CLAIM_LEDGER.yaml

schema_version: 1
protocol_id: AP-V1-PROTOCOL-20260922-R2
claims:
- claim: The frozen Track-A AP-EV challenger did not beat the retained business heuristic.
  supporting_artifact: artifacts/evidence/maven_locked_result.json
  evidence_class: FICTITIOUS_PUBLIC_CRM_NOT_PRODUCTION
  safe_wording: On 403 one-shot LOCKED Maven opportunities, AP-EV captured 24.2% of 60-day realized value at 10% capacity versus 29.2% for the frozen heuristic; the difference was -5.0 points with 95% account-bootstrap CI [-14.5, 8.1], so RETAIN_BASELINE.
  prohibited_stronger_wording: AccountPulse causes revenue lift or is validated for production enterprise deployment.
- claim: Track-A calibration and cold-account promotion support were insufficient.
  supporting_artifact: artifacts/evidence/maven_locked_result.json
  evidence_class: FICTITIOUS_PUBLIC_CRM_NOT_PRODUCTION
  safe_wording: Isotonic improved Brier versus raw XGBoost but had locked calibration slope 0.244, and no cold accounts appeared in LOCKED; G7 failed and G9 was not evaluable.
  prohibited_stronger_wording: Track-A is calibrated across deployment populations or robust to new accounts.
- claim: Implemented an evidence-controlled, CUDA-qualified predictive decision pipeline.
  supporting_artifact: artifacts/bootstrap/post_sync_gpu_smoke.json; artifacts/evidence/accountpulse.duckdb
  evidence_class: LOCAL_SYSTEMS_AND_PUBLIC_BENCHMARK
  safe_wording: Built and tested a reproducible research pipeline with explicit PIT,
    maturity, GPU, and claim controls.
  prohibited_stronger_wording: Production-deployed enterprise revenue system.
- claim: RFM retained over fusion on the UCI decision endpoint.
  supporting_artifact: artifacts/evidence/evaluation_result.json
  evidence_class: PUBLIC_REAL_WHOLESALE_HEAVY_RETAIL
  safe_wording: Fixed fusion reached 57.5% ValueCapture@10% versus 57.3% for RFM,
    but the paired 95% CI for the difference [-1.0%, 1.9%] crossed zero; retained
    RFM.
  prohibited_stronger_wording: Fusion materially increased enterprise value capture.
- claim: Sparse AP-EV did not beat propensity on Olist.
  supporting_artifact: artifacts/evidence/olist_evaluation_result.json
  evidence_class: PUBLIC_REAL_SPARSE_FUNNEL
  safe_wording: Sparse AP-EV captured 10.4% versus 17.4% for XGBoost at 10% capacity;
    the estimated difference was -7.1% with uncertainty crossing zero.
  prohibited_stronger_wording: AccountPulse-EV improved sales caused by outreach.
- claim: Adaptive fusion improved discrimination but not the primary capacity decision.
  supporting_artifact: artifacts/evidence/ranking_metrics.parquet
  evidence_class: PUBLIC_REAL_WHOLESALE_HEAVY_RETAIL
  safe_wording: Adaptive fusion achieved ROC AUC 0.869 and Brier 0.154, while ValueCapture@10%
    was 57.4%.
  prohibited_stronger_wording: Adaptive fusion is the production champion.
- claim: Criteo response targeting and uplift targeting selected different populations.
  supporting_artifact: artifacts/evidence/causal_policy_comparison.parquet
  evidence_class: RANDOMIZED_PUBLIC_CAUSAL_BENCHMARK
  safe_wording: On verified Criteo v2.1, the validation-selected S-learner and response
    model shared 64.2% of their top-10% lists; AIPW incremental visit value favored
    uplift by 0.000933 per eligible user, with 95% bootstrap CI [0.000633, 0.001255].
  prohibited_stronger_wording: Uplift targeting will improve enterprise sales outreach
    or transports to Maven, Olist, UCI, or production populations.