MGID Source Auditor

Audits selective bidding entities and recommends blocks or QF changes. All rules, thresholds, and impact formulas live in this skill (and reference.md).

Tools: tool reference
Assumptions: threshold assumptions

Prerequisite: tracking gate pass.

Required inputs

Input	Required	Default
Campaign ID(s)	Yes	simulator scope list, user ids, or all active when run standalone
Date range	Yes	last 7 days (max 90)
Target CPA	Yes	`get_campaign` → `bidding.price` if strategy=cpa, else ask user

MCP tools

Tool	Purpose
`get_campaign`	`sourcesOptimization`, `bidding.strategy`, `bidding.price`, type (push?)
`get_selective_bidding_statistics`	Performance by entity
`get_quality_factors`	Current QF for proposed from/to values
`get_source_filters`	Existing blocks (avoid duplicate recs)

Campaign type gate (before anything else)

type in (content, product) → audit normally
type == push               → audit with entity: providers
type in (search_feed, newsletter) → SKIP, report as unsupported

search_feed and newsletter campaigns have no source/widget entities, so selective-bidding rules are meaningless there. List them in the output as skipped with the reason — never silently drop or silently audit them. See the campaign-type table in tool reference.

Entity selection

sourcesOptimization == true  → entity: sources
sourcesOptimization == false → entity: widgets
campaign type push           → entity: providers

Never fetch sources and widgets for the same campaign in one audit pass unless comparing explicitly.

Bidding strategy → action

`bidding.strategy`	Poor performer action
`cpc`	ASSUMPTION: QF −15% (max −25%) when CPA > 1.5× target
`cpa` (CPA Tune)	ASSUMPTION: block when CPA > 2× target — unless learning phase
`maxConversion`	ASSUMPTION: block if CPA > 3× target; else QF −10%

CPA Tune learning phase

Before recommending blocks on cpa campaigns, check assumptions.md. If learning → suppress with cpa_tune_learning; emit monitor alert instead.

Minimum data (ASSUMPTION)

Clicks ≥ 50 OR conversionsBuy ≥ 5
Spend ≥ $10 for block actions

Poor performer triggers (ASSUMPTION)

Condition	Action
CPA > 2× target (CPA Tune, post-learning)	block
CPA > 1.5× target (manual CPC)	adjust_qf
Spend ≥ $50, conversionsBuy = 0, tracking OK	block or QF
Top-3 converter by volume + CPA ≤ target	skip block

Full decision tree: reference.md

Impact estimation

Formulas in reference.md — all ASSUMPTION.

Output

Chat: entity counts + top N actions table.

Return: Recommendation[] (block_source, block_widget, block_provider, adjust_qf).

Apply

Action	Tool
QF change	`update_quality_factors`
Block/unblock	`update_source_filters`

Write scope required. No batch-* tools.

Source Auditor — Reference

Assumptions registry: threshold assumptions

All thresholds below are ASSUMPTION until backtested.

Decision tree

tracking gate pass?
  no → suppress
min clicks/spend met?
  no → suppress (insufficient_data)
cpa strategy + learning phase?
  yes → suppress block (cpa_tune_learning); alert only
bidding.strategy == cpc?
  CPA > 1.5× target → adjust_qf −15%
bidding.strategy == cpa? (post-learning)
  CPA > 2.0× target → block
maxConversion?
  CPA > 3.0× target → block; else CPA > 1.5× → adjust_qf −10%

Top converter downgrade

Sort entities by conversionsBuy desc within campaign.

Top 3 AND CPA ≤ target → skip block
Top 3 AND CPA > target but < 3× target → adjust_qf −10% instead of block

Field mapping (`get_selective_bidding_statistics`)

Metric	Fields
ID	`sourceId` / widget uid in response
Name	`sourceName`
Spend	`spent.amount`
Clicks	`clicks`
Conversions	`conversionsBuy`
CPA	`conversionsCostBuy.amount`
ROAS	`roas`

Impact formulas (ASSUMPTION)

Block:

wasted_share = clamp(1 - target_cpa / actual_cpa, 0, 0.95)
daily_savings = (entity_spend_7d / 7) × wasted_share
range = [daily_savings × 0.85, daily_savings]

QF −N%:

spend_delta ≈ -(entity_daily_spend × N / 100)

Confidence matrix (ASSUMPTION)

Case	Confidence	Risk
Block, clicks ≥ 100, CPA > 2×	High	Low
Block, clicks 50–99	Medium	Low
QF −15%	Medium	Low
Block during CPA Tune learning	—	Suppressed