Compounding Expertise Lab
An interactive research instrument for stress-testing whether accumulated graded experience can become durable competitive Power in an AI application.
The problem
Compounding Expertise is a powerful thesis, but it can easily become a story that proves itself. Investors and operators need a structured way to represent the thesis, identify unresolved debates, distinguish evidence from assumptions, and test when historical scorebooks can be compressed or relearned.
The thesis
Accumulated graded experience may become Power only when the product owns a continuous decision-to-outcome learning loop, the cases remain fresh and transferable, and challengers cannot cheaply compress, simulate, or relearn the useful knowledge.
System flow
Step 1
Thesis capture
Capture company, product, customer, workflow, principal decision, and current investment thesis.
Step 2
Key debates
Represent the 2-4 load-bearing questions with bull case, bear case, evidence needed, and belief-revision triggers.
Step 3
Provenance diagnostic
Score HELMER, SUN, and WOLFE dimensions with explicit confidence and evidence status.
Step 4
Lagged compounding simulator
Compare incumbent and challenger scenarios using matured feedback delay, staleness, transferability, information value, and base model capability.
Step 5
Memo synthesis
Generate a concise, uncertainty-preserving memo that surfaces unresolved debates rather than hiding them behind a simplistic moat score.
Case-study artifacts
Workflow
Inputs, Key Debates, Diagnostic, Simulator, and Memo.
No moat score
The tool preserves uncertainty instead of collapsing assumptions into fake precision.
Stress test
Asks whether historical scorebook knowledge can be compressed or relearned.
Decision frame 1
Input
Company thesis, workflow, decision surface, debates, and evidence-backed assessments.
Decision frame 2
Decision
Evaluate whether the scorebook, capture point, or another Power appears defensible.
Decision frame 3
Output
Scenario comparison, unresolved debates, evidence requests, and copyable memo.
Evidence links
Shows company thesis setup and synthetic example loader.
Shows belief revision questions and evidence requests.
Shows HELMER, SUN, and WOLFE provenance-aware assessment.
Shows lagged feedback compounding and crossover detection.
Shows uncertainty-preserving synthesis.
Guardrails
AI suggestions are analysis, not evidence.
Observed and sourced statuses require explicit user input.
The simulator is exploratory and is never presented as an empirical law.
System architecture
1. Thesis capture
Capture company, product, customer, workflow, principal decision, and current investment thesis.
2. Key debates
Represent the 2-4 load-bearing questions with bull case, bear case, evidence needed, and belief-revision triggers.
3. Provenance diagnostic
Score HELMER, SUN, and WOLFE dimensions with explicit confidence and evidence status.
4. Lagged compounding simulator
Compare incumbent and challenger scenarios using matured feedback delay, staleness, transferability, information value, and base model capability.
5. Memo synthesis
Generate a concise, uncertainty-preserving memo that surfaces unresolved debates rather than hiding them behind a simplistic moat score.
KPI callouts
Frameworks
HELMER, SUN, and WOLFE provenance stays visible throughout the workflow.
Evidence states
Observed, sourced, assumed, and unknown are separated explicitly.
Simulator
Feedback delay is modeled as maturation lag, not a simple haircut.
Commercial framing
The Lab is built for rigorous venture/product conversations: it makes the thesis editable, debate-driven, falsifiable, and honest about uncertainty.
What I built
I built the Prisma schema, five-stage workflow, synthetic example loader, belief-revision debate model, provenance-aware diagnostic, lagged toy simulator, crossover detection, memo synthesis, documentation, and unit tests.