CHALLENGES: The selected dataset’s 20% pattern-overlap proxy is below the approximately 50% pooling threshold. Selective-minus-local value ranges from -0.009 to 0.054 across the tested boundary; material benefit begins around 50% shared structure. Works when: Shared decision structure is at or above roughly 50% under the displayed assumptions. Fails when: Unconditional pooling can destroy value through approximately 25% shared structure.

Treat cross-customer learning as a conditional mechanism with an estimable transfer boundary, not as an automatic consequence of having more customers. Next experiment: Estimate shared structure and segment-level harm on newly held-out customers using the selected dataset’s actual features and actions.

Limitation: Automated sensitivity results are conditional simulations calibrated from the selected dataset’s size, group structure, missing feedback, and descriptive time shift. They identify mechanism boundaries; they do not convert simulated outcomes into observed company evidence.

28/59 decision-class/case-type patterns occur across segments; 225/291 eligible cases belong to shared patterns.

Measures whether segments share potentially reusable decision structure.

Limitation: Shared structure is a prerequisite for transfer, not its performance result.

No material pooling advantage. 88 held-out cases; local Brier 0.246, pooled 0.243.

Segment transfer probe v1: earliest 70% train, latest 30% test; only feedback available before each test decision is used. Predict CORRECT versus other resolved grades from decision class and case type. Same fixed smoothing in each arm; lower Brier score is better. A 0.005 Brier difference defines the descriptive effect band, not statistical significance.

Limitation: Groups are customer segments, not identified customers. This tests transfer of grade prediction; improving chosen actions requires an action-policy experiment.

Grade prediction tested; decision-policy improvement remains untested.

Use identified customers and compare local versus pooled policies on held-out decisions, measuring correctness and economic outcomes.

Limitation: Missing evidence cannot support the proposition.

Reconstruction has not been tested: no challenger with restricted data access or a calibration budget has been evaluated.

The grade-prediction models share historical data access. Their scores cannot establish how cheaply a competitor could reproduce the capability.

Limitation: Compare against a challenger restricted to public features, rules and a fixed calibration budget.

SUPPORTS: The challenger retains a 0.014 value gap after 500 calibration cases. The incumbent gap changes from 0.023 with no calibration to 0.014 at 500 cases. Works when: The challenger remains below the incumbent throughout the tested budget. Fails when: The failure boundary exceeds the tested 500-case budget.

The simulated experience advantage resists the tested calibration budget, but public priors and richer challenger strategies remain untested. Next experiment: Run a restricted-access challenger on actual policy artifacts, public data, synthetic cases, and a fixed labeled calibration budget.

Limitation: Automated sensitivity results are conditional simulations calibrated from the selected dataset’s size, group structure, missing feedback, and descriptive time shift. They identify mechanism boundaries; they do not convert simulated outcomes into observed company evidence.

Relearning difficulty: UNKNOWN

Frames the rebuildability hypothesis.

Limitation: Assumption is not a challenger benchmark.

No sourced contractual rights evidence.

Need retention, derived-feature, evaluation, and cross-customer training rights.

Limitation: Missing evidence cannot support the proposition.

No compression/relearning benchmark.

Need evidence that a capable challenger cannot reproduce performance cheaply.

Limitation: Missing evidence cannot support the proposition.

291/300 cases graded; 291/300 outcomes represented; median feedback 23 days.

Supports graded experience capture within the selected dataset.

Limitation: Still does not prove that captured grades improve future behavior.

Purpose-built AI agents and dispute-management software for banks, credit unions, and fintechs; public materials describe dispute model/API, win-score prediction, fraud functionality, and built-in Reg E, Reg Z, Nacha, and card-network expertise.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: MEDIUM. Review source scope before treating as decisive.

$25M Series A announced August 7, 2025; $33.5M total funding; AI agents analyze evidence, predict outcomes, and automate lifecycle from intake to chargeback filing; proprietary first-party fraud score; company-reported customer results.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: MEDIUM. Review source scope before treating as decisive.

Company-published customer story reports estimated $875K annual savings, resolution time improved from 90 days to 12 days, 95% chargeback win rate, writeoffs down 72%, costs down 85%, and monthly capacity from 1,200 to 4,000 transactions.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: MEDIUM. Review source scope before treating as decisive.

Company-published customer story reports fraud loss reduced 51%, cost per dispute reduced 90%+ from $37, chargeback floor reduced 85%, real-time merchant-response/chargeback tracking, and Visa transaction metadata used in fraud investigation.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: MEDIUM. Review source scope before treating as decisive.

Filene Research Institute public report summary describes testing Casap's fraud dispute platform in sandbox and live environments, with participating credit unions reporting usability, reduced manual effort, and confidence supporting compliance/dispute outcomes.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: HIGH. Review source scope before treating as decisive.

Filene reports testing Casap with five credit unions; one institution had been live on the platform for more than a year; staff reported a 122% improvement in ease of use versus prior systems; reported dispute-processing error frequency declined 63%.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: HIGH. Review source scope before treating as decisive.

Company funding announcement claims Casap's platform becomes more effective at fraud detection with customer interaction.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: MEDIUM. Review source scope before treating as decisive.

Privacy policy says Casap may use collected information to develop/improve products and conduct analytics, including aggregated non-specific information. It also says third-party personal data uploaded through the service is processed on behalf of customers and according to their instructions.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: MEDIUM. Review source scope before treating as decisive.

Primary Venture Partners public materials identify Casap in its portfolio/insights context and state Primary led the seed round and invested in every round.

Public-source evidence record for the Casap Public Evidence Analysis. Treat as sourced context for diligence, not production case-level evidence.

Limitation: Confidence: LOW. Review source scope before treating as decisive.

Validation selected base-rate; final-test Brier 0.246 versus base-rate 0.246 on 59 cases. The final-set leader is blended (0.223): the ranking changed across windows. The validation-selected model is retained, not replaced after seeing test outcomes.

Tests whether past experience predicts grades on later cases. It does not measure improvement from choosing different actions.

Limitation: Compare decisions made with and without the selected learner on a new evaluation batch.

Correct-grade rate: early 0.621, late 0.576.

Descriptive time shift may reflect case mix, policy or grading changes; it is not attributed to learning.

Limitation: Compare matched decision types across time and identify policy/model versions.

291/300 graded cases.

Grades exist to learn from if connected to updates.

Limitation: Grade coverage does not establish that grades improve future decisions.

With a fixed pooled estimator and evaluation set, Brier 0.224 → 0.217 → 0.215 → 0.224 as training grows through 58 / 116 / 174 / 232 cases. The learning curve reverses direction: more history is not consistently better.

A fixed-test learning curve measures whether additional historical cases help this estimator; results may be non-monotonic.

Limitation: Repeat on a fresh time window and examine changing case mix.

SUPPORTS: Full history remained competitive throughout the tested drift range. The best selected-history gain versus full history moves from -0.009 at zero drift to 0.009 at maximum tested drift. Works when: Full accumulated history remains useful across the tested range. Fails when: A staleness boundary was not reached in the tested range.

The learning system needs explicit recency and selection logic; accumulating every case is not itself the advantage. Next experiment: Backtest rolling time windows on the selected dataset and compare stable versus rapidly changing decision classes.

Limitation: Automated sensitivity results are conditional simulations calibrated from the selected dataset’s size, group structure, missing feedback, and descriptive time shift. They identify mechanism boundaries; they do not convert simulated outcomes into observed company evidence.

300 cases. HIGH DIVERSITY; MODERATE REPETITION; NOVELTY DECLINING.

Pattern diversity and cohort novelty describe the supply of new information available to learn from.

Limitation: A held-out learning curve is needed to connect novelty to performance gain.

Update: UNKNOWN; deploy: UNKNOWN

Closed-loop architecture is relevant but does not prove difficult-to-copy Process Power.

Limitation: This is deterministic Power Map input, not a standalone proof of durable Power.

UNKNOWN

No evidence currently shows that grades alter future model or policy behavior.

Limitation: Missing evidence cannot support the proposition.

UNKNOWN

No evidence currently shows learned improvements reach production.

Limitation: Missing evidence cannot support the proposition.

No controlled before/after learning comparison.

Compare held-out decisions with and without updates from accumulated grades.

Limitation: Missing evidence cannot support the proposition.

No compression/relearning benchmark.

Need evidence that a capable challenger cannot reproduce performance cheaply.

Limitation: Missing evidence cannot support the proposition.