Ad account auditor

Ad Account Auditor

Audit one paid-media account or portfolio for incremental contribution and operating quality under declared constraints. Platform-reported ROAS is one input, never the objective or truth set by itself.

When This Must Trigger

• Before launching, materially increasing spend, or changing a risky bid/targeting strategy.
• When tracking, attribution inflation, unsafe placements, claims, or wasted spend are in doubt.
• When the user requests a ROAS/RQS account audit from their exports.

Quick Start

Audit this USD account for direct response using 7-day click, 3-day lag, and $120 CAC ceiling.
Run the incremental-profit profile against the holdout and order-ID exports.

Skill Contract

Reads: one normalized account/portfolio evidence set. Writes: only a permissioned v3 artifact. Done when: required context and all 20 states are explicit, vetoes use verified evidence, and scorer output is reported without executing spend changes.

This skill judges. conversion-signal-qa, attribution-reconciler, campaign-architect, ad-creative-builder, and budget-pacing-monitor build/fix the inputs. Never enable campaigns, change bids, upload audiences, or scale budgets without separate explicit approval.

Data Sources

Need	Preferred evidence
Delivery/spend	Campaign, query, placement, audience, and change-history exports
Outcome truth	Deduplicated order/lead IDs from ecommerce, analytics, or CRM
Economics	Currency, margin/contribution, CAC/payback constraint
Attribution	Platform + own-data timestamps/IDs, normalized windows and lag
Safety/claims	Placement report, rendered ad/landing, approved claim/disclosure state from offer-claims-registry (the paid claims SSOT)
Incrementality	Holdout/geo split/causal test, otherwise explicitly labeled proxy

Instructions

Runtime and Setup

Read ../../../references/auditor-runbook.md, scoring-semantics.md, roas-benchmark.md, and the ROAS catalog entry. Standalone installs use bundled immutable references/auditor-runtime.md; never fetch mutable main. Before deterministic calls, follow runtime-invocation.md, resolve AARON_SKILLS_ROOT="${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || true)}", and require the scorer, validator, and typed catalogs. If unavailable, return score_state: NOT_SCORED / score_confidence: not_scored with no gate verdict or persistent artifact.

Declare profile (direct-response|prospecting|incremental-profit), target, currency, attribution window, conversion lag, business constraint, goal, and observation date. If any required context is missing, return NEEDS_INPUT/UNDECIDED.

Evidence and Scoring

1. Normalize currency, windows, IDs, lag, and portfolio scope before comparing metrics.
2. Score all 20 R1..S5 criteria from the benchmark with source/date/type/confidence.
3. Use Unknown for missing own-data truth, placement exports, or reconciliation. No data is not a veto and cannot be N/A merely because access is inconvenient.
4. Verify vetoes:
   • ROAS-R1: instrumentation demonstrably fails the named own-data truth set.
   • ROAS-R2: material double-counting/inflation is demonstrated.
   • ROAS-O1: material claim/disclosure failure against the offer-claims-registry approved state.
   • ROAS-O2: applicable platform/restricted-category violation.
   • ROAS-A1: placement evidence demonstrates a material safety breach.
5. Run the typed scorer. Report estimated/proxy incrementality as such; do not call platform attribution causal.

§2 ROAS Worked Examples

• Complete direct-response profile, raw 78, no veto/fail: DONE/SHIP, final 78.
• Complete profile, raw 78, one verified R1 failure: DONE_WITH_CONCERNS/FIX, final 59.
• Complete profile, verified R1 and R2 failures: DONE/BLOCK, raw retained, no final score.
• Missing placement report: A1 Unknown, NEEDS_INPUT/UNDECIDED, no overall score.

§3 ROAS Guardrails

• High reported ROAS can reflect under-spend, branded-demand capture, or attribution inflation.
• Learning-phase disruption is an S2 finding, not an automatic veto.
• ATT/modeled data may reduce confidence; it does not automatically fail R1.
• Frequency, creative fatigue, and audience saturation require separate evidence.
• Never compare cross-platform returns before normalizing currency/window/lag and deduplicating outcomes.

§5 ROAS Translation

Lead with business impact and evidence. On trace request, qualify ROAS-R1/R2/O1/O2/A1; do not expose bare IDs that collide with RAMP/ECHO/TALE.

Report and Verdict

Begin with the auditor-runbook's exact typed conversation header. Never replace status, verdict, or score_state with prose; list each explicitly missing qualified item as ``ID: `unknown``` before findings.

Show verdict, profile/context, score or coverage/interval, confidence, R/O/A/S detail, reconciliation table, verified critical controls, Unknown evidence, and prioritized fix/owner/rerun condition. The scorer owns status/verdict and the 59 ceiling.

Validation Checkpoints

• Scope/currency/window/lag/constraint/goal are explicit.
• Own-data outcome truth is separated from platform self-report.
• All 20 items have valid states and provenance; Unknown is not renormalized.
• Veto failures are positively verified.
• No spend/account mutation occurred without separate approval.

Persistence

Persist only after explicit authorization to memory/audits/ad/YYYY-MM-DD-<topic>.md. Assemble and validate the complete v3 draft with validate-audit-artifact.py against that intended --relative-path, persist only through one full-content Write, then revalidate the target as required by the auditor runbook. Edit/shell/MCP mutations of the reserved sink are unsupported. Do not autonomously write hot cache, claims, candidates, or account state.

Reference Materials

• ROAS benchmark
• Measurement protocol
• Auditor runbook
• Scoring semantics

Next Best Skill

• Tracking: conversion-signal-qa
• Claims/disclosures: offer-claims-registry — the approved claim/disclosure state behind ROAS-O1
• Attribution: attribution-reconciler
• Structure/audience: campaign-architect
• Pacing: budget-pacing-monitor