derivatives: semantic-layer

Drafts only; nothing auto-posts. Numbers trace to filled-intake-A.md; illustrative figures are labelled where used. Run checks.py on this file with the same record.

linkedin

draft 1: closer

Your CEO waits days for a number a system could compute in seconds.

Not because the data is missing. Because the question has to travel: CEO to COO to rollout lead to centre manager. Every hop costs a day and a little truth, and the answer comes back pre-negotiated.

We keep hearing the same request across industries right now: GoStudent's 100-plus-location tutoring network, a vehicle manufacturer's dealer organisation, an e-bike retailer with 100 plus points of sale. Different sectors, identical problem. Leadership steers with numbers that arrive late and flattered.

The fix is not another dashboard. It is one canonical definition per metric, computed from data nobody thinks to fake, with every source's disagreement shown instead of averaged away.

We wrote up the pattern, including the part where it gets political.

{{POST_URL}}

draft 2: architect

A dashboard cannot fix a metric nobody defined.

The failure chain is always the same: definitions live in people's heads, targets bend them quietly, reports inherit the bending. So we build the boring, strict version instead:

One canonical definition per metric, versioned and owned. "No-show = booked meeting, no check-in event within 30 minutes of start. v1.2. Owner: COO office."

Computed from machine exhaust: calendar events, chat timestamps, CRM state changes, payment records. Byproduct data, which nobody falsifies.

Reconciliation as the product. Illustrative: bookings say 100 meetings, chat analysis finds 82, CRM shows 15 no-shows, revenue implies 78 sessions. The divergence IS the finding.

And one door in: agents query the semantic layer, never raw tables. A model on raw tables improvises a fresh definition per question.

Full pattern, including what we deliberately refuse to build:

{{POST_URL}}

draft 3: operator

The centre manager is not lying.

His target is 200 booked meetings. A no-show, reclassified as a reschedule, protects that target. Nobody ever wrote down what a no-show is, so the definition quietly became his call.

That is not a character flaw. It is what any undefined metric does to any team with targets: it invites everyone to hit the number by moving what the number measures.

The uncomfortable part of fixing it: a shared data layer deletes the information asymmetry parts of middle management run on. Expect resistance. Design the ingestion for it, and treat human input as the least trusted source.

How we approach it, and the first step you can do without us this week:

{{POST_URL}}

newsletter

Subject: Your metric has no definition. Your team knows.

Body: A CEO asks how many no-shows a two-month-old branch had. The question relays through three people and comes back days later, softened at every hop, because nobody ever wrote down what a no-show is. We keep meeting this pattern across tutoring, automotive and retail: dashboards answer "what" while definitions still live in people. The fix is one canonical, versioned definition per metric, computed from data nobody thinks to fake. Here is the full pattern, including the political part: {{POST_URL}}

sales snippet

If your leadership waits days for front-line numbers that arrive pre-negotiated, the problem is rarely the data. It is that no metric has a written, owned definition, so every report gets quietly reinterpreted on its way up. We build semantic layers: canonical definitions, computed from system data, disagreements shown rather than averaged, queryable by AI in seconds.

metric: evaluation gate before widening: 30–50 leadership questions verified against trusted reports (pilot design)

x thread

1. A CEO wants one number: no-shows at the branch that opened two months ago. The question travels CEO → COO → rollout lead → centre manager. Days pass. What comes back is a negotiation.
2. The manager isn't lying. His target is 200 booked meetings, and nobody wrote down what a no-show is. Undefined metrics invite everyone to hit targets by moving what the target measures.
3. We keep hearing this across industries: GoStudent (100+ tutoring locations), a manufacturer's dealer org, a 100-plus-store e-bike retailer. All have dashboards. Dashboards answer "what". Definitions still live in people.
4. What works is boring and strict: one canonical definition per metric. "No-show = booked meeting, no check-in within 30 min. v1.2. Owner: COO office." Answers computed, never collected.
5. Prefer machine exhaust over human input: calendar events, chat timestamps, CRM state changes, payments. Byproduct data. Nobody falsifies a byproduct.
6. Show the disagreement (illustrative): bookings 100, chat analysis 82, CRM 15 no-shows, revenue ~78 sessions. Don't blend it away. The divergence points at the numbers being managed.
7. First step needs no vendor: pick your three most-leaned-on metrics, write one-sentence definitions with owner and version, count the systems that could compute each independently. The gap is the honest size of your problem. {{POST_URL}}