Budget Optimization

For Sponsored Products budget-change causal review, base/effective daily budgets must be observed finite nonnegative values and the average-daily-budget policy must be known. Missing, Unknown/Unsupported, boolean, negative or non-finite values cannot establish a stable control or Treatment Isolated; do not infer zero from missing connector evidence.

Connector capability gate

When a conclusion depends on live MCP/API/connector data, load ../../references/connector-capability.md before interpreting missing or empty fields. If a machine-readable capability snapshot is available, identify the exact required capability IDs and run ../../scripts/evaluate_connector_capability_gate.py before metric interpretation.

• Pass — continue with the Skill's normal evidence, sufficiency, lineage and safety checks.
• Degraded — do not make a high-confidence dependent recommendation; keep the result to Directional, Hold, Alternate Source, Missing Data, or Manual Review.
• Blocked — do not treat the dependent observation as action-safe until the capability is resolved or an allowed alternate source is verified.
• missing_evidence_policy = never_zero — Partial, Unsupported, Unknown, or absent connector capability is never a numeric zero, unchanged state, or proof that the platform lacks the feature.

This gate is read-only and does not authorize live Amazon Ads mutation.

Campaign objective gate

Before reallocating budget based on apparent efficiency, load ../../references/campaign-objective.md when campaign roles differ. Cross-campaign allocation is an objective-allocation problem, not only an ACOS ranking problem.

• Do not drain Discovery merely because current ACOS is worse than a mature Profit campaign without checking the learning mandate and guardrails.
• Do not fund Growth merely because budget exhausts; require marginal headroom and explicit growth economics.
• Do not cut Defense solely from one efficiency ratio without strategic-coverage and cannibalization context.
• Do not starve an Experiment in a way that breaks the declared design.
• If campaign objective is Unknown, avoid reallocation that assumes campaigns have interchangeable jobs.

No universal campaign-type budget percentage is defined by this repository.

Budget control-state causal gate

When the question is whether a Sponsored Products budget change worked, whether a budget increase caused a result, or whether a surprising spend change came from a manual budget edit, load ../../references/control-state-comparability.md and use decision_surface=sponsored_products_budget_change when the required evidence is available.

For this causal surface, budget_or_pacing must distinguish the source-supported base_average_daily_budget, the effective_daily_budget that was actually in force for the comparison window, and active_budget_rules. If the connector cannot observe budget-rule state, keep material-control coverage Partial/Unknown; do not encode missing rule evidence as an empty list or unchanged state.

For this registered surface, overlapping bid controls include placement, audience and Sponsored Products video bid adjustments when applicable; missing video-control evidence is not a zero boost or proof of ineligibility.

Use ../../scripts/compare_control_state.py only as a read-only evidence classifier. Treatment Isolated means the evidenced registered controls isolate the intended budget treatment; it does not prove incrementality or authorize another budget mutation.

核心原则

“预算花完”不是自动加预算信号；“预算没花完”也不等于 campaign 有问题。预算动作必须结合边际效率、业务目标和预算池约束。Campaign 预算只是约束层级之一；账户级、portfolio 或外部预算 cap 可能才是真正瓶颈。

Progressive loading

存在多个 Campaign 竞争同一个业务预算、portfolio cap、Sponsored Products account-level daily budget cap、外部 pacing pool，或其他上游 spend constraint 时，再加载：

• references/portfolio-budget-conflicts.md

只有方向性预算结论需要转成具体金额时，再加载：

• ../../references/action-sizing.md

普通单 Campaign 预算诊断若不需要具体动作幅度，不加载这些 reference。

输入

需要：

• daily/lifetime budget；
• spend；
• budget status / budget utilization（如有）；
• campaign performance；
• 目标 ACOS/ROAS/利润；
• 时间窗口。

推荐：

• average time in budget / out-of-budget coverage；
• estimated missed impressions/clicks/sales（如平台提供，按模型估计处理）；
• platform recommended budget（如有，作为参考而非默认动作）；
• Sponsored Products account-level daily budget cap / portfolio cap / business budget；
• 小时级 pacing；
• placement；
• inventory；
• promotion plan；
• campaign priority/role；
• 同账户可调配预算池、protected floor 或外部 pacing 约束。

Configured budget vs effective budget

Keep these observations separate before diagnosis:

• base average daily budget — the configured campaign budget from a trusted source;
• effective daily budget — the budget in force after source-supported automated budget rules for the relevant date/time;
• observed spend / serving coverage — realized delivery evidence such as spend, budget status and average time in budget;
• budget-rule state — schedule/performance rule status and effective interval when the active source supports it;
• upstream constraint state — portfolio/account/business/external caps and pacing when decision-relevant.

Amazon currently documents Sponsored Ads budget rules that can automatically increase daily budget based on schedules/events or performance conditions, and multiple applicable rules can combine. A performance-triggered rule is an adaptive control: its activation may be related to the same performance signal being evaluated, so it is a potential control/confounding change rather than proof that a manual action caused the outcome.

Amazon also documents Sponsored Ads daily budgets as average daily budgets whose realized spend can vary by day under the applicable account setting/policy. Therefore:

spend on one day > base average daily budget does not by itself prove a manual budget increase, executor drift, overspend, or policy violation.

If effective budget or rule state is unavailable, say what is unknown and keep the conclusion Directional / Hold / Manual Review as appropriate. Missing rule evidence is never zero rules.

诊断

Efficient and budget constrained

高效且频繁因预算受限停止投放，是优先增预算候选。但还需检查：

• 库存；
• 是否已有足够规模；
• 预算增加后边际流量是否可能明显变差；
• 大促是否即将结束；
• 是否存在更上游的 account/portfolio/business cap；
• 如果总预算固定，新增预算从哪里来以及来源 Campaign 的机会成本。

average time in budget < 100%、out-of-budget status 或稳定日内提前耗尽可支持“delivery constrained”的判断，但不自动证明“profitable headroom”。

Inefficient and budget constrained

不能因为花完预算就增加。先解决 keyword/search term/bid/placement/retail efficiency。

Efficient and not spending

预算不是主要瓶颈。检查 bid、流量规模、相关性、排名、广告状态和 targeting coverage。

Inefficient and not spending

低优先级；可能需要结构整改而非预算动作。

Modeled opportunity evidence

平台提供的 estimated missed impressions/clicks/sales、recommended budget 等可作为有用的机会信号，但必须与观测数据分开：

• spend / budget / average time in budget 属于 serving/current-state 证据；
• missed opportunity / recommended budget 属于模型估计或推荐；
• 模型估计不等于保证增量，不替代利润、库存、重叠流量和上游 cap 检查；
• 无法解释估计方法、范围或适用窗口时，降低 confidence，而不是把预测值当成事实。

Pacing

如有小时数据，检查：

• 是否上午过早耗尽；
• 高转化时段是否没有预算；
• 是否存在大促导致的临时加速；
• 日内波动是否只是单天随机性；
• account/portfolio cap 是否在相同时间窗口先于 campaign budget 成为瓶颈。

没有小时数据时，不声称可以精确做 dayparting。

上游预算约束

当存在 account-level / portfolio / business / external cap 时：

1. 先识别实际 constraint hierarchy；
2. 记录每层 cap、当前 spend 与可信 remaining headroom；
3. 若上游 cap 已绑定，不能把单 Campaign 提额写成“可独立获得更多 delivery”；
4. 要么给出平衡的 source → destination 重分配，要么把上游 cap 变更作为单独的业务预算决策；
5. 若 headroom/source budget 不清楚，输出 Directional / Hold / Manual Review。

预算重分配

当总预算固定时：

1. 明确总预算池、业务目标和不可随意移动的 protected spend；
2. 识别低边际回报预算；
3. 识别高边际回报且存在真实 headroom 的 Campaign；
4. 把建议写成来源 Campaign → 目标 Campaign 的平衡转移，而不是彼此冲突的独立增预算建议；
5. 同时估计目标收益与来源机会成本；
6. 采用与证据质量、风险和可逆性匹配的变更幅度；
7. 设验证窗口观察 pool-level 结果与边际效率。

变更幅度 Guardrail

不使用通用固定百分比作为默认预算调整幅度。需要给出具体预算金额时，使用 ../../references/action-sizing.md，并让建议幅度受以下因素共同约束：

• 可用的 marginal headroom 证据；
• 数据量和归因成熟度；
• 库存与促销窗口；
• 账户/portfolio/business cap 的真实可用 headroom；
• 预算池可移动空间；
• Campaign 的业务角色和 protected floor；
• 最近是否存在未完成验证的 Bid / Placement / Budget / Structure 变更；
• 动作可逆性与潜在损失上限；
• caller/account 显式提供的 max change、spend-at-risk 或其他政策限制。

证据弱时优先 Directional、Probe/Shadow、Experiment 或 Hold，而不是套用固定比例。无法证明具体 proposed_value 时保留方向和缺失约束，不制造精确金额。

输出

Budget map

每个 campaign 标记：

• constrained/not constrained；
• constraint level：campaign / portfolio / account / business / external / unknown；
• efficient/inefficient；
• average time in budget / serving coverage（如有）；
• observed vs modeled headroom evidence；
• role/priority；
• marginal headroom；
• pool constraint / protected floor（如适用）；
• recommended direction。

Reallocation plan

列出来源 campaign → 目标 campaign，以及：

• 预算前后（仅当具体幅度可被证据支持）；
• active upstream caps / remaining headroom；
• action class / sizing basis；
• pool total reconciliation；
• 观测证据与模型估计分栏；
• 来源机会成本；
• 预计目的；
• confidence；
• guardrail；
• validation window。

禁止

• 不把预算增长当作销售增长保证；
• 不把 estimated missed sales/clicks/impressions 当成保证增量；
• 不把 platform recommended budget 当成无需校验即可执行的默认值；
• 不在上游 account/portfolio cap 已绑定时，把单 Campaign 提额描述为独立可实现的 delivery 增长；
• 不把平均 ROAS/ACOS 当作下一单位预算的边际回报；
• 不在固定总预算池下对所有受限 Campaign 同时给出无法调和的增预算建议；
• 不从高效品牌防御或其他 protected Campaign 抽走关键预算而不提示风险；
• 不在库存紧张时建议大幅扩量；
• 不同时大幅改 budget、bid、placement 而没有分阶段验证计划；
• 不把第三方/平台示例百分比直接当作本账户默认预算动作幅度；
• 不引入未经授权的真实账户写入。

Active budget-rule member evidence

A complete list of active budget rules is not proof that its rows are usable. For a registered Sponsored Products budget-change causal review, require each repository-normalized rule member to have a unique nonblank rule_id and finite nonnegative observed increase_percent; also check schedule/event bid-rule members separately. Unknown/Unsupported/Partial, duplicate/missing IDs, boolean, negative or non-finite modifiers yield Unknown, not an assumed zero or stable budget. An observed zero and a source-supported empty list remain distinct from missing evidence. This comparator is read-only and does not prove rule activation or effective-time coverage; see ../../references/control-state-comparability.md.

Optional budget-rule activation evidence

When a normalized active budget-rule member explicitly contains activation_status, it must be active, inactive, or scheduled; unknown, unsupported, malformed or one-sided status evidence on a shared rule identity is Unknown in a causal comparison. Both-absent legacy status is not proof that a rule did not execute during the window. See ../../references/control-state-comparability.md. Do not infer rule execution from configured modifiers, observed spend or snapshot recency.

For explicitly supplied normalized budget-rule activation_effective_at, reject malformed, future, status-less or one-sided transition timestamps; do not infer missing from zero or assume unchanged execution. Verified timestamp changes are competing controls. This read-side evidence check is not a guarantee of schedule/event window coverage; see ../../references/control-state-comparability.md.