What Is the Condrox Evidence Engine?
The Condrox Evidence Engine is an automated, deterministic verification layer that produces academic-grade evidence for every major capability of the system. Rather than relying on hand-written claims, the engine executes the actual Condrox AI code against a fixed suite of scenarios and records the measured outputs. This approach ensures that every reported metric corresponds to a real execution trace that can be inspected and reproduced.
The engine evaluates seven dimensions of evidence, organized as Phases 1 through 7:
- System Verification (Phase 1): boot probes, health checks, and API status.
- Scenario Runs: end-to-end intent and mode classification accuracy across the scenario suite.
- Conversational Evidence (Phase 2): empathy, relevance, consistency, multi-turn stability, and context preservation.
- Reasoning Evidence (Phase 3): step-by-step structure, logical coherence, explanation quality, planning, code analysis, and correctness signals.
- Agent & Pipeline Evidence (Phase 4): agent usage, metadata propagation, retrieval-augmented generation, memory, personality, and pipeline stage traces.
- Full-System Integration Evidence (Phase 5): combined multi-turn conversation, reasoning, agent, retrieval, and API health evaluated within a single scenario.
- Academic & Lab Reports (Phase 6): PDF-compatible HTML reports with methodology, references, signature, and reproducibility information.
- Evidence Portal (Phase 7): a static dashboard providing a report library, score-history chart, version history, and system signatures.
Latest Report Snapshot
The WARN status reflects remaining reasoning-metric limitations (StepByStepStructure and ExplanationQuality on code_analysis). The full-system release status is READY, with a complete-session score of 92.94%. These are measured observations, not concealed failures. The engine reports them transparently so that subsequent improvements can be tracked against a documented baseline.
Generated Artifacts
Each evidence run produces the following artifacts in Reports/evidence/. The artifacts span machine-readable payloads, dashboard-style reports, and three academic editions tailored to different review contexts:
| Artifact | Description |
|---|---|
*.json | Machine-readable full evidence payload |
*.html | Dashboard-style HTML report |
*_lab_edition.academic.html | Lab-grade report with methodology and signature |
*_academic_edition.academic.html | Academic edition with reasoning focus |
*_peer_review_edition.academic.html | Peer-review edition with conversation evidence |
index.html | Evidence Portal: report library and score history |
Reproducibility
The reports are fully reproducible. Running the following command in the project root regenerates every artifact using the same scenario suite and measurement rules:
python scripts/generate_evidence.pyTo verify a previously generated report against the current runtime, use the verification command:
condrox verifyYou can also open the portal directly from the Condrox runtime manager:
condrox evidence-portalKey Source Files
core/evidence/evidence_collector.py: orchestrates all evidence collection across phases.core/evidence/conversation_evaluator.py: computes conversation quality metrics.core/evidence/reasoning_evaluator.py: computes reasoning quality metrics.core/evidence/agent_pipeline_evaluator.py: measures agent and pipeline utilization.core/evidence/full_system_evaluator.py: performs end-to-end integration evaluation.core/evidence/academic_report_generator.py: renders lab-grade academic reports.core/evidence/portal_generator.py: builds the static evidence portal.scripts/generate_evidence.py: CLI entry point for evidence generation.