feat: Add counter-evidence generation feature to challenge research f…

app/research_workflow/companion_routes.py

@@ -111,6 +111,34 @@ def companion_ask(project_id):
         return json_error(str(e), status_code=500)
 
 
+# =========================================================================
+# Counter-Evidence (Devil's Advocate)
+# =========================================================================
+
+@research_workflow_bp.route('/companion/<int:project_id>/counter-evidence', methods=['POST'])
+@login_required
+def companion_counter_evidence(project_id):
+    """Generate counter-evidence challenging a research finding."""
+    project = get_user_resource_or_403(ResearchProject, project_id, current_user.id)
+
+    data = request.json or {}
+    finding = data.get('finding', '').strip()
+    research_question = data.get('research_question', '').strip() or None
+    step_index = data.get('step_index', project.current_step_index)
+
+    if not finding:
+        return json_validation_error('Finding is required')
+
+    try:
+        argos = ArgosService(user_id=current_user.id)
+        context = argos.build_research_context(project_id, step_index=step_index)
+        counter_evidence = argos.generate_counter_evidence(context, finding, research_question)
+        return json_success('Counter-evidence generated', data={'counter_evidence': counter_evidence})
+    except Exception as e:
+        logger.error(f"Counter-evidence generation failed: {e}")
+        return json_error(str(e), status_code=500)
+
+
 # =========================================================================
 # Session Wrap-Up
 # =========================================================================

app/services/ai/prompts/companion/counter_evidence.yaml

@@ -0,0 +1,64 @@
+name: "counter_evidence"
+description: "Surface disconfirming evidence and questions that challenge a research finding"
+version: "1.0"
+category: "companion"
+
+preferred_provider: "gemini"
+model: "gemini-3-pro-preview"
+max_tokens: 6000
+temperature: 0.3
+
+system_context: |
+  You are a devil's advocate for an investor's research finding.
+  Your role is to surface what would have to be TRUE for the finding to be WRONG.
+
+  CRITICAL RULES:
+  1. You surface DISCONFIRMING EVIDENCE, ALTERNATIVE EXPLANATIONS, and QUESTIONS.
+     You NEVER give investment opinions or recommendations.
+  2. Never say "you should sell", "this is a bad investment", or "the thesis is wrong".
+     Say "this finding would not hold if..." and "the data that would disprove this is...".
+  3. Distinguish clearly between what you know as fact and what is an untested assumption.
+     Label anything you cannot verify as "unverified".
+  4. Attack the finding, never the investor. Reference their history as facts:
+     "Your mistake log shows a similar assumption on..." not "You tend to...".
+  5. If the finding is already well-supported by the prior findings, say so plainly
+     rather than manufacturing weak objections.
+
+template: |
+  COMPANY: {company_name}
+  SECTOR: {sector_name}
+  CURRENT STEP: {step_name}
+
+  FINDING TO CHALLENGE:
+  {finding}
+
+  RESEARCH QUESTION IT ANSWERS:
+  {research_question}
+
+  OTHER FINDINGS SO FAR:
+  {prior_findings}
+
+  CURRENT THESIS:
+  {investment_thesis}
+
+  KEY FLAGS:
+  Red flags: {red_flags}
+  Green flags: {green_flags}
+
+  INVESTOR'S HISTORY:
+  Past mistakes: {mistake_summary}
+  Behavioral patterns to watch: {pattern_summary}
+
+  Challenge the finding with:
+  1. ASSUMPTIONS BEING MADE: What the finding takes for granted but has not verified
+  2. DISCONFIRMING EVIDENCE TO LOOK FOR: Specific data, filings, or metrics that would
+     contradict the finding if they exist
+  3. ALTERNATIVE EXPLANATIONS: Other readings of the same evidence
+  4. QUESTIONS TO ANSWER NEXT: Concrete questions that would settle whether it holds
+
+  Be specific and concrete. No opinions, no recommendations. Facts, gaps, and questions only.
+
+output_format: |
+  Structured text with the four numbered section headers above.
+  Two to four bullet points per section, one sentence each.
+  Every bullet should be checkable against a source the investor can go find.

app/services/argos/companion.py

@@ -183,6 +183,38 @@ def ask_companion(
             logger.error(f"Companion chat failed: {e}")
             return f"Could not process question: {e}"
 
+    def generate_counter_evidence(
+        self,
+        context: CompanionContext,
+        finding: str,
+        research_question: str = None,
+    ) -> str:
+        """Play devil's advocate against a finding. Disconfirming evidence and questions, no opinions."""
+        try:
+            prompt_data = prompt_service.get_prompt_with_metadata(
+                'companion', 'counter_evidence',
+                company_name=context.company_name,
+                sector_name=context.sector_name,
+                step_name=context.step_name,
+                finding=finding,
+                research_question=research_question or context.research_questions,
+                prior_findings=context.prior_findings,
+                investment_thesis=context.investment_thesis,
+                red_flags=context.red_flags,
+                green_flags=context.green_flags,
+                mistake_summary=context.mistake_summary,
+                pattern_summary=context.pattern_summary,
+            )
+            model_enum, provider_enum = resolve_model_provider(
+                prompt_data.get('metadata', {}), user_id=self.user_id, prompt_category='companion',
+            )
+            return ai_service.generate_text(
+                prompt_data['prompt'], model=model_enum, provider=provider_enum,
+            )
+        except Exception as e:
+            logger.error(f"Counter-evidence generation failed: {e}")
+            return f"Could not generate counter-evidence: {e}"
+
     def wrap_up_session(
         self,
         context: CompanionContext,