/deepline-scoring

name: deepline-scoring
disable-model-invocation: false
description: 'Use when discovering niche signals, auditing ICP or won/lost evidence, rescoring accounts, or building account and lead scoring Plays. Triggers on fit scoring, engagement scoring, backtesting, holdout evaluation, predictive lift, external proxies, and scoring leakage. Skip pure outreach copy or contributor skill installation tasks.'

Deepline Scoring

Quick Start

npm install -g deepline
deepline -h

Find evidence for the customer's decision. Use approved rules or a separately evaluated model for scoring. Phrase matches and prevalence ratios alone cannot supply scoring weights.

For account-fit and propensity work, always collect relevant website language, materialize account-level phrase/concept features, and show their model role and score contributions. Industry, size and geography alone are a baseline. If website evidence or approved scoring mappings are missing, disclose the gap rather than silently substituting that baseline. Follow [website account features](references/account-website-features.md); preserve an approved incumbent when auditing or replaying it.

| Task | Read | | -------------------------------------- | ------------------------------------------------------------------------------------------------------------------ | | Find phrases and buyer language | [Keyword catalog](references/keyword-catalog.md), [buyer-language research](references/buyer-language-research.md) | | Build account-level website features | [Account website features](references/account-website-features.md) | | Extract typed flags, choices or scores | [Jev feature extraction](references/jev-feature-extraction.md): `scripts/jev_features.py` | | Build or audit a score | [Scoring delivery](references/scoring-delivery.md) | | Test rules or evaluate outcomes | [Testing and evaluation](references/testing-and-evaluation.md) | | Backtest or score the full population | [Backtesting task](references/backtesting-task.md) | | Create an artifact-backed scorecard | [Scorecard creation pattern](references/scorecard-creation.md) | | Train a lift or propensity scorecard | [Scorecard training](references/scorecard-training.md): `scripts/train_scorecard.py`, `scripts/scorecard/` | | Pull accounts and sources | `scripts/hubspot_accounts.py` (HubSpot labels), `scripts/collect_sources.py` (website + jobs via Deepline tools) | | Debug disputed features or routing | [Scoring diagnostics](references/scoring-diagnostics.md) | | Verify technology | [Technology evidence](references/technology-evidence.md) | | Estimate staffing or demand | [Capacity evidence](references/capacity-evidence.md) |

Read `deepline-gtm` before collection and `deepline-plays` before authoring. Verify the workspace, current provider schema and price. Pilot one or two rows and pass the [quality gate](references/quality-gate.md) on the generated outputs before scaling within the approved budget. Keep exports and receipts in a persistent project directory. Reuse collected evidence when rescoring.

Workflow

1. Define the product, decision date, prediction horizon, population, analysis unit and requested outputs. Before paid collection, identify the executable model and independent expected scores for parity, or historical evidence and untouched labels for predictive evaluation. Missing prerequisites leave that test blocked; an authorized research run can still proceed. Keep fit, engagement, capacity and coverage separate. Use only requested dimensions: `account_fit`, `account_engagement`, `lead_fit`, `lead_engagement`. A combined priority policy must preserve its components. 2. Resolve identities, parent groups and conflicting outcomes. Split discovery and validation by time and parent before selecting features. Preserve the full requested population. Open accounts and random alternatives have unknown outcomes; lookalikes are not wins. Separate acquisition, renewal and expansion. 3. Research workflows, problems, roles, systems and counterexamples across relevant public sources. Follow observed buyer phrases and source URLs. Mine discovery documents without labels, then review concepts and aliases against their evidence. Keep rare and inconclusive candidates. Cover billing models, revenue types and evidence for B2B/B2C/mixed customer models alongside product-specific concepts. Materialize matches, nonmatches and unknowns for every account; cohort summaries are not an account feature table. Extract website and business-model judgments as typed Jev features ([Jev feature extraction](references/jev-feature-extraction.md)) rather than free-form LLM prompts or keyword hits alone; count roster titles with deterministic regexes. Exclude report prose and outcome summaries from the corpus. 4. Check collection with the [quality gate](references/quality-gate.md). Report coverage by source and outcome before citing lift. Compare a coverage-only model and evaluate features where both outcomes have adequate observed data. Imputation or dropping missingness flags can still encode collection bias. Keep failed, partial and empty results distinct. 5. Check what each feature measures. Reviews are not calls, openings are not hires, and software mentions or portal links d