PUMA is an applied Agentic Science system positioned at LeCun’s Level 2 of AI scientific capability

Core Insight

LeCun (2026) identifies three levels of AI scientific capability: (1) assistant, (2) model, (3) agent. PUMA’s triage and estimation systems operate at Level 2 — AI as a scientific model that captures domain regularities (PM issue patterns) better than simple baselines. PUMA’s Stage 5 Smart PMO approaches Level 3 — autonomous orchestration of PM workflow cycles. This positioning places PUMA within the validated Agentic Science trajectory, providing theoretical and empirical justification for its claims.


Evidence Chain

LevelGeneral Science ExamplePUMA PM Equivalent
L1: AssistantClaude writing literature reviewPUMA using Claude/NotebookLM for SLR
L2: Scientific ModelAlphaFold predicting protein structuresPUMA triage agent classifying issue priority
L2: Scientific ModelGraphCast forecasting weatherPUMA estimation agent predicting story points
L3: Autonomous AgentAI Scientist generating ML papersPUMA Stage 5 Smart PMO orchestrating sprints

Why Level 2 is Sufficient for PUMA’s Claim

Felin & Holweg (2024) argue AI cannot generate novel causal theories — and they are largely correct for Level 3 claims. But PUMA’s claim is Level 2:

  • “LLM agents can classify Jira issue priority more accurately than majority-class baseline”
  • “LLM agents can estimate story points more accurately than historical average”

These are prediction tasks with ground truth — exactly the domain where Level 2 AI models (AlphaFold, GraphCast, GNoME) have proven highly effective. The argument against Felin & Holweg does not need to go to Level 3 for PUMA’s validation.


Thesis Section Mapping

PUMA SectionPEC2 Connection
1.1 Context and JustificationAI science trajectory → PM automation as same pattern
1.3 Ethical-Social ImpactKlinger (2025): social requirements for responsible AI in science
2. Materials and MethodsLeCun three-level framework as theoretical grounding
4. Conclusions and Future WorkStage 5 approaching Level 3; reproducibility as scientific contribution

MOCs