From conversation to typed fact, on a weekly clock.

Stage 02 is how a raw transcript becomes a row in your fact corpus. The enrichment pipeline runs every accepted conversation through a typed schema: ten parallel dimensions, each producing rows with a bi-temporal window, a segment vector, a salience count, a confidence score, and a citation back to the chunk that produced it. The corpus refreshes on a weekly clock by default. You can tune the cadence per dimension if your vertical asks for it. Comparability is the product.

This page is the product view of that stage: what the schema looks like, what a refresh does, and what you control from the console. The pipeline mechanics (the hallucination filter, the cross-contamination check, the deterministic scoring, the prompt integrity hash) are documented under Methodology and under Trust & Security.

Each conversation produces rows across ten parallel dimensions. For retail the operative set is: pain points, opportunity scoring, brand-perception map, AI-readiness scoring, retention drivers, market signal, regulatory and compliance risk, training gap, customer-segment behavior, and competitor mention map. The schema is vertical-specific in surface form. Two or three dimensions swap out per vertical (for consulting, customer-segment behavior is replaced with engagement-economics tells; for restaurants, brand-perception map is replaced with shift-level operational variance). The shape stays the same: ten typed slots.

Stability is the whole point. A pain point recorded in one market in May reads the same way as a pain point recorded in another in November, because both are rows in the same schema produced by the same protocol. That property is what makes cross-market benchmarking possible. Surveys cannot do this, because the instrument changes with every translation. Focus groups cannot do this, because group dynamics warp comparability. The interview agent holds the protocol constant across every conversation, which lets the schema stay constant on top of it.

A refresh cycle is a scheduled batch of new conversations that the platform recruits, runs, and enriches. The default cadence is weekly per active cohort. When the cycle completes, every new fact lands in the corpus with its own bi-temporal window. Facts whose validity has expired since the last cycle move into the historical record. The dashboards re-compute against the refreshed corpus.

The refresh trigger is not a calendar reminder. The console surfaces a per-dimension staleness clock: when twenty percent of the brand-perception facts have aged out of validity in the last sixty days, the console reports that the dimension is due for refresh. This is the recurring-revenue spine of the platform: stale data is a product signal you can see in the dashboard, and it tells you exactly when to commission more capture.

Decay is calibrated per dimension and per vertical. Brand-perception facts decay in weeks. Market signal decays in weeks. Customer-segment behavior decays in weeks to months. Pain points and opportunity scoring decay in months. AI-readiness scoring decays in quarters. Retention drivers decay in quarters. Training gap decays only when the training catalog or product changes.

The platform ships with a baseline calibration that covers retail, consulting, restaurants, and pharmacy. From there, the Expert-in-the-Loop refines the half-lives for your vertical. When a domain expert on your team tells the platform that a particular fact class moves faster or slower than the baseline, the calibration updates and the corpus replays under the new decay rule. The prior calibration stays in the audit trail.

Every conversation lands in the corpus with a 0 to 100 composite quality score, computed by a deterministic formula. The formula combines richness (how many pain points, opportunities, surprises, contradictions, and verbatims the conversation produced), depth (turn count, longest-turn length, duration), and execution (how well the agent followed the six audited KPIs). A score of 71 means three pain points, two opportunities, one surprise, and full coverage on the brand-perception KPI. The breakdown is visible per conversation.

Conversations below a threshold are flagged as outliers and excluded from executive-facing aggregates. The intelligence record is preserved for audit; it does not disappear. The full description of the scoring formula, the weights, and the outlier thresholds is on the Methodology page.

When a client domain expert on your team tells the platform that an interpretation is wrong in their context, the system finds every conversation in the corpus that matches that context and re-pipelines them. It re-analyzes, re-tags, and re-weights every match. The expert’s correction propagates across the entire historical corpus. The audit trail keeps every prior interpretation alongside the new one.

There are two surfaces for making a correction. The first is a dashboard annotation: the expert sees the interpretation, the source verbatim, and the methodology description, and overrides where needed. The second is the interview agent directly, in voice. The expert talks to the agent as a coaching surface, and the refinement enters the same correction loop. The detail is on the Trust & Security page.

For every enriched conversation, the corpus gains roughly one hundred fact rows across the ten dimensions. Each row carries the dimension, the claim, the source chunk identifier, the bi-temporal window, the segment vector, the salience and confidence scores, and the embedding for vector retrieval. The same row reaches the dashboard, the brand-partner API, and your AI agents through the MCP surface. Citation is identical across the three.

At 1,000 conversations the corpus holds around 100,000 fact rows. At 50,000 conversations on weekly refresh, the corpus holds millions of rows with active staleness management. The architecture handles all three scales on the same store. The decision to design for the largest case at the smallest deployment is what keeps the corpus useful as you grow. Bolting capacity on later means re-ingesting the historical record, which is where most platforms collapse.