls, human interventions, " "method evaluation, rejected claims and process signals. It never participates " "in the scientific evidence gate." ) t1, t2, t3, t4 = st.columns(4) t1.metric(model_calls_label, t_stats["model_calls"]) t2.metric("Human events", t_stats["human_events"]) t3.metric("Rejected claims", t_stats["rejected_claims"]) t4.metric("Process signals", t_stats["process_signals"]) st.dataframe( [{"Node type": k, "Count": v} for k, v in t_stats["nodes_by_type"].items()], hide_index=True, width="stretch" ) st.download_button( download_label, data=trajectory_json(trajectory), file_name=file_name, mime="application/json", key=f"{key}_trajectory_download", ) _hero_path = ASSETS / "flatland-hero.svg" if _hero_path.exists(): _hero = base64.b64encode(_hero_path.read_bytes()).decode("ascii") st.markdown(f'', unsafe_allow_html=True) st.caption("META GLOBAL AI DEVELOPER HACKATHON 2026 / INDIE DEVELOPER / PUSH THE FRONTIER") st.title("FrontierLab / Research Intelligence") st.caption(f"Research → Evidence → Hypothesis → Evaluation → Planning · v{__version__}") deployment_path = Path(__file__).resolve().parent / "deployment.json" on_public_space = deployment_path.exists() if on_public_space: deployment = json.loads(deployment_path.read_text(encoding="utf-8")) st.caption(f"Deployed source: {deployment['source_sha'][:12]} · {deployment['version']} · {deployment['space_id']}") st.write("An evidence-grounded scientific research workflow that **records disagreement instead of hiding it**.") # Recorded real M2 core (2026-10-28): recomputed from docs/runs/*.json on every render. # Zero model calls and zero network requests; see frontierlab/recorded_replay.py. try: replay = build_recorded_replay() replay_error = None except (OSError, ValueError, KeyError, TypeError, PermissionError) as exc: replay, replay_error = None, exc REPLAY_TITLE = "Replay recorded real M2 case (no model call)" SYNTHETIC_BANNER = ( "SYNTHETIC / FICTIONAL MODE. Every publication, effect size, benchmark case, failure injection, " "method strategy and plan in this mode is invented for testing. None of it is real scientific " "evidence or a literature citation, and no model or laboratory experiment runs. It exists to show " "the gates failing closed. For real recorded results, see “" + REPLAY_TITLE + "” at the top." ) with st.container(border=True, key="judge_path"): st.markdown( "**Judge path · start here (about 3 minutes)**\n\n" f"1. **Read the recorded real case.** “{REPLAY_TITLE}” is open just below: two published " "abstracts, real Meta Muse Glimmer outputs recorded on 2026-10-28, researcher-reviewed. " "Every number is recomputed from the recorded files; nothing calls a model.\n" "2. **Run the gates yourself.** Keep *Synthetic audit demonstration* selected in the sidebar and " "click **Run research audit**. The data is fictional and the gates fail closed by design.\n" "3. **Inspect provenance.** In either path, open *Research Operating Interface*, *Evidence graph* " "and *Trajectory*, and download the JSON receipts." ) with st.expander("Start here · " + REPLAY_TITLE, expanded=True): if replay is None: st.error(f"Recorded replay unavailable: {type(replay_error).__name__}: {replay_error}") else: r_calls = replay["recorded_calls"] r_run = replay["run"] r_sel = replay["selection"] r_trace = replay["trace_coverage"] r_ready = replay["readiness_summary"] st.info( f"RECORDED REAL RUN · REPLAY ONLY. Recorded on {r_calls['recorded_date_taipei']} (Asia/Taipei): " f"{r_calls['extraction_calls']} Evidence Extraction calls and {r_calls['continuation_calls']} " f"reviewed-continuation calls ({' → '.join(r_calls['continuation_roles'])}) to {r_calls['model']} " f"via {r_calls['provider']} on {len(r_run.papers)} published abstracts; claims reviewed and method " f"{r_sel['selected_candidate_id']} selected by {r_sel['reviewer']}. This panel recomputes every number " f"from docs/runs/*.json with the same code used for live runs and makes " f"{replay['new_model_calls']} model calls. Coverage measures provenance, not scientific accuracy. " "Planning is NOT approved." ) k1, k2, k3, k4 = st.columns(4) k1.metric("Source trace coverage", _pct(r_trace["source_trace_coverage"])) k2.metric("Strict EXTRACTED coverage", _pct(r_trace["strict_extracted_coverage"])) k3.metric("Planning obligations", r_ready["required_obligations"]) k4.metric("Model calls in this replay", replay["new_model_calls"]) st.caption( f"{r_trace['source_traceable_artifacts']} / {r_trace['material_artifacts']} artifacts reach a known paper; " f"{r_trace['strict_extracted_artifacts']} / {r_trace['material_artifacts']} have an all-EXTRACTED path. " "Sources: " + ", ".join(p.paper_id for p in r_run.papers) + ". " f"Recorded runs: [extraction]({r_calls['extraction_run_url']}) · " f"[Skeptic → Hypothesis]({r_calls['continuation_run_url']})." ) rt1, rt2, rt3, rt4, rt5 = st.tabs(["Research Operating Interface", "Planning readiness", "Evidence graph", "Trajectory", "Skeptic → Hypothesis"]) with rt1: render_operating_status(replay["operating_status"], key="replay", download_label="Download recorded operating-status JSON", file_name="research_operating_status_recorded_20261028.json") with rt2: render_planning_readiness(replay["planning_readiness"], r_ready, key="replay", download_label="Download recorded planning-readiness JSON", file_name="planning_readiness_recorded_20261028.json") with rt3: render_graph_overview(replay["evidence_graph"], replay["graph_summary"]) st.caption("No plan exists in the recorded case: planning approval was never granted.") st.download_button("Download recorded graph.json", data=graph_json(replay["evidence_graph"]), file_name="graph_recorded_20261028.json", mime="application/json", key="replay_graph_download") with rt4: render_trajectory(replay["trajectory"], replay["trajectory_summary"], key="replay", model_calls_label="Model calls in recorded receipt", download_label="Download recorded trajectory.json", file_name="trajectory_recorded_20261028.json") st.caption( "The trajectory is rebuilt from the extraction receipt (recorded 2026-10-28). The " f"{r_calls['continuation_calls']} continuation calls are preserved separately in " "docs/runs/nim_reviewed_continuation_20261028.json. No call is made during replay." ) with rt5: st.caption("Recorded verbatim on 2026-10-28 after researcher claim review. Not independently " "verified; scientific accuracy is not measured.") st.markdown("**Adversarial Skeptic Agent (recorded)**") st.json(r_run.skeptic, expanded=False) st.markdown("**Hypothesis Agent (recorded)**") st.json(r_run.hypotheses, expanded=False) LOCAL_MODE = "Live Europe PMC + local Meta model" NIM_MODE = "Live Europe PMC + Meta via NVIDIA" HF_MODE = hf_providers.MODE_LABEL MODE_LABELS = ["Synthetic audit demonstration", "Live Europe PMC metadata (no model calls)", LOCAL_MODE, NIM_MODE, HF_MODE] hf_available, hf_reason = hf_providers.live_status() def _mode_display(label: str) -> str: if label == LOCAL_MODE and on_public_space: return label + " · not available on the public Space" if label == NIM_MODE and not research_enabled(): return label + " · not available on the public Space (backend not enabled)" if label == HF_MODE and not hf_available: return label + " · unavailable (no Hugging Face token configured)" return label nim_authorized = False hf_consent = False hf_session_budget = st.session_state.setdefault("hf_session_budget", hf_providers.SessionBudget()) with st.sidebar: st.header("Engine") mode_label = st.radio("Evidence mode", MODE_LABELS, index=0, format_func=_mode_display) mode = "demo" if mode_label.startswith("Synthetic") else "live" use_local_meta = mode_label == LOCAL_MODE use_nim = mode_label == NIM_MODE use_hf = mode_label == HF_MODE st.info( "Live Meta Muse calls are available through Hugging Face Inference Providers when the app " "owner has configured a token, with per-session and daily call caps. The direct billed " "Meta Model API adapter remains off." ) if mode == "demo": st.caption("Synthetic mode: invented publications only (see the banner in the main panel).") elif use_nim: if not research_enabled(): st.warning("Not available on the public Space: NVIDIA research inference has not been enabled by the app owner.") nim_authorized = st.checkbox( "Allow this research question and retrieved abstracts to be sent to NVIDIA for Meta inference", value=False, disabled=not research_enabled(), key="nim_research_consent") st.caption(f"Remaining shared model calls in this server process: {remaining_calls()} / 8. Failed requests consume the allowance; no automatic retries.") elif use_hf: if not hf_available: st.warning("Live Meta Muse is unavailable: " + hf_reason + ".") hf_consent = st.checkbox(hf_providers.CONSENT_LABEL, value=False, disabled=not hf_available, key="hf_live_consent") st.caption( f"Primary: {hf_providers.PRIMARY_MODEL} · automatic fallback: {hf_providers.FALLBACK_MODEL}. " f"Session calls left: {hf_session_budget.remaining} / {hf_session_budget.limit} · " f"server calls left today: {hf_providers.daily_budget().remaining} / {hf_providers.daily_budget().limit}. " "Every attempt, including a fallback, counts. Both models reason before answering, so " "one call can take up to 90 s (fallback 75 s). At most " f"{hf_providers.MAX_LIVE_ABSTRACTS} abstracts, each truncated to " f"{hf_providers.MAX_ABSTRACT_CHARS} characters. Quotes still pass the verbatim evidence gate " "and every step stays human-gated." ) elif use_local_meta: if on_public_space: st.warning("Not available on the public Space: this mode needs a Muse Glimmer checkpoint served on the same machine. A run here fails safely without any model call.") else: st.warning("Experimental real open-weight model mode. Requires an actual Muse Glimmer checkpoint served on THIS machine; charges no Meta API credits. Never expose the local server publicly.") else: st.info("Europe PMC metadata only. No model inference, no paid API requests.") st.caption("No autonomous wet-lab execution, clinical advice or unverified study claims.") band("I · Flatland — question and evidence", "point") st.subheader("1 · Research question") if mode == "demo": st.warning(SYNTHETIC_BANNER) with st.form("research_form"): question = st.text_area("Scientific research question", value=QUESTION, height=85) query = st.text_input("Europe PMC search query", value=QUERY, disabled=(mode == "demo")) capped = use_nim or use_hf n = st.slider("Maximum live abstracts", 1, 2 if capped else 10, 2 if capped else 4, disabled=(mode == "demo")) hf_blocked = use_hf and not (hf_available and hf_consent and hf_session_budget.remaining > 0) run_clicked = st.form_submit_button("Run research audit", width="stretch", type="primary", disabled=(use_nim and not nim_authorized) or hf_blocked) if run_clicked: with st.spinner("Retrieving evidence and applying verification gates..."): try: st.session_state.pop("frontier_result", None) st.session_state.pop("frontier_model_client", None) client = None client = (NvidiaResearchClient(authorized=nim_authorized) if use_nim else hf_providers.HFMuseResearchClient(authorized=hf_consent, session_budget=hf_session_budget) if use_hf else OpenWeightClient() if use_local_meta else None) st.session_state["frontier_model_client"] = client result = run_research(question, query, mode=mode, model_client=client, max_papers=n) st.session_state["frontier_result"] = result except HFProvidersUnavailable as exc: st.error(str(exc)) except Exception as exc: friendly = europepmc_error_message(exc) st.error(friendly if friendly else f"Research run failed: {type(exc).__name__}: {exc}") if use_hf and client is not None and "frontier_result" not in st.session_state: for line in hf_providers.failure_summary(getattr(client, "provenance", None)): st.caption("Failed attempt · " + line) with st.expander("Live Meta Muse call receipts (failed run)", expanded=True): st.json(getattr(client, "provenance", {})) with st.expander("Published-literature pilot · 12 cases awaiting researcher annotation"): pilot = load_dataset() st.caption("Four published antibody-engineering papers; short abstract excerpts only. Draft reference labels are excluded from the review download. Researcher annotations remain pending.") recorded_path = Path(__file__).resolve().parent / "docs/runs/meta_nim_pilot_20261028.json" if recorded_path.exists(): recorded = json.loads(recorded_path.read_text(encoding="utf-8"))["receipt"] st.info("RECORDED REAL META RUN · 12 structured predictions, 8 matched quotations and 4 insufficient-evidence responses. This panel makes no live model call; accuracy awaits researcher review.") st.download_button("Download recorded Meta predictions", data=json.dumps(recorded["predictions"], ensure_ascii=False, indent=2), file_name="meta_pilot_predictions_20261028.json", mime="application/json") ablation = run_recorded_pilot() with st.expander("Gate ablation · recorded 12-case Glimmer batch", expanded=False): st.warning( "POST-GENERATION REPLAY ONLY. The recorded model prompt already requested evidence quotes, " "so this does not isolate the causal effect of quote-grounded prompting. " "The 12-case human review is still pending; no accuracy is reported." ) g1, g2, g3, g4 = st.columns(4) g1.metric("Raw decisive", ablation["arm_a_fixed_model_outputs"]["decisive"]) g2.metric("Quote-gated decisive", ablation["arm_b_source_quote_post_gate"]["decisive"]) g3.metric("New abstentions from gate", ablation["effects"]["quote_gate_new_abstentions"]) g4.metric("Human reviews pending", ablation["arm_c_source_quote_plus_human_review"]["pending_review"]) st.caption( f"Recorded Meta call: {ablation['recorded_model_call']['usage']['total_tokens']} tokens · " f"{ablation['recorded_model_call']['elapsed_seconds']:.3f}s · " f"new model calls for this replay: {ablation['new_model_calls']}." ) st.caption( "In this clean recorded batch, all eight decisive outputs already had valid anchored quotes, " "so the post-processing source/quote gate introduced zero additional abstentions. " "That is a result, not evidence that the gate is unnecessary." ) st.download_button( "Download recorded gate-ablation JSON", data=json.dumps(ablation, ensure_ascii=False, indent=2, sort_keys=True), file_name="gate_ablation_recorded_20261028.json", mime="application/json", ) st.download_button( "Download literature reviewer packet", data=json.dumps(reviewer_packet(pilot), ensure_ascii=False, indent=2), file_name="frontierlab_real_excerpt_review.json", mime="application/json") reviewer_name = st.text_input( "Reviewer name for an uploaded 12-case packet", key="pilot_review_reviewer", help="Used for provenance only; FrontierLab does not verify identity." ) uploaded_review = st.file_uploader( "Upload edited 12-case reviewer packet (JSON)", type=["json"], key="pilot_review_upload", help="Keep the source/case fields unchanged; edit only label and rationale fields." ) if uploaded_review is not None: try: review_upload = json.loads(uploaded_review.getvalue().decode("utf-8")) canonical_review = canonicalize_review_upload( pilot, review_upload, reviewer=reviewer_name ) review_stats = canonical_review["review_stats"] review_labels = labels_from_review_packet(pilot, canonical_review) reviewed_ablation = run_recorded_pilot(review_labels=review_labels) arm_c = reviewed_ablation["arm_c_source_quote_plus_human_review"] if canonical_review["status"] == "COMPLETE_RESEARCHER_REVIEW": st.success( "12-case researcher review packet is complete and source-bound. " "It remains researcher-supplied review, not independent gold." ) else: st.info( f"Partial researcher review: {review_stats['reviewed']} / " f"{review_stats['total']} cases completed." ) rv1, rv2, rv3 = st.columns(3) rv1.metric("Cases reviewed", review_stats["reviewed"]) rv2.metric("Arm C reviews completed", arm_c["reviews_completed"]) rv3.metric( "Reviewer changes to surviving decisions", arm_c["review_corrections_to_surviving_decisions"], ) st.caption( f"Arm C pending reviews: {arm_c['pending_review']}. " "Accuracy remains unavailable because independence_verified=false." ) st.download_button( "Download canonical source-bound review JSON", data=json.dumps(canonical_review, ensure_ascii=False, indent=2, sort_keys=True), file_name="frontierlab_real_excerpt_review_completed.json", mime="application/json", ) except (ValueError, UnicodeDecodeError, json.JSONDecodeError) as exc: st.error(f"Reviewer packet rejected: {exc}") st.markdown("[Pilot protocol and source attribution](https://github.com/Jaycee871/FrontierLab-Meta-Hackathon-2026/blob/main/docs/REAL_LITERATURE_PILOT.md)") review_path = Path(__file__).resolve().parent / "docs/runs/nim_claim_review_completed_20261028.json" if review_path.exists(): reviewed = json.loads(review_path.read_text(encoding="utf-8")) st.success("RECORDED CORE RUN · Jaycee871 reviewed both claims as SUPPORTED on October 28, 2026. These judgments cover two claims, separate from the 12-case pilot.") st.caption("The original extraction receipt is preserved. No hypothesis, method selection or plan was generated by recording this review; this panel makes no model call.") st.dataframe([{"Case": c["case_id"], "Claim": c["claim"], "Human judgment": c["review_label"], "Reason": c["review_note"]} for c in reviewed["cases"]], hide_index=True, width="stretch") st.download_button("Download completed core-workflow human review", data=review_path.read_text(encoding="utf-8"), file_name=review_path.name, mime="application/json") result = st.session_state.get("frontier_result") if result: active_client = st.session_state.get("frontier_model_client") if result.backend == "nvidia-nim" and (not use_nim or not nim_authorized): active_client = None if result.backend == hf_providers.BACKEND and (not use_hf or not hf_consent or not hf_available): active_client = None stats = summary(result) a, b, c, d = st.columns(4) a.metric("Sources", stats["sources"]) b.metric("Claims: passage checks / synthetic", stats["claims"]) c.metric("Conflicts", stats["conflicts"]) d.metric("Rejected claims", stats["rejected"]) st.caption(f"Decision gate: **{stats['decision']}** · Backend: {result.backend} · Model: {result.model} · Data source: {result.mode}") if result.inference: with st.expander("Model request receipts"): st.json(result.inference) with st.expander("Process metrics · audit only", expanded=False): pm = process_metrics(result) m1, m2, m3, m4 = st.columns(4) m1.metric("Completed stages", f"{pm['stage_completion']['completed']} / 5") m2.metric("SCR", f"{pm['scr'] * 100:.0f}%") m3.metric("Observed human actions", pm["human_interventions"]["observed_total"]) aerr = pm["nim"]["aerr"] m4.metric("AERR", "N/A" if aerr is None else f"{aerr * 100:.0f}%") st.caption( "Process metrics describe workflow activity, not model accuracy. " "Observed human actions use recorded intervention events when available; legacy runs use a conservative fallback. Identity is not verified." ) taxonomy = pm["error_taxonomy"] e1, e2, e3 = st.columns(3) e1.metric("Governance / evidence blocks", taxonomy["governance_or_evidence_blocks"]) e2.metric("Runtime failures", taxonomy["runtime_failures"]) e3.metric("Scientific tensions", taxonomy["scientific_tensions"]) if pm["nim"]["applicable"]: st.caption( f"NIM calls: {pm['nim']['attempted_calls']} · " f"tokens: {pm['nim']['total_tokens']} · " f"recorded model time: {pm['nim']['elapsed_seconds']:.3f}s" ) with st.expander("Research Operating Interface · operational status", expanded=False): roi = research_operating_status(result) synthetic_note = None if result.mode == "demo": synthetic_strict = roi["traceability"]["strict_extracted_coverage"] synthetic_note = ( "The synthetic fixture contains no evidence quotes, so strict EXTRACTED coverage is " f"{_pct(synthetic_strict)} by design (fail-closed)." ) if replay is not None: synthetic_note += ( " The recorded real M2 core measures " f"{_pct(replay['trace_coverage']['source_trace_coverage'])} source / " f"{_pct(replay['trace_coverage']['strict_extracted_coverage'])} strict: see " f"“{REPLAY_TITLE}” at the top of the page." ) render_operating_status(roi, key="live", download_label="Download operating-status JSON", file_name="research_operating_status.json", synthetic_note=synthetic_note) for warning in result.audit.warnings: if result.mode == "demo" and warning.startswith("SYNTHETIC FIXTURE"): continue # disclosed once by the synthetic banner st.warning(warning) if result.mode == "demo" and mode != "demo": st.warning(SYNTHETIC_BANNER) if result.mode == "demo": with st.expander("Synthetic audit fixtures (FICTIONAL) · heterogeneity, claim benchmark, failure lab", expanded=False): fx1, fx2, fx3 = st.tabs(["Wolfram-validated heterogeneity (FICTIONAL numbers)", "Claim-to-passage audit benchmark (FICTIONAL · hand-labeled)", "Research Failure Lab · 12 fictional failure injections"]) with fx1: stats_q = inverse_variance_heterogeneity([0.4, -0.3], [0.15, 0.1]) st.caption("SYNTHETIC: These invented effect sizes do NOT represent observed studies.") x, y, z = st.columns(3) x.metric("Pooled effect (not a conclusion)", f"{stats_q['pooled_effect']:.4f}") y.metric("Cochran Q", f"{stats_q['Q']:.4f}") z.metric("I²", f"{stats_q['I2'] * 100:.1f}%") st.caption("High inconsistency: do NOT force consensus. Independently checked with Wolfram Language; no online Wolfram call is made by this app.") with fx2: fixture_papers, fixture_cases = benchmark() verified = audit_predictions(fixture_papers, fixture_cases) st.caption("All ten examples and predictions are HAND-AUTHORED FICTION. Scores do not represent Meta/NVIDIA model accuracy, nor prove scientific validity.") x, y, z = st.columns(3) x.metric("Known source IDs", f"{verified['source_id_coverage']:.0%}") y.metric("Verbatim excerpt matches", f"{verified['verbatim_passage_coverage']:.0%}") z.metric("Synthetic gold-label agreement", f"{verified['label_accuracy_vs_independent_gold']:.0%}") st.caption("A quotation can exist yet be used to support a FALSE statement; genuine semantic verification needs independently annotated evidence.") st.json({"cases": verified["cases"], "missing_conflict_detection_example": conflict_transparency([], [("SYNTHETIC:A","SYNTHETIC:B")])}) with fx3: lab = suite_report() st.caption( "OFFLINE HAND-AUTHORED POLICY TESTS ONLY: no actual Meta inference, " "real biomedical model call, independent scientific assessment or " "human learning experiment has occurred." ) u, v, w = st.columns(3) u.metric("Policy test cases", lab["total"]) v.metric("Expected guard decisions matched", lab["matched_expectations"]) w.metric("Real model calls", lab["model_calls"]) st.caption( "Cross-stage scenarios: stale source snapshots, unsupported excerpts, " "unverified BioNeMo origin, train/test peptide leakage, and human teachability evidence. " "A quote that exists can still be semantically wrong." ) st.json(lab["cases"]) st.download_button( "Download fictional failure-injection results", data=json.dumps(lab, indent=2, ensure_ascii=False), file_name="frontierlab_v011_failure_lab_fictional.json", mime="application/json", ) with st.expander("2 · Evidence catalogue", expanded=True): for p in result.papers: st.markdown(f"**{p.title}** · {p.paper_id} · {p.year}") st.caption(f"{p.authors} · {p.source} · {'SYNTHETIC / NO LINK' if p.synthetic else p.url}") st.write(p.abstract) st.divider() with st.expander("Evidence claims and citation checks", expanded=True): if not result.audit.verified_claims: st.info("No claims have passed the source-ID gate. This is not evidence that a scientific claim is false.") for claim in result.audit.verified_claims: st.markdown(f"**{claim.stance.upper()}** · {claim.paper_id} · {claim.claim_key}") st.write(claim.statement) if claim.passage_matched: st.caption(f"Verbatim abstract excerpt (NOT semantic verification): {claim.evidence_quote}") else: st.caption("No passage match recorded (synthetic fixture only).") st.caption(f"Limitations: {claim.limitation}") st.caption("Valid source IDs and verbatim excerpts are NOT independent scientific fact-checking or semantic entailment. Compare claims against the original article and methods.") if result.audit.rejected_claims: st.json(result.audit.rejected_claims) with st.expander("Evidence conflicts and skeptical review", expanded=True): if result.audit.conflicts: st.error("Incompatible directions detected. Do not report consensus.") st.json(result.audit.conflicts) else: st.info("No machine-detected contradictions; this is not proof of agreement.") st.json(result.skeptic) if result.mode == "live" and result.audit.verified_claims: band("II · The Sphere — human claim review", "sphere") with st.expander("Researcher claim review · required before synthesis", expanded=True): st.warning( "Verbatim excerpt matching does NOT establish entailment. " "Choose a human judgment for each claim. PENDING items and fewer than two " "supported distinct sources block research hypothesis generation." ) packet=review_packet(result) previous={c["case_id"]:c for c in (result.review or {}).get("cases", [])} st.download_button( "Download source-bound blank reviewer packet", data=json.dumps(packet, ensure_ascii=False, indent=2), file_name="frontierlab_human_review_blank.json", mime="application/json" ) decisions=[] with st.form("claim_reviewer_form"): for case in packet["cases"]: cid=case["case_id"] earlier=previous.get(cid, {}) st.markdown(f"**{cid} · {case['paper_title']}**") st.caption(f"Source ID: {case['paper_id']} · {case['source']}") st.write("**Model claim:** " + case["claim"]) st.code(case["evidence_quote"], language="text") st.caption("The quotation exists in the source abstract. Its meaning has NOT been verified.") opts=["PENDING","SUPPORTED","CONTRADICTED","INSUFFICIENT"] default=earlier.get("review_label","PENDING") verdict=st.selectbox( "Reviewer judgment", opts, index=opts.index(default) if default in opts else 0, key=f"claim_review_{cid}" ) note=st.text_input( "Reason for verdict (required if contradicted or insufficient)", value=earlier.get("review_note",""), key=f"claim_reason_{cid}" ) decisions.append({"case_id":cid,"review_label":verdict,"review_note":note}) st.divider() submitted=st.form_submit_button("Save human review judgments", type="primary") if submitted: try: save_review(result, decisions) st.session_state["frontier_result"]=result st.success("Human review saved in this session; download JSON to preserve it.") except ValueError as exc: st.error(f"Review validation failed: {exc}") if result.review: st.json({"review_status":result.review.get("status"), "review_statistics":result.review.get("review_stats"), "independent_expert_adjudication_verified":False}) st.download_button( "Download human-annotated review JSON", data=json.dumps(result.review,ensure_ascii=False,indent=2), file_name="frontierlab_human_review_annotated.json", mime="application/json" ) if result.review.get("status")=="READY_FOR_HUMAN_GATED_SYNTHESIS": st.info("Review complete. Hypothesis synthesis requires an actually connected, available model and further human scrutiny.") if st.button("Resume model synthesis after human review", type="secondary", disabled=active_client is None): try: continue_after_review(result, model_client=active_client) st.session_state["frontier_result"]=result st.success("Provisional human-gated draft created; scientific validation is still needed.") except HFProvidersUnavailable as exc: st.error(str(exc)) except (ValueError, PermissionError, RuntimeError) as exc: st.error(f"Synthesis remains blocked: {type(exc).__name__}: {exc}") band("III · Spaceland — hypothesis, evaluation and plan", "cube") with st.expander("3 · Falsifiable hypotheses", expanded=True): st.json(result.hypotheses) render_evaluation(result, model_client=active_client) st.subheader("5 · Research plan approval gate") try: require_evaluation(result) evaluation_ready = True except (ValueError, PermissionError): evaluation_ready = False if evaluation_ready and result.plan is None: try: readiness = build_planning_readiness(result) ready_stats = readiness_summary(readiness) with st.expander("Planning readiness · pre-approval contract", expanded=False): render_planning_readiness(readiness, ready_stats, key="live", download_label="Download planning-readiness JSON", file_name="planning_readiness.json") except (ValueError, PermissionError) as exc: st.info(f"Planning readiness not available: {exc}") approved = st.checkbox("I have reviewed the evidence, conflicts and limitations. Allow a DRAFT research plan.") if st.button("Generate cited plan", disabled=not approved or not result.hypotheses or not evaluation_ready or (result.mode == "live" and active_client is None), type="primary"): try: planner = active_client if result.mode == "live" else None make_plan(result, model_client=planner, approved=approved) st.success("Draft created. Human scientific review is still required.") except Exception as exc: st.error(f"Planning blocked: {exc}") if result.plan: st.json(result.plan) # Rebuilt from the current run on every render, so reviews, hypotheses and plans stay in sync. graph = refresh_evidence_graph(result) g_stats = graph_summary(graph) with st.expander("6 · Evidence graph · provenance of every plan step", expanded=False): render_graph_overview(graph, g_stats) steps = [n for n in graph["nodes"] if n["type"] == "plan_step"] if steps: gates = {n["id"]: step_gate(graph, n["id"]) for n in steps} st.dataframe([{"Step": n["index"], "Text": n["text"], "Path gate": "ALL EXTRACTED" if gates[n["id"]]["approvable"] else "BLOCKED", "Researcher approval": "RECORDED" if n["id"] in graph["approvals"] else "—", "Reasons": "; ".join(gates[n["id"]]["reasons"])} for n in steps], hide_index=True, width="stretch") eligible = [n["id"] for n in steps if gates[n["id"]]["approvable"] and n["id"] not in graph["approvals"]] if eligible: with st.form("plan_step_approval_form"): chosen = st.selectbox("Plan step with an all-EXTRACTED evidence path", eligible, format_func=lambda sid: next(f"Step {n['index']}" for n in steps if n["id"] == sid)) reviewer = st.text_input("Reviewer") rationale = st.text_input("Approval rationale (8+ characters)") if st.form_submit_button("Record researcher approval of this step", type="secondary"): try: approve_plan_step(graph, chosen, reviewer=reviewer, note=rationale) record_human_event(result, "PLAN_STEP_APPROVED", chosen, actor_label=reviewer, note=rationale) st.session_state["frontier_result"] = result st.success("Step approval recorded in this session; it remains a draft, not a validated protocol.") except (ValueError, PermissionError) as exc: st.error(f"Step approval blocked: {exc}") else: st.info("No plan step is currently eligible for approval under the EXTRACTED-only path rule.") else: st.caption("No plan steps yet. Steps appear here once a draft plan exists.") st.download_button("Download graph.json", data=graph_json(graph), file_name="graph.json", mime="application/json") trajectory = build_trajectory_graph(result) t_stats = trajectory_summary(trajectory) with st.expander("7 · Trajectory audit graph · process trace", expanded=False): render_trajectory(trajectory, t_stats, key="live", model_calls_label="Model calls", download_label="Download trajectory.json", file_name="trajectory.json") with st.expander("Audit trail", expanded=False): st.json(result.log) st.download_button("Download complete provenance JSON", data=export_json(result), file_name="frontierlab_audit.json", mime="application/json") else: st.info("Begin with the synthetic audit or Europe PMC. Model-backed research requires live Meta Muse via Hugging Face (owner-configured token and your consent), a configured local server or an explicitly enabled NVIDIA backend.")