Decision engine screens

What it is#

The screens belong to the Marketing Systems OS package, which a platform administrator enables per organization; it is off by default and needs the social intelligence package. With it off, every screen below disappears from the menu and a typed address is refused. By default the screens are in the owner and admin menus; an admin can grant them to other roles. The mechanism behind them is on The decision engine.

Two ways it learns, and one that is not learning

The third row is not learning. It is a fixed rule with one multiplier that can only go down, to a floor of a quarter; it has no probability, no interval and no trained model (Signals and the morning brief). Beta: dismissals arrive only through the platform console or the API today (status).

Only confirmed outcomes are learned

The engine's own verdict on a decision (hit, partial, miss) never enters the track record. It counts only once an owner or org admin ratifies or overrides it. The background worker may write proposals and measurements; it never writes an approval, an activation, a rejection or a confirmation. Each of those is a person's action, behind a permission check, with an audit-log entry.

Learning per business domain

A decision is recorded against a lever, and levers come from the domains the organization has activated, not from a fixed commerce list. A lever no domain declares gets no default priority. Two domains may use the same lever name without colliding. The general playbook path records the chosen playbook as its lever, so an organization that does not use the commerce lever plane still builds a decision history. Off-policy evaluation still uses the commerce lever list by default.

Why it does not claim a causal effect

Every decision records the candidate set, the choice, its true probability, the logging policy, and a measurement window declared before any result existed; a build gate replays recorded decisions and passes only with zero divergences. Selection, however, is deterministic, so every recorded probability is certainty. Off-policy evaluation exists and refuses on such logs rather than printing a number, and no learned policy is promoted without one. Outcome readouts are before-and-after comparisons on first-party data, labelled as confounded: directional evidence, not a measured causal effect. See Measurement and honest numbers.

It does not act on its own

The engine explains a receipt and drafts a typed proposal. A person approves it, and an approved action reaches a person as manual steps; the platform does not change a marketplace listing, an ad account or a budget. Automatic execution is closed by policy.

When it declines to answer

Below its evidence floor the engine returns "not enough data" or "no reliable signal" instead of a number. A generated sentence whose number does not trace to a supplied fact is dropped. A signal with too few periods is recorded as not evaluable, not as "nothing to report".

How it works#

Menu group Marketing Systems OS ("Hệ vận hành Marketing") holds six screens; the Action Queue and the weekly memo sit under AI Assistant.

Analysis & Suggestion

Marketing Systems OS > Analysis & Suggestion ("Phân tích & đề xuất"), /marketing-systems. It opens on Loop state: four cells for read (signals computed, latest day), fire (what fired and what was withheld, with the reason), decide (weekly runs: proposed and abstained, with the reason) and answer (proposals waiting for a person, and the oldest one's age), plus recent weekly runs. Below it, an organization with marketplace orders sees the diagnostic chain: detected signals, decomposition, diagnostic scan, channel contribution, lever scan, demand shift, combo analysis, benchmark triangulation and social report. Each culprit the scan names links to the one proposal waiting for it in the Action Queue. An organization measured on leads or deals sees a chain built on its own conversion instead of the order axis.

Social conversation

Marketing Systems OS > Social conversation ("Thảo luận mạng xã hội"), /marketing-systems/signals. Weekly conversation volume per lever from the organization's social listening feed: the input the off-take loop reads, not its result. Above the raw series, the lever the current policy suggests trying next period.

Decisions & Lessons

Marketing Systems OS > Decisions & Lessons ("Quyết định & Bài học"), /marketing-systems/decisions. The decision inbox (waiting, running, history) and the decision memory timeline: signal, proposal, decision, outcome, lesson, next time. A calibration card compares the engine's stated confidence with the realised hit rate. Each card answers "what happened last time this fired" from the frozen receipt of the earlier decision. Owners and org admins approve, activate, reject and confirm; other roles read, and the server enforces the same rule as the hidden buttons.

Outcome measurement

Marketing Systems OS > Outcome measurement ("Đo lường kết quả"), /marketing-systems/measure. Every activated proposal against first-party results: expected, measured, the difference and the verdict. A system-health block shows drift, the champion and challenger comparison, and ledger completeness, and says whether the loop is drifting or short of data. The champion and challenger cell compares the current policy with a computed variant of itself; it is not a second policy serving decisions.

Industry playbook library

Marketing Systems OS > Industry playbook library ("Thư viện playbook ngành"), /marketing-systems/playbooks. The playbooks the loop draws from, read-only here. Each card links to its machine rule (name, version, status) and shows the pooled track record once enough confirmed outcomes exist; below that it shows counts, not a rate. The library is filtered to the organization's activated domains, and the number of hidden playbooks is stated on screen. The catalogue itself is on Marketing playbooks.

Operating settings

Marketing Systems OS > Operating settings ("Cấu hình vận hành"), /marketing-systems/settings. The loop's thresholds: verdict tolerance, diagnostic scan severity, the off-take cold-start point, and the conversion the organization measures decisions on. It also lists recent weekly runs with how many proposals each made or skipped, and why each skip happened. Members with the organization-settings capability (owner and admin by default) edit; others view.

Action Queue

AI Assistant > Action Queue ("Trợ lý AI > Hàng đợi hành động"), /action-queue. The actions an approved decision produces, waiting for approval, with running experiments and the monthly approval rate. The catalogue of available actions follows the organization's planes: an organization without commerce data never sees commerce suggestions such as restocking a best-selling SKU. Opening it needs both the AI management and the action management capabilities.

Weekly action memo

AI Assistant > Weekly action memo ("Trợ lý AI > Ghi chú hành động tuần"), /decisions/weekly-memo. A weekly digest of findings across brand, channel, creator, price, inventory and service, graded info, watch or critical, and deduplicated before display. Anyone with dashboard access reads it; publishing or regenerating needs the decision management capability, and without it the buttons are not shown.

What it reads and what it never reads#

• Reads: the organization's own first-party planes, its declared conversion, its social listening feed as an input signal, and its members' confirmations.
• Never reads as an outcome: third-party market estimates or social signals; they can prompt a hypothesis, not score a decision.
• Never writes outside the platform.

Limits#

• The weekly proposal scan runs only for organizations that enabled the package and declared a conversion.
• Proposals in the learning state stay there until the conversion feed returns data for their window.
• The rule editor is not on these screens; see Marketing playbooks.

Related#

• The decision engine
• Measurement and honest numbers
• Signals and the morning brief
• Marketing playbooks
• Read and act on the morning brief
• Project status