Methodology
How the intelligence is produced.
Predictively uses a constrained, deterministic approach to intelligence generation. Every design decision exists to make the output more reliable, more auditable, and less susceptible to the hallucination and vagueness patterns that make unconstrained AI research difficult to use professionally.
Generation process
Each brief follows a documented production sequence.
Source data ingestion
Verified operator data from the Ontario Gaming Commission is ingested through a deterministic pipeline. Period keys, metric labels, and fiscal calendar handling are standardised at ingestion. No manual editorial decisions are made at this layer.
Source pack construction
A structured source pack is assembled from the ingested data. This includes monthly timelines, quarter-to-date year-over-year comparables, trend signals derived from the data series, decision signals, and evidence bullets. The source pack is the only permitted evidence base for the generation step.
Constrained generation
The generation engine operates against a fixed JSON schema and a system prompt that prohibits invented figures, unsupported causal claims, and generic language. Every string field must be commercially specific and traceable to the source pack. The engine cannot output unknown schema keys or null values unless the schema explicitly permits them.
Quality scoring
The generated brief is evaluated against a scoring schema. Each of five dimensions receives a score from 0 to 10: specificity, commercial utility, credibility, actionability, and non-generic language. The overall average must meet a configured threshold. Briefs below the threshold are not released.
Targeted regeneration
If a brief fails the quality threshold, the weak sections identified by the scoring pass are regenerated while the stronger sections are preserved. Up to two regeneration attempts are made before a full regeneration is triggered.
Structured rendering
The validated JSON payload is rendered into the fixed four-module briefing format. The same source data powers both the intelligence brief and the data visualisation layer, maintaining source lineage across both surfaces.
Operating principles
Six principles that govern the production methodology.
Evidence first
Every claim in the brief must be anchored to a specific figure or pattern in the source pack. Generalised market commentary is treated as a quality failure.
No hallucination surface
The generation step cannot invent regulators, market events, figures, or causal claims not present in the source pack. The schema constraint makes hallucination detectable.
Comparable period discipline
Quarter-to-date comparisons use the same number of months from the prior-year equivalent quarter. Partial periods are never compared against complete ones.
Fixed output format
The briefing format does not change between cycles. Users learn the structure once, then read it faster on every subsequent brief.
Brevity constraint
Every string field in the brief is constrained to one or two sentences. The value is in precision and density, not length.
Auditable lineage
The source pack, generation schema, scoring payload, and rendered brief are maintained in the same production run. Any output can be traced back to the specific source data that produced it.
Data handling