Decisions and Limitations

This file records maintained behavior that is easy to misinterpret. The scripts and tests are authoritative when prose and implementation disagree.

Decisions

1. Compute and commit are separate. The scripts create local proposals and reports only. Live mutation remains an integration action after approval.
2. Account context is part of the proposal. Advertiser, profile, marketplace, currency, and timezone are required inputs and are rechecked before commit.
3. Targets are exact. A campaign must match Ad Type × Strategy × Product Line. The allocator never borrows another product line's target.
4. ROAS excludes immature days. roas_data_through must be at least two days before as_of; spend may be newer but must be labeled preliminary when it is.
5. Amazon budget behavior is explicit. The nominal pool reserves the current overdelivery setting. Active Amazon budget rules block the run until modeled.
6. Ceilings are hard constraints. Floors are reserved first. An impossible floor/ceiling combination is an error; overlapping caps may leave part of the pool unallocated.
7. Change caps cannot override ceilings. If a lower change bound would violate a ceiling, the run errors rather than publishing an internally inconsistent proposal.
8. Exhaustion affects the score before allocation. Chronic early exhaustion multiplies the efficiency score by the configured pacing-health boost.
9. Pausing is a state action. Reaching the monthly hard ceiling emits pause_campaign, never a zero-dollar budget update.
10. Reports distinguish targets from proposals. Today's forward value is the constrained proposal; later values are planning targets recalculated daily.

Known limitations

• The allocator supports campaign-level daily monetary budgets for SP, SB/SBV, and SD. It does not model DSP, lifetime, portfolio, or shared budgets.
• Amazon budget rules are detected but not modeled; the safe behavior is to stop.
• The overdelivery reserve is conservative. It protects the daily target but does not predict auction delivery or guarantee exact monthly spend.
• Product-line and strategy ceilings are daily shares of the nominal pool. A month-to-date share constraint would require bucket-level historical spend.
• Saturating overlapping ceilings can conservatively leave money unallocated even when another mathematically feasible mix might spend more.
• Future report values assume today's seasonality and event inputs and are not a commitment or delivery forecast.
• Live tool names and payloads are deployment-specific and must be discovered from the connected integration.

Future extensions

• Model active Amazon budget rules instead of blocking them.
• Add month-to-date bucket spend for monthly-share ceilings.
• Add an optimizer that maximizes score under intersecting caps while preserving the current deterministic, stdlib-only contract.
• Define an unattended-mutation policy only after explicit operational review.