Implement Scene 7 evidence goal flow

server/evidenceJudge.test.ts

@@ -0,0 +1,49 @@
+import { afterEach, describe, expect, it, vi } from 'vitest'
+import { createEvidenceJudgeFromEnv, EvidenceJudgeError, validateEvidenceVerdict } from './evidenceJudge.js'
+
+const evidence = 'Patent applicant Nils Aall Barricelli describes an improved chest of drawers with rotating compartments.'
+
+describe('semantic evidence judge', () => {
+  afterEach(() => {
+    delete process.env.EVIDENCE_JUDGE_PROVIDER
+    delete process.env.EVIDENCE_JUDGE_MODEL
+    delete process.env.ANTHROPIC_API_KEY
+    delete process.env.EVIDENCE_JUDGE_MAX_CHARACTERS
+    vi.restoreAllMocks()
+  })
+
+  it('validates a supported verdict only when its quotation exists in the OCR', () => {
+    expect(validateEvidenceVerdict({ subject:'target',supports_claim:true,evidence_excerpt:'Nils Aall Barricelli describes an improved chest of drawers',confidence:.94 }, evidence)).toEqual({
+      subject:'target', supportsClaim:true, evidenceExcerpt:'Nils Aall Barricelli describes an improved chest of drawers', confidence:.94,
+    })
+    expect(() => validateEvidenceVerdict({ subject:'target',supports_claim:true,evidence_excerpt:'invented quotation',confidence:.99 }, evidence))
+      .toThrow(EvidenceJudgeError)
+  })
+
+  it('is disabled safely without explicit provider configuration', async () => {
+    const judge = createEvidenceJudgeFromEnv()
+    expect(judge.enabled).toBe(false)
+    await expect(judge.judge({ targetSubject:'Nils', assertion:'was an inventor', evidenceText:evidence })).rejects.toMatchObject({ code:'provider_unavailable' })
+  })
+
+  it('uses a constrained Anthropic tool response and truncates untrusted OCR', async () => {
+    process.env.EVIDENCE_JUDGE_PROVIDER = 'anthropic'
+    process.env.EVIDENCE_JUDGE_MODEL = 'configured-cheap-model'
+    process.env.ANTHROPIC_API_KEY = 'test-secret'
+    process.env.EVIDENCE_JUDGE_MAX_CHARACTERS = '1000'
+    const fetcher = vi.fn(async (_url: string | URL | Request, init?: RequestInit) => {
+      const request = JSON.parse(String(init?.body))
+      expect(request.model).toBe('configured-cheap-model')
+      expect(request.tool_choice).toEqual({ type:'tool',name:'record_evidence_verdict' })
+      expect(String(request.messages[0].content)).not.toContain('x'.repeat(1001))
+      expect(new Headers(init?.headers).get('x-api-key')).toBe('test-secret')
+      return new Response(JSON.stringify({ content: [{ type:'tool_use',name:'record_evidence_verdict',input:{
+        subject:'target',supports_claim:true,evidence_excerpt:'Nils Aall Barricelli describes an improved chest of drawers',confidence:.93,
+      } }] }), { status:200,headers:{'content-type':'application/json'} })
+    })
+    const judge = createEvidenceJudgeFromEnv(fetcher)
+    const verdict = await judge.judge({ targetSubject:'Nils Aall Barricelli', relatedSubject:'his father', assertion:'was an inventor', evidenceText:`${evidence}${'x'.repeat(5000)}` })
+    expect(verdict).toMatchObject({ subject:'target',supportsClaim:true,confidence:.93 })
+    expect(fetcher).toHaveBeenCalledOnce()
+  })
+})