Pre-send deliverability scoring AI. Latin mailbox tuned. Trained on 11 million B2B emails.

"What's the model card publication standard? Is it actually inspectable or marketing fluff?"

Model card v3.4.2 documentation follows Mitchell et al. 2019 model card standard plus Arnold et al. 2019 factsheet additions. The card covers nine sections per the standard: model details (architecture, version, owner, license, intended use), intended use cases plus out-of-scope use cases, factors evaluated (geographic, vertical, sender size cohort), metrics (precision, recall, F1, AUC-ROC, Brier score plus per-dimension breakdown, plus disaggregated metrics across factors), evaluation data (corpus composition, label distribution, validation methodology, temporal hold-out specification), training data (corpus stratification, class imbalance handling, feature engineering approach), quantitative analyses (calibration, fairness across cohorts, performance under distribution shift), ethical considerations (Latin mailbox provider feature engineering disclosure, English-language bias acknowledgment, Spanish dialect coverage, training data privacy posture), caveats and recommendations (limitations, when to use advisory mode, when to override, when to escalate to manual review). The card is available under NDA on Pro and Enterprise tiers because the corpus stratification details could permit adversarial gaming; the architecture, validation methodology, and calibration metrics are publicly documented in this page for evaluation purposes. Independent ML reviewer access is available on Enterprise tier with mutual NDA.

"How do you handle model drift? Mailbox provider behavior changes constantly."

Model drift management operates across three timescales. Timescale 1 within-week behavioral shifts: dimension-level feature engineering hotfixes deployed without full retrain when new mailbox provider behavior is detected; the v3.3 release in March 2026 was a hotfix-style deployment of the AI-similarity feature in response to the Google February 2026 AI spam update detected within 6 weeks of the update going live. Timescale 2 model retraining cadence: full retrain approximately every 8 weeks against the latest corpus window, with validation metrics published per release version. The 8-week cadence balances retraining cost against drift accumulation; faster cadence increases compute cost without proportionate benefit, slower cadence lets drift accumulate beyond acceptable threshold. Timescale 3 architecture refresh: architecture changes (e.g., switch from LightGBM to XGBoost, addition of attention mechanism in neural network, ensemble weight rebalancing) deployed approximately annually based on systematic A/B testing of architecture variants; v3.0 to v3.4 represents one architecture generation with v4.0 currently in development for Q3 2026 deployment. Model performance monitoring operates continuously: drift detection alerts trigger when validation precision drops below 92 percent or AUC-ROC drops below 0.94 on rolling 4-week window; the alerts trigger emergency retraining outside the 8-week cadence when triggered. Two such alerts have triggered emergency retraining since v3.0 launch (March 2025): once in October 2025 due to Microsoft Outlook reputation system update, once in February 2026 due to Google AI spam update.

"What's the false positive cost? If the Scorer hard-blocks a legitimate campaign, what happens?"

False positive cost is real and managed at three layers. Layer 1 calibrated threshold: default threshold 65 is calibrated to keep false positive rate under 4 percent of all campaigns scored at default threshold; 7 percent total catch rate at hard-block on Starter covers both true positives (campaigns that would have damaged reputation) and false positives (campaigns that would have been fine if sent). The 4 percent false positive rate matches the precision metric (94.7 percent precision means 5.3 percent of flagged campaigns are false positives among the flagged set). Layer 2 manual override on Pro and Enterprise: senders can override soft warning with explicit risk acknowledgment when they have judgment that the Scorer false-positived their campaign. Override is logged to audit trail and counts against future threshold calibration; the audit trail surfaces senders with high override rates for review. Layer 3 escalation to manual review on Starter: senders without manual override authority can escalate to EMP support for manual review of hard-blocked campaigns. Manual review turnaround is 4 business hours typical, 24 business hours maximum. About 12 percent of hard-blocked campaigns escalate to manual review; 73 percent of escalated campaigns are released after manual review, 27 percent are confirmed as risk and the sender modifies before send. The false positive cost is the friction in escalation workflow plus the time-to-send delay; the cost is meaningful but smaller than the cost of reputation damage from missed true positives.

"How does the Scorer handle multi-language campaigns? Mixed Spanish and English content?"

Multi-language handling operates at three levels. Level 1 language detection: every campaign gets language classification before content embedding, with confidence threshold 0.85 required for single-language treatment. Below threshold (multi-language content detected), the content gets segmented into language-homogeneous chunks for separate scoring then aggregated. Level 2 per-language model weighting: the content embedding network was trained with weighted Spanish-English dual-language corpus (47 percent Spanish, 38 percent English, 15 percent multi-language). Spanish content scoring uses Latin Spanish vocabulary patterns (73 percent Latin weighting in Spanish corpus); English content scoring uses standard US English patterns. Level 3 multi-language penalty: campaigns with significant code-switching (Spanish-English mixing within single message) receive a small additional risk score because mixed-language content has higher historical bounce rates in B2B Latin segments (recipients often filter mixed-language as automated translation spam). The penalty is calibrated at 3-5 points addition to base score; senders running intentional bilingual campaigns can override the penalty on Pro and Enterprise tiers. About 14 percent of EMP campaigns are detected as multi-language; mostly Spanish-English mixing in Latin B2B campaigns where technical terminology defaults to English (cloud, SaaS, API, dashboard, etc.).

"What's the data privacy posture for the training corpus? Are tenant emails in your training data?"

Training corpus privacy posture documented per ML transparency standards. Three categories of data in the corpus. Category 1 metadata-only signals: bounce rate, complaint rate, engagement rate, mailbox provider response codes per delivery, IP reputation trajectory, list freshness statistics, time-of-send distribution. Metadata-only signals are aggregated across campaigns and do not contain message content or recipient PII. Category 2 anonymized content embeddings: content from campaigns sent through EMP infrastructure is processed through embedding generation (subject line embeddings, content structure embeddings, link density patterns) and only the embeddings are retained for training. Original content text is not stored in training corpus after embedding generation. Tenant identification is removed from embeddings before corpus merge. Category 3 explicitly-opted-in benchmark data: a subset of tenants (currently 47 tenants representing 14 percent of corpus volume) have explicitly opted into benchmark data sharing as part of their tenant agreement, contributing labeled examples for model improvement in exchange for early access to new model versions and per-tenant performance reporting. Opt-in is reversible at any time; data contributed before opt-out remains in corpus per the agreement terms but no new data is contributed post opt-out. Tenants that have not explicitly opted in contribute only Category 1 metadata signals (no content, no PII). The corpus privacy posture is documented in the model card under NDA and reviewable by tenant compliance counsel during procurement.

"Can we run the Scorer on our own infrastructure? Self-hosted deployment?"

Self-hosted deployment is not supported. The decision is structural rather than commercial. Reason 1 model size and infrastructure dependency: the Scorer model plus feature engineering pipeline plus inference infrastructure exceeds 8 GB compressed and requires GPU acceleration for sub-second latency at production volume; the deployment footprint is non-trivial for clients without dedicated ML serving infrastructure. Reason 2 model maintenance lifecycle: the 8-week retraining cadence plus hotfix feature engineering deployments require continuous model versioning, deployment automation, rollback capability, and validation metrics monitoring; self-hosted deployment would require client-side investment in MLOps infrastructure that EMP centralizes for operational economy. Reason 3 training corpus locality: the model trains on EMP portfolio delivery telemetry; self-hosted deployment would either require shipping the corpus (unwieldy and creates corpus sprawl) or training a new model on client-side data alone (degrades performance compared to the centralized corpus that aggregates 11M events across the portfolio). The closest pattern to self-hosted is the Enterprise tier hybrid pattern: own MTA infrastructure, EMP-hosted Scorer API, the API call adds round-trip latency (typically 200-400ms over the inherent 643ms Scorer compute time, so 800-1100ms total) but preserves the centralized model benefits. About 3 percent of Enterprise tier clients use the hybrid pattern; the rest use platform-native deployment for the lower latency.