David Wolfe — AI-driven revenue systems

TOOL

Compounding Expertise Lab

Does experience become a moat?

Compounding Expertise Lab tests whether an AI company can turn repeated real-world decisions and outcomes into expertise that competitors cannot easily reproduce.

Start new company analysis Read methodology

Current analysis

Casap — Public Evidence Analysis — Sep 2026

Company/testCasap — Public Evidence Analysis — Sep 2026Test typeCustom company analysisActive CaseSetNo CaseSet selectedProvenanceNo case provenance available

Continue analysis Start new company analysis

Understand -> Observe -> Hypothesize -> Test -> Decide

How the Lab works

Define the system

What decisions does the product make and in what environment?

Inspect the experience

Examine cases, decisions, human overrides, outcomes, and grades.

Identify what must be true

Surface the 2-4 uncertainties that determine whether learning really compounds.

Stress-test the moat

Compare accumulated experience against stronger models, faster learners, feedback delay, staleness, and other competitive forces.

Decide what you believe

Identify where Power may reside, what remains uncertain, and what evidence to collect next.

Choose a Canonical Test

Each test isolates a different claim about Compounding Expertise. Together they are designed to show where the theory works, where it breaks, and where other forms of Power may dominate.

Bundled case-level records are SYNTHETIC ILLUSTRATIVE DATA - NOT COMPANY DATA.

The real-company examples are theory test fixtures. Their scorebook rows are designed to stress-test the framework, not to describe actual company operations.

POSITIVE TEST

Casap

Canonical question: Can repeated, economically consequential, objectively graded decisions create expertise that a capable challenger cannot quickly reproduce?

What this test isolates: Does CE work under favorable conditions?

Principal decisionDispute / fraud / exception resolution action.Grade objectivityRelatively highFeedback speedRelatively fast - days/weeks as an archetype assumptionCase frequencyHigh / repeated workflowPrimary Power hypothesisCompounding Expertise, potentially reinforcing Network Economies and Process Power.Synthetic case count28 rows

Company basisPublic-company archetype; source verification pending.Archetype assumptionsCanonical metadata is analytical test-fixture metadata, not verified company measurement.Case dataSYNTHETIC ILLUSTRATIVE DATA - Synthetic disputes scorebook, not Casap data

Secondary detail

Economic stakes: High

Transfer potential: Plausibly high but requires evidence

Historical dependence: Potentially important

Expected behavior: Compounding Expertise should be plausible if cross-customer transfer and resistance to compression are demonstrated.

Lab failure condition: If the Lab rejects CE merely because case count is modest, or accepts CE without testing transferability/compressibility, the framework is behaving poorly.

Competing Power hypothesis: Workflow distribution, process execution, or data rights may matter as much as the scorebook.

BOUNDARY TEST

Listen Labs

Canonical question: Is accumulated proprietary knowledge the same thing as accumulated graded expertise?

What this test isolates: When is accumulated knowledge not graded expertise?

Principal decisionResearch insight / recommendation used to support later product, marketing, or strategy decisions.Grade objectivityLow-to-medium / often indirectFeedback speedDelayed and sometimes ambiguousCase frequencyPotentially high research volume, but not necessarily high graded-decision volumePrimary Power hypothesisPotentially valuable proprietary knowledge / filing cabinet rather than a true scorebook.Synthetic case count24 rows

Secondary detail

Economic stakes: Medium / context-dependent

Transfer potential: Uncertain

Historical dependence: Questionable

Expected behavior: The Lab should not infer strong CE merely from large amounts of proprietary research or customer knowledge.

Lab failure condition: If interview/research volume alone creates a strong CE conclusion without an observable decision/outcome/grade loop, the Lab has confused the filing cabinet with the scorebook.

Competing Power hypothesis: Brand, workflow ownership, research distribution, or proprietary access may dominate.

SUBSTITUTION / COMPRESSION TEST

Aaru / model-first research

Canonical question: Can sufficiently capable models or simulation substitute for real-world experience that historically had to be accumulated case by case?

What this test isolates: Can model/simulation capability substitute for accumulated experience?

Principal decisionPrediction/simulation of likely human or market response to a proposed action.Grade objectivityPotentially high when simulation predictions are later compared with actual outcomesFeedback speedDepends on prediction horizonCase frequencyPotentially high through simulationPrimary Power hypothesisModel/simulation capability may compress the value of historical scorebooks.Synthetic case count22 rows

Secondary detail

Economic stakes: Use-case dependent

Transfer potential: Potentially central

Historical dependence: Core question - potentially lower if model priors/simulation substitute for experience

Expected behavior: Increasing base-model/simulation capability should be capable of reducing the modeled advantage of historical experience.

Lab failure condition: If the Lab always rewards the larger historical scorebook and cannot represent model/simulation substitution, the theory test is incomplete.

Competing Power hypothesis: Base model capability, simulation quality, or synthetic data generation may dominate proprietary experience.

ALTERNATIVE POWER TEST

Maybern

Canonical question: Does an AI company need Compounding Expertise at all, or can durable Power reside in deterministic domain infrastructure?

What this test isolates: Can durable Power reside somewhere other than CE?

Principal decisionStructured fund/finance calculation or controlled operational action.Grade objectivityVery highFeedback speedImmediate/shortCase frequencyRepeated but not necessarily strategically important for learningPrimary Power hypothesisDeterministic rails, Process Power, Switching Costs, domain infrastructure.Synthetic case count26 rows

Secondary detail

Economic stakes: High

Transfer potential: Not necessarily the primary strategic variable

Historical dependence: Potentially secondary

Expected behavior: The Lab should be capable of concluding weak/modest CE but potentially strong business Power elsewhere.

Lab failure condition: If weak CE automatically produces a weak-company conclusion, the Lab is improperly treating CE as the only form of Power.

Competing Power hypothesis: Compounding Expertise may be weak/modest while business Power resides elsewhere.

NEGATIVE CONTROL

Creative Marketing Agent

Canonical question: When can enormous volumes of apparently graded data fail to create durable expertise?

What this test isolates: Can huge volumes of apparently graded data still fail to create durable expertise?

Principal decisionCreative / message / audience / campaign allocation choice.Grade objectivitySuperficially measurable but causally noisyFeedback speedFast for clicks/conversions; slower for durable economic outcomesCase frequencyVery highPrimary Power hypothesisPossibly none from CE alone.Synthetic case count32 rows

Company basisEntirely synthetic archetype.Archetype assumptionsCanonical metadata is analytical test-fixture metadata, not verified company measurement.Case dataSYNTHETIC ILLUSTRATIVE DATA - Synthetic creative marketing agent scorebook

Secondary detail

Economic stakes: Medium

Transfer potential: Potentially broad but highly context-sensitive

Historical dependence: Large datasets may exist but contain substantial redundancy

Expected behavior: Large data volume should not automatically produce strong CE when outcomes are confounded, nonstationary, redundant, or easily learned by frontier models.

Lab failure condition: If the Lab declares strong CE primarily from volume and fast click feedback, the analytical framework has failed its negative control.

Competing Power hypothesis: Workflow distribution, brand, media buying economics, or creative process may matter more.

Compare all canonical tests

Test	Canonical question	Grade objectivity	Feedback	Case frequency	Historical dependence	Primary Power hypothesis
POSITIVE TEST Casap	Can repeated, economically consequential, objectively graded decisions create expertise that a capable challenger cannot quickly reproduce?	Relatively high	Relatively fast - days/weeks as an archetype assumption	High / repeated workflow	Potentially important	Compounding Expertise, potentially reinforcing Network Economies and Process Power.
BOUNDARY TEST Listen Labs	Is accumulated proprietary knowledge the same thing as accumulated graded expertise?	Low-to-medium / often indirect	Delayed and sometimes ambiguous	Potentially high research volume, but not necessarily high graded-decision volume	Questionable	Potentially valuable proprietary knowledge / filing cabinet rather than a true scorebook.
SUBSTITUTION / COMPRESSION TEST Aaru / model-first research	Can sufficiently capable models or simulation substitute for real-world experience that historically had to be accumulated case by case?	Potentially high when simulation predictions are later compared with actual outcomes	Depends on prediction horizon	Potentially high through simulation	Core question - potentially lower if model priors/simulation substitute for experience	Model/simulation capability may compress the value of historical scorebooks.
ALTERNATIVE POWER TEST Maybern	Does an AI company need Compounding Expertise at all, or can durable Power reside in deterministic domain infrastructure?	Very high	Immediate/short	Repeated but not necessarily strategically important for learning	Potentially secondary	Deterministic rails, Process Power, Switching Costs, domain infrastructure.
NEGATIVE CONTROL Creative Marketing Agent	When can enormous volumes of apparently graded data fail to create durable expertise?	Superficially measurable but causally noisy	Fast for clicks/conversions; slower for durable economic outcomes	Very high	Large datasets may exist but contain substantial redundancy	Possibly none from CE alone.

Public evidence

Analyze real company

Start from public company information, customer results, product documentation, and explicit evidence gaps. These analyses do not include production case-level data unless a real CaseSet is later imported.

PUBLIC COMPANY ANALYSIS

Casap — Public Evidence Analysis — Sep 2026

PUBLIC SOURCES - NO PRODUCTION CASE DATA

This analysis uses sourced public records and preserves unknowns. It does not create synthetic production cases or reuse the Casap canonical fixture.

CompanyCasap TechnologiesCategoryAgentic AI / dispute management / fraud operations / payment operationsStageSeries AExperience dataNo production CaseSet loadedInitial CE postureStructurally plausible, empirically unprovenPrimary diligence gapLearning causality and cross-customer transfer

Research integrity

Structured assessment, not empirical measurement.

This Lab represents, measures, stress-tests, and can falsify the Compounding Expertise thesis. Scores are explicit judgments with evidence provenance, not validated moat measurements.