Skip to main content
A consequence link is the edge between a decision and the financial event that decision caused. Links carry four fields that make the attribution actionable: which pair they connect, how strong the attribution is, how the attribution was established, and what kind of financial impact it represents. Without links, a revenue event is just revenue. With a link pointing at a decision, that revenue is attributable to an agent with a provenance trail you can show to a skeptic — or an auditor. The ledger records both automatic links created by the attribution engine and manual links created by a human confirming a suggestion on the dashboard. The attribution_method field records which — so you always know whether a number came from a machine score or a human judgment.

​
Shape

{
  "id": "lnk_...",
  "org_id": "o_...",
  "decision_id": "d5e7a3b0-...",
  "financial_event_id": "fev_abc",
  "correlation_strength": 0.91,
  "attribution_method": "ml_inferred",
  "link_type": "revenue",
  "notes": "attribution engine: score=0.91",
  "created_at": "2026-04-16T12:40:00.000000Z",
  "deleted_at": null
}

# ``decision_id`` is a UUID string — use whichever variable your
# caller exposes (``receipt["decision_id"]`` from ``log_decision``
# in non-batched mode, or a stored id from a prior write).
pq.link_outcome(
    "d5e7a3b0-0000-0000-0000-000000000001",
    "fev_abc",
    correlation_strength=1.0,
    link_type="revenue",
    notes="direct cause — pricing API response produced this charge",
)
When you create a link via the SDK with a known decision_id and financial_event_id, set correlation_strength: 1.0 and let the attribution_method default to direct. Reserve lower correlation values for cases where you know the relationship is partial.

​
Attribution methods

MethodWho creates itHow the engine decides
manualHuman via the dashboardReviewed and confirmed by a person
directSDK, passing a known idExplicit caller-supplied link
temporalCorrelation workerTime proximity and metadata overlap
ml_inferredAttribution engine (auto)Composite score at or above auto-link threshold
ml_suggestedAttribution engine (review)Score in the suggested band — awaits confirmation
The engine’s scoring combines temporal decay, Jaccard metadata overlap, a Bayesian agent-pattern prior, and amount plausibility. Links with a score in the ml_suggested band appear in the dashboard review queue rather than being applied automatically. When the attribution engine surfaces a suggested link, you confirm or reject it from the Links tab in the dashboard. Confirming promotes the attribution_method to manual and makes the link active. Rejecting soft-deletes the link — it is excluded from P&L queries but remains in forensic exports so an external reviewer can see the prior suggestion was retracted.
Soft-deleted links are never hard-deleted. The audit trail is always intact, even after a rejection.

revenue

The decision drove positive revenue — a successful upsell, a completed checkout.

cost

The decision drove a cost you would not otherwise have incurred — extra API calls, ops labour.

liability

The decision created legal or compliance exposure — an erroneous claim denial, a regulated miscommunication.

neutral

Attribution matters for the audit trail but the dollar sign is zero — for example, a canceled transaction.
The sign of a link’s contribution to the P&L comes from the linked event’s amount field, not from the link_type. The type is metadata for filtering and reporting, not an arithmetic override. A revenue link to a negative-amount event will subtract from the revenue total.
LIABILITY and NEUTRAL link types are excluded from the revenue and cost sums in the AI P&L — liability flows into the exposure column only, and neutral links are audit metadata with no P&L effect.