Skip to content
MSO CloudDocumentation

From signal to decision

On this page

In short#

A measured number becomes a ranked morning-brief line through a fixed chain: collect, evaluate, suppress, rank, record. Every evaluation is written to an append-only ledger, including silent ones, and human feedback can only lower a rank.

1 · CollectBind pack signals,activated domains2 · EvaluateFired, insufficient,quiet or unreadable3 · SuppressDismissed, cooldown,inhibited4 · RankImpact × weight ×severity × feedback5 · Morning briefRanked lines foractivated domainsSignal ledger · append-onlyEvery evaluation, fired or withheld, with its reason from a closed set; only four feedback fields may changeevery evaluationdismiss, actedweight ×¼ to ×1Feedback can only demote. Only the background worker inserts instances.

The word "signal" names two objects, and keeping them apart is what makes replay possible. A decision episode pins both versions, because pinning one cannot reproduce the other. The decision half of the loop is on The decision engine.

Object One row is Used for
Feature store The scalar a rule predicate reads at a period Deciding whether a rule matches
Signal ledger One evaluated binding at one as-of day: fired or withheld, and why Answering "why was I not told"

Collect#

The collector reads signal declarations from the organization's installed dataset pack (newest version wins) and binds each at that version. Only activated domains are evaluated.

  • Window: the dataset-wide total over the last twelve periods, with a cell cap per read.
  • Sensitive fields: a compensation-sensitive field is removed by the read-scope resolver before any cell is fetched, so a salary column cannot reach a brief.
  • One address per line: a brief line declares a dataset plus measure key, or a catalog measure reference; both or neither is rejected by schema. Line kinds are append-only.
  • Shared rules: brief facts and signal series share one implementation of scope and cell picking.

Evaluate, then suppress#

Evaluation fires or returns insufficient or quiet; the collector adds unreadable. Suppression then reads the ledger's own history in a fixed order, so the recorded reason names the person before it names the clock.

Evaluatethe type's comparisonover 12 periodsFiredthreshold crossedInsufficienttoo few periodsQuietnothing crossedUnreadableno series or bad params1 · Dismissed?a person, within 30 days2 · Cooldown?same key, declared periods3 · Inhibited?a higher key, same passyesnonoDismissednames the personyesCooldownnames the clockyesInhibitednames the covering keyyesRankedinto the morning briefnoSignal ledger, append-only: all seven outcomes, each with its reason from a closed set; the worker inserts, the web runtime may only set four feedback fields.Suppression order is fixed so the recorded reason is the same on every run: a person's dismissal before the machine's cooldown, cooldown before inhibition.A withheld instance suppresses nothing itself; only a fired one starts a cooldown.
Seven outcomes, one ledger. Only the right-most survivor reaches the brief; the other six are written with their reason, never dropped.

The thirty-day dismissal window is one constant; the database predicate derives from it. A database check holds the reason set, and a withheld instance is written with its reason, never dropped.

The ledger#

The ledger is append-only at the database: a trigger rejects any update outside four feedback fields (who dismissed, when, the reason, when marked acted). A web request therefore cannot manufacture a signal instance. Rows carry a brand, so the ledger is isolated per brand as well as per organization; its idempotency key treats an empty brand as a value, so a re-run cannot duplicate an organization-level row.

Writer May
Background worker Insert instances
Web runtime Read, and update the four feedback fields only

Rank, and the weight that corrects it#

The score is a product of four integers. The feedback weight is a rebuildable cache over the ledger, never a source of truth; rebuilds delete stale rows one key at a time, so one brand's rebuild cannot remove another brand's row. It is a hand-written rule with no probability or interval, not a learned policy.

score =Impactbasis pointsfrom the measure×Weightdeclared per signal1 to 1000, default 100×Severityinfo 1.0warning 1.5 · critical 2.0×Feedbackper key, last 90 daysclamped 0.25 to 1Feedback multiplier0.25 floor (database check)0.51.0 ceiling: feedback never promotesAt least one 'marked acted' in 90 days× 1No 'acted', two or more dismissals in 90 days× 0.5Anything else, including no feedback× 1One dismissal is not evidence; 'acted' wins over 'dismissed'.The multiplier is a rebuildable cache over the ledger, never a source.Raising a signal means changing its declared weight, not clicking.
The ranking number. Integers in, integer out; an organization with no feedback ranks exactly as it did before the feedback loop existed.

Where a brief alert comes from#

Where a brief alert comes fromDataset rulesEvaluated every 15 minutesOrg on/off switch in AlertsAlert workflowsGraph: trigger, detect, condition,alert or silence; runs before the briefalert_eventsRows of enabled, org-visiblealert_rules only; one ledgerdeduplicated per rule and periodMorning briefPer person and cadence:daily 07:00, Mon 07:00, 1st 07:00mute and snooze per readerIn-app notificationThe floor: one brief.readyper person, cadence, dayOptional channelsEmail, Telegram, Zalo, pushneeds platform setupPlatform-audience rules show to super admins only. Unconfigured org rules are switched off at deploy.
The brief shows only alert_events of enabled rules. Nothing outside this path becomes a brief line.

An alert in the brief is an open or acknowledged incident (alert_events) of an enabled, org-visible rule, filtered by the audience fence, the reader's brand grant and the reader's own mute and snooze. A detection that is not such a row never becomes a bullet. One incident stands per rule, brand and entity: a still-true run updates it, a run back inside the clear band resolves it, and only an opened incident starts the playbooks subscribed to the rule. The brief reads the last closed ICT day; a GMV, return-rate or contribution-margin alert links to that day's commerce report. Dataset rules are evaluated by the 15-minute job; a rule owned by an alert workflow is evaluated only by the workflow runtime, in the daily brief job before the brief is written. Admins switch each org-visible rule on or off, audited; platform rules are refused. Briefs are written per person: daily at 07:00, weekly on Mondays at 07:00, monthly on the 1st at 07:00, Vietnam time.

Two cycles, gated separately#

Cycle Does Runs when
Record Collect, write the ledger, rebuild the feedback weight The signals layer is on and at least one domain is activated
Decision Fit, review and validate a policy; propose and measure decisions The Marketing Systems OS package is enabled and a conversion is declared

The record cycle runs inside the per-organization 07:00 brief job; a declining gate returns "skipped", not an error. Only that job writes the ledger; the in-product refresh re-ranks and writes nothing. One module wires the collector to the brief generator; the job and the refresh both build from it inside one organization-scoped isolation context. A failing read degrades to pack declaration order; a failing write is reported and never costs the brief.

From a signal to a decision#

A decision is recorded against a lever, validated against the levers the organization's activated domains declare. An undeclared lever gets no prior; policy models carry their domain in their key; the general playbook path records playbook:<rule id>. Off-policy evaluation still defaults to the commerce lever list.

When the loop writes an episode it links the fired instance that stood for the episode's domain at or before the decision's period: never a suppressed instance, never a later one, never an invented binding version. Without feedback the weight stays at one; an owner or org admin marks a brief line done on the Morning brief page, and dismissal is in the signal ledger of the platform console.

What it reads and what it never reads#

Direction Scope
Reads Installed pack declarations, dataset rollups for activated domains, the ledger's history, human feedback
Never reads Compensation-sensitive fields, unactivated domains, platform-infrastructure signals (super admins only)

Where it stops#

The chain recommends and a person approves. Human feedback can only lower a rank, and signals outside activated domains are never read. Dated status per capability: Project status.

Decisions#

Decision Why Enforced by
The value a rule reads and the record of an evaluation are two stores Merged, a replay pins one version and silently reads another The episode schema carries both pins; replay refuses an episode without a rule version
Every evaluation is written, quiet and withheld ones included "Why did the brief not tell me" must have an answer after the fact A database check on the reason set and an append-only trigger
Feedback multiplies a rank by at most one, floor one quarter A weight that could promote turns a few clicks into a fabricated priority A clamp in code and a database range check
Only the worker identity inserts signal instances A web request that could insert instances could manufacture evidence Runtime grants in the privilege manifest, checked by the runtime-privilege audits