PART 06 — PHASE 2

The Analyst + proposal service.

Keep analysis read-only and recommendations non-executing. A separate, tightly scoped write service may apply only the exact change a named human approved, after revalidating live state and every safety condition.

55–75 Min Model-agnostic Controlled decisions

Reviewed 31 Aug 2026. Educational control pattern, not an instruction to change spend. Advertising APIs, attribution behavior, metrics, and policies change. Validate against current platform documentation, your contracts, and your own economics. All numbers and scenarios here are hypothetical.

Service 03 · Read-only analysis

The Analyst

Reads allowlisted campaign snapshots, checks freshness and completeness, identifies unusual movements, and prepares a cited summary. It has no campaign-write credential and cannot pause, launch, rotate, or alter budgets.

Identity · analytics reader Scope · read only Output · evidence packet Decision · human

An alert is a question, not a verdict

Freshness
Is the data ready?
Report extraction time, timezone, platform delay, attribution window, late conversions, refund/cancellation state, and missing partitions.
Baseline
Is the comparison appropriate?
Use a store-approved baseline that accounts for campaign age, weekday, promotion, inventory, bidding changes, and local seasonality.
Sample
Is there enough information?
Expose denominators and uncertainty. A short window or sparse conversions may support “watch,” not a spend decision.
Alternatives
What else could explain it?
Site outage, tracking loss, inventory, price, creative approval, auction change, payment failure, delivery mix, or ordinary variance.
Economics
Does the metric match the business?
ROAS alone omits margin, returns, COD delivery, discounts, fees, taxes, and incremental demand. Use approved contribution economics.
Calibration
How often is the alert useful?
Track false alarms, missed incidents, operator overrides, and performance by store, platform, objective, and seasonal period.
{ "status": "watch | investigate | insufficient_data", "observed_at": "timestamp + timezone", "data_freshness": {"last_complete_interval": "...", "known_lag": "..."}, "observation": "what changed, with numerator and denominator", "baseline": "approved comparison and why it applies", "uncertainty": "interval, sparse-data warning, or qualitative limitation", "alternative_explanations": ["tracking", "inventory", "auction", "variance"], "sources": ["immutable snapshot IDs"], "suggested_next_check": "read-only verification", "proposed_change": null }
Service 04 · Change proposal

The Proposal service

Turns a complete evidence packet into zero or more bounded options. It cannot execute them. There is no universal ROAS, CPA, pacing, frequency, sample-size, or percentage rule: each store approves decision policy from its data and risk tolerance.

Identity · no write scope Output · proposed diff Default · no change Expiry · required
Make the decision inspectable
  • Tenant, platform account, campaign, field, old value, and proposed value
  • Evidence snapshot IDs, freshness, uncertainty, and alternatives
  • Business rationale, expected observation, cap, and expiry
  • Monitoring window, stop conditions, and rollback value
  • A unique idempotency key and proposal hash
Separate advice from authority
  • Approve itself or treat a model score as approval
  • Bundle unrelated campaigns, fields, stores, or destinations
  • Use stale attribution or assume a causal explanation
  • Change spend, status, bid, creative, audience, or schedule
  • Reuse approval after any bound value or live state changes

The exact approval envelope

{ "tenant_id": "tenant_demo", "platform_account_id": "acct_demo", "resource_id": "campaign_demo", "field": "daily_budget", "expected_current_value": "HYPOTHETICAL_OLD_VALUE", "approved_value": "HYPOTHETICAL_NEW_VALUE", "absolute_cap": "STORE_APPROVED_CAP", "valid_after": "2026-08-31T10:00:00Z", "expires_at": "2026-08-31T10:10:00Z", "proposal_hash": "sha256:...", "idempotency_key": "single-use-random-id", "approver_id": "named-human", "rollback_value": "HYPOTHETICAL_OLD_VALUE" }

A separate write service, closed by default

1 · Authenticate
Use a dedicated, least-privilege write identity
Store its secret in a managed secret service. Isolate it from the model runtime and analytics reader. Enforce tenant and account allowlists server-side.
2 · Verify approval
Validate signature, approver authority, hash, scope, and expiry
Reject missing, altered, broad, already-used, future-dated, or expired approvals. A chat reaction alone is not a sufficient authorization artifact.
3 · Re-read live state
Compare the current value with the approved precondition
Abort if campaign status, budget, currency, objective, schedule, tenant, account, cap, policy state, or data health differs from what the human reviewed.
4 · Apply once
Send only the approved field and value
Use the idempotency key, a short timeout, bounded retries for safe failure modes, destination rate limits, and no model-generated API payload at execution time.
5 · Verify and monitor
Read back, log, observe, stop, or roll back
Append request/response hashes, actor, timestamps, before/after state, and correlation ID to an immutable audit log. If verification or a stop condition fails, disable further writes and execute the preapproved rollback path.

A proposal suggests a small budget change after a metric moves. Before execution, the current-state check finds that an operator already paused the campaign. The write service rejects the stale approval, records PRECONDITION_FAILED, alerts the operator, and performs no write.

The drill passes only if replaying the same request also performs no write and the campaign remains unchanged.

Pre-production evidence gates

Gate
Evidence
Release rule
Synthetic
Tenant confusion, stale approval, forged signature, cap breach, replay, rate limit, timeout
Every unsafe case fails closed
Shadow
Proposals compared with real human decisions; no write credential present
Representative error and override analysis reviewed
Canary
One approved tenant/account, narrow field set, small business cap, staffed monitoring
Security and business owners sign the release
Recovery
Kill switch, credential revocation, stop-rule alert, rollback, reconciliation, incident communications
Drill completed and evidence retained

Budget and campaign writes can create immediate financial and reputational harm. Keep human approval for every production change. If the approval service, audit log, fresh-state read, platform API, data-quality check, or monitoring channel is unavailable, the safe result is no change.

“Reversible” does not mean harmless: auction dynamics, missed demand, learning state, customer experience, and already-spent budget may not be recoverable. Test rollback and containment before enabling the first write.

← Part 5 — The Creative Part 7 — Audience + Reporter →