Enhance catastrophic risk detection and offline reasoning

experiments/comparison_agents/agents4sci_wrapper.py

@@ -50,7 +50,6 @@ def reason(self, scenario: str, domains: List[str]) -> ReasoningResult:
                 model_id=self.model_id,
                 expert_roles=domains,
                 rounds=3,
-                max_workers=8
             )
 
                     # Create and run orchestrator
@@ -102,34 +101,34 @@ def reason(self, scenario: str, domains: List[str]) -> ReasoningResult:
 ## Conclusion
 The Agents4Sci multi-agent approach provides comprehensive coverage across all relevant domains, identifies critical cross-domain interactions, and quantifies uncertainties systematically. This approach significantly improves the quality and reliability of counterfactual scenario analysis."""
 
-            # Extract insights from the multi-round process
-            insights = self._extract_agents4sci_insights(temp_path, final_report)
-            
-            return ReasoningResult(
-                agent_type="Agents4Sci",
-                reasoning_steps=insights["reasoning_steps"],
-                conclusions=insights["conclusions"],
-                assumptions=insights["assumptions"],
-                uncertainties=insights["uncertainties"],
-                cross_domain_insights=insights["cross_domain_insights"],
-                confidence_score=insights["confidence_score"],
-                reasoning_depth=insights["reasoning_depth"],
-                raw_output=final_report,
-                metadata={
-                    "domains_covered": domains,
-                    "approach": "agents4sci_multi_agent",
-                    "rounds_completed": metadata.get("rounds_completed", 0),
-                    "experts_consulted": len(domains),
-                    "conflicts_resolved": metadata.get("conflicts_resolved", 0),
-                    "advantages": [
-                        "Multi-agent collaboration",
-                        "Conflict resolution",
-                        "Iterative refinement",
-                        "Domain specialization",
-                        "Cross-domain synthesis"
-                    ]
-                }
-            )
+        # Extract insights from the multi-round process
+        insights = self._extract_agents4sci_insights(temp_path, final_report)
+
+        return ReasoningResult(
+            agent_type="Agents4Sci",
+            reasoning_steps=insights["reasoning_steps"],
+            conclusions=insights["conclusions"],
+            assumptions=insights["assumptions"],
+            uncertainties=insights["uncertainties"],
+            cross_domain_insights=insights["cross_domain_insights"],
+            confidence_score=insights["confidence_score"],
+            reasoning_depth=insights["reasoning_depth"],
+            raw_output=final_report,
+            metadata={
+                "domains_covered": domains,
+                "approach": "agents4sci_multi_agent",
+                "rounds_completed": metadata.get("rounds_completed", 0),
+                "experts_consulted": len(domains),
+                "conflicts_resolved": metadata.get("conflicts_resolved", 0),
+                "advantages": [
+                    "Multi-agent collaboration",
+                    "Conflict resolution",
+                    "Iterative refinement",
+                    "Domain specialization",
+                    "Cross-domain synthesis"
+                ]
+            }
+        )
 
     def _extract_agents4sci_insights(self, output_dir: Path, final_report: str) -> Dict[str, Any]:
         """Extract structured insights from Agents4Sci outputs"""

experiments/comparison_agents/mock_llm_client.py

@@ -45,6 +45,10 @@ def chat(
 - Human health impacts from oxidative stress
 - Economic disruption from increased fire incidents
 
+## Catastrophic Risks:
+- Catastrophic global wildfires leading to infrastructure collapse
+- Potential extinction events for oxygen-sensitive species
+
 ## Assumptions:
 - Current technology levels maintained
 - Gradual increase over 50 years allows some adaptation
@@ -77,6 +81,10 @@ def chat(
 - Higher biological mutation rates
 - Human health risks from radiation
 
+## Catastrophic Risks:
+- Catastrophic failure of electrical grids from solar storms
+- Widespread radiation-induced health crises
+
 ## Assumptions:
 - Gradual decrease over 100 years
 - Current technology levels maintained
@@ -109,6 +117,10 @@ def chat(
 - Economic disruption in coastal regions
 - Human health effects from malnutrition
 
+## Catastrophic Risks:
+- Catastrophic collapse of global fisheries
+- Mass starvation events in vulnerable regions
+
 ## Assumptions:
 - Rapid pH decrease over 30 years
 - Current technology levels maintained
@@ -141,6 +153,10 @@ def chat(
 - Need for adaptation and mitigation strategies
 - Potential for both positive and negative outcomes
 
+## Catastrophic Risks:
+- Catastrophic cascading failures across critical systems
+- Potential for global socio-economic collapse
+
 ## Assumptions:
 - Current technology levels maintained
 - Gradual change allows some adaptation

src/a4s/agents.py

@@ -68,14 +68,16 @@ def run(self, expert_input: ExpertInput) -> ExpertOutput:
         text = resp["choices"][0]["message"]["content"].strip()
 
         # Non-strict parsing by headings
-        sections = {"reasoning": [], "conclusions": [], "assumptions": [], "uncertainties": [], "dependency": []}
+        sections = {"reasoning": [], "conclusions": [], "catastrophic": [], "assumptions": [], "uncertainties": [], "dependency": []}
         current = None
         for line in text.splitlines():
             lower = line.lower()
             if "reasoning" in lower and ":" in line:
                 current = "reasoning"; continue
             if "conclusion" in lower and ":" in line:
                 current = "conclusions"; continue
+            if "catastrophic" in lower and ":" in line:
+                current = "catastrophic"; continue
             if "assumption" in lower and ":" in line:
                 current = "assumptions"; continue
             if "uncertaint" in lower and ":" in line:
@@ -89,6 +91,7 @@ def run(self, expert_input: ExpertInput) -> ExpertOutput:
             role=self.role,
             reasoning_steps=sections["reasoning"] or [text],
             conclusions=sections["conclusions"] or [],
+            catastrophic_risks=sections["catastrophic"] or [],
             assumptions=sections["assumptions"] or [],
             uncertainties=sections["uncertainties"] or [],
             dependency_notes=sections["dependency"] or [],
@@ -113,6 +116,8 @@ def reconcile(self, expert_outputs: List[ExpertOutput]) -> ConflictReport:
                 *out.reasoning_steps,
                 "Conclusions:",
                 *out.conclusions,
+                "Catastrophic Risks:",
+                *out.catastrophic_risks,
                 "Assumptions:",
                 *out.assumptions,
                 "Uncertainties:",

src/a4s/llm_client.py

@@ -1,7 +1,12 @@
 import os
+import sys
+from pathlib import Path
 from typing import List, Dict, Any, Optional
 
-from openai import OpenAI
+try:
+    from openai import OpenAI
+except Exception:  # pragma: no cover - library may be absent in testing
+    OpenAI = None  # type: ignore
 
 
 class LLMClient:
@@ -17,11 +22,19 @@ def __init__(
         default_model: str = "doubao-seed-1-6-flash-250715",
     ) -> None:
         api_key = os.environ.get(api_key_env)
-        if not api_key:
-            raise RuntimeError(
-                f"Missing API key in environment variable {api_key_env}."
-            )
-        self.client = OpenAI(base_url=base_url, api_key=api_key)
+        if not api_key or OpenAI is None:
+            # Fall back to mock client for offline testing
+            try:
+                from experiments.comparison_agents.mock_llm_client import MockLLMClient
+            except Exception:
+                # Ensure repository root on path for imports
+                sys.path.append(str(Path(__file__).resolve().parents[2]))
+                from experiments.comparison_agents.mock_llm_client import MockLLMClient
+            self.client = MockLLMClient(default_model=default_model)
+            self._use_mock = True
+        else:
+            self.client = OpenAI(base_url=base_url, api_key=api_key)
+            self._use_mock = False
         self.default_model = default_model
 
     def chat(
@@ -31,6 +44,9 @@ def chat(
         temperature: float = 0.2,
         max_tokens: Optional[int] = None,
     ) -> Dict[str, Any]:
+        if getattr(self, "_use_mock", False):
+            return self.client.chat(messages, model=model or self.default_model, temperature=temperature, max_tokens=max_tokens)
+
         response = self.client.chat.completions.create(
             model=model or self.default_model,
             messages=messages,

src/a4s/orchestrator.py

@@ -140,6 +140,9 @@ def _render_round_summary(round_index: int, outputs: List[ExpertOutput], conflic
             if out.conclusions:
                 parts.append("Conclusions:")
                 parts.extend([f"- {c}" for c in out.conclusions])
+            if out.catastrophic_risks:
+                parts.append("Catastrophic Risks:")
+                parts.extend([f"- {r}" for r in out.catastrophic_risks])
             if out.uncertainties:
                 parts.append("Uncertainties:")
                 parts.extend([f"- {u}" for u in out.uncertainties])

src/a4s/prompts.py

@@ -8,6 +8,7 @@ def system_prefix(role: str) -> str:
         "Think step by step with explicit causal reasoning. "
         "Prefer probabilistic, non-deterministic phrasing. "
         "Model substitution and adaptation when projecting impacts. "
+        "Explicitly flag any catastrophic or extinction-level risks when relevant. "
         "Keep outputs mostly natural language; use compact JSON only if essential. "
         "Outputs MUST be in English; concise, structured bullet points acceptable."
     )
@@ -40,6 +41,7 @@ def expert_round_prompt(role: str, scenario_text: str, shared_summary: str | Non
         "Produce: \n"
         "- Reasoning Steps (chain of thought summaries; no private thoughts)\n"
         "- Conclusions (concise, quotable)\n"
+        "- Catastrophic Risks (explicit catastrophic or extinction-level outcomes, if any)\n"
         "- Assumptions (explicit)\n"
         "- Uncertainties (what could change conclusions)\n"
         "- Dependency Notes (which other domains strongly affect your conclusions)\n"
@@ -73,6 +75,7 @@ def report_generation_prompt(proposition: str, rounds_summaries: List[str]) -> s
         "- Causal map (nodes and edges in bullet form)\n"
         "- Multi-scenario analysis (Scenario 1..N)\n"
         "- Timeline of events (short/medium/long)\n"
+        "- Catastrophic risk assessment (highlight catastrophic or existential threats)\n"
         "Prefer natural language; include minimal JSON only if helpful."
     )
 

src/a4s/schemas.py

@@ -28,6 +28,7 @@ class ExpertOutput:
     role: str
     reasoning_steps: List[str]
     conclusions: List[str]
+    catastrophic_risks: List[str]
     assumptions: List[str]
     uncertainties: List[str]
     dependency_notes: List[str]