BossAgent Glossary & Activation Formulas

Version: 1.0.0 Last Updated: February 5, 2026 Scope: Activation logic terms, causal metrics, confidence weighting, SQL/Cypher reference

Glossary (Plain English)

Term	Symbol	Definition
Conversion probability	`p_conv`	Probability a user converts (baseline or with a treatment).
Incremental lift	`delta_p_conv` (`Dp_conv`)	Incremental lift in conversion probability from a treatment vs. its baseline/control.
Path ROI	`ROI_path`	Return on investment for a specific journey/path (touchpoints from first exposure to conversion).
Incremental path ROI	`DROI_path` (`DROI_path`)	Incremental ROI of a path vs. its baseline (e.g., removing one edge/touch or vs. control).
Edge-level uplift	`edge_uplift`	The incremental lift attributed to a single edge (touchpoint -> next touchpoint or touchpoint -> conversion) holding the rest of the path constant.
Confidence weight	`w_conf`	A 0-1 weight capturing certainty (sample size, variance, recency, coherence, etc.).
Activation score	`activation`	The gated score BossAgent uses to approve actions (budget shifts, bids, status) from ROI and confidence.

Dp_conv = p_treatment - p_control

edge_uplift(A->B) = Dp_conv estimated for that edge
                     (marginal, ablation, Shapley, or causal)

ROI_path = (Revenue_path - Cost_path) / Cost_path

DROI_path = ROI_path(treatment) - ROI_path(control)

Punishes weak links by using a geometric mean with tunable alpha weights:

w_combined = exp( SUM_i( alpha_i * ln(max(w_i, epsilon)) ) )

where SUM(alpha_i) = 1

Default alpha weights:

activation = max(0,
  (beta1 * DROI_path + beta2 * edge_uplift + beta3 * Dp_conv)
  * w_combined
  - policy_threshold
)

Default coefficients:

Param	Default	Description
`beta1`	0.60	Weight on incremental path ROI
`beta2`	0.25	Weight on edge-level uplift
`beta3`	0.15	Weight on conversion lift
`policy_threshold`	0.60	Minimum gated score to act

Stability first: Compute delta metrics on rolling windows (e.g., 14-28d) and add minimum support filters (n >= X).
Guardrails: Require w_combined >= c_min and cap per-step changes (+/-15%) when activation > 0.
Explainability: Persist {d_roi, d_edge, d_p, w_combined, threshold} with each decision for audit/UI chips.

File	Purpose
`data-pipelines/dbt/models/activation_score/mart_activation_score.sql`	BigQuery view computing activation_score end-to-end
`miz-oki-adk-agents/boss/config/neo4j_activation_loader.cypher`	Neo4j script to store and compute activation scores on graph edges
`miz-oki-adk-agents/boss/config/kg_guardrails.yaml`	Guardrail rules referenced by activation gating