How the Compounding Expertise Lab works
The Lab separates three sources: Hamilton Helmer's Seven Powers, Ben Sun's Compounding Expertise thesis, and David Wolfe's extensions around transferability, heterogeneity, marginal information gain, compressibility, causal quality, nonstationarity, and learning velocity.
The V0.2 UX flow is Understand, Observe, Hypothesize, Test, Decide: Overview, System & Environment, Scorebook, Key Debates, Diagnostic, Simulator, Conclusion.
Epistemic model
Observed / derived
Directly calculated from rows or observed in operating evidence.
Sourced
Supported by external evidence or explicit references.
Endogenous assumption
A company-controlled variable currently assumed by the user.
Exogenous assumption
A market or environment variable largely outside company control.
Why Scorebook comes early
The Lab asks users to inspect experience before scoring abstract dimensions. Scorebook rows show whether the system saw a case, made a decision, received human intervention, observed reality, and converted that reality into a grade.
Toy simulator
The simulator uses monthly steps. New cases enter a pending queue and become effective only after `ceil(feedbackDelayDays / 30)` monthly steps. Effective graded experience decays by monthly staleness.
N_eff(t+1) = (1 - staleness_rate) * N_eff(t)
+ cases_whose_feedback_matures_at_t
E(t) = base_capability
+ learning_efficiency
* information_value
* transferability
* ln(1 + N_eff(t))This is an exploratory toy model, not an empirical law or forecast.
Scorebook inspection
The scorebook is inspectable because case-level evidence is the object under test: case, decision, outcome, and grade. Missing outcomes and unresolved grades remain visible because coverage and latency are themselves diagnostics.
Bundled example rows are synthetic illustrative fixtures, not actual company data.
Canonical tests
Why five examples exist
The examples are a canonical test suite: positive test, boundary test, substitution/compression test, alternative Power test, and negative control. They are designed to expose where CE works, where it breaks, and where other forms of Power may dominate.
Synthetic fixture policy
Real-company archetypes may frame the analysis, but bundled case rows are synthetic illustrative data. They test the theory and must not be treated as actual company operations.
CaseSets and source systems
CE is an analytical layer over experience produced by operating systems. The intended loop is: source system -> run / experiment -> CaseSet -> CE analysis, with a return path back to the source system to change policy or assumptions and produce a new CaseSet.
A CaseSet is not automatically a scorebook. It becomes scorebook-like only when it contains decisions, observed outcomes, meaningful grades, and enough provenance to connect learning back to the generating process.
Internal source links use a lightweight return context such as returnTo=compounding-expertise, analysisId, and caseSetId. Full production ingestion remains future work.
Structural model
The Lab separates company identity, exogenous compounding opportunity, endogenous learning architecture, and competitive architecture. Useful learning is not automatically defensible learning.
Cases preserve the sequence: context -> agent decision -> human intervention -> action actually taken-> outcome -> grade. Action timestamps and source-record lineage are kept separate from decision and outcome timestamps so future source systems can link CE evidence back to operational records.
Evidence statuses
Observed
Directly seen in product, customer, or operating data.
Sourced
Supported by cited materials or explicit external evidence.
Assumed
A live hypothesis used for structured reasoning.
Unknown
Not yet supported enough to score confidently.
Deferred analytical layers
Information theory / Shannon layer
Entropy, redundancy, marginal information gain, cross-customer transfer, knowledge compressibility, and decision-relevant value of information are future work.
Autonomy frontier
Empirical decision classes, success/error distributions, downside risk, override value, evidence thresholds, and migration from expert-only to autonomous action are intentionally deferred.