Metric Definition First, Number Second (2026)
By William Zhu & the InfiniSynapse Data Team · Published: 2026-09-02 · Last updated: 2026-09-02 · Last verified: 2026-09-02 · Next review: 2026-12-02 · Editorial standards · Corrections
Table of Contents
- TL;DR
- What a metric definition is before any digit
- A four-layer harness framework
- How this work differs from a semantic-layer purchase
- Tool landscape for outlines and bound sentences
- How to bind the sentence, then compute
- Desk sample: an illustrative account-segment pack
- Scorecard: ready to compute or still drafting
- Failure modes
- Frequently Asked Questions
- Conclusion
TL;DR
Direct answer: A metric definition is the bound sentence that names grain, filter, and window before any query runs. The model decides whether it can think like a person. The harness decides what kind of person it becomes. Sequence the work as direction, then steps, then the metric definition, then the number. A correct digit with no sentence cannot enter a decision.
What you'll learn:
- What a metric definition is, and why it is not a semantic-layer buying page
- A four-layer frame: direction, steps, sentence, number
- How a Data Agent path must stay readable by operations and compliance
- How to state source, table reason, filters, sort, and juxtaposition first
- An illustrative account-segment desk with 8 / 42 / 88 objects, not people
- When to refuse the number until the metric definition is inspectable
A Data Agent sits closer to operations and compliance than a Code Agent. The path has to be readable by business, not only by engineers who can parse a blob of SQL. If you skip that bound sentence, you skip the job.
What a metric definition is before any digit
Key Definition: A metric definition is the written sentence that binds grain (what one row is), filter (who is in), and window (which time or event range) before calculation. It is the third object in the harness sequence—after direction and steps, before the number—and it must be inspectable, editable, and reusable.
Independent published context (retrieved 2026-09-02; separate from this page’s desk composite): the NIST AI Risk Management Framework treats measurement and documentation as controls, not afterthoughts. UK NCSC guidelines for secure AI system development ask teams to keep human-readable intent around model output. Those pages did not run this desk and are not a product award.
The bound sentence is not a slide adjective. “Active accounts” is a nickname. The sentence says “one row is a billing account; include accounts with a paid invoice in the last 90 days; exclude internal tenants; window is calendar month in UTC.” Until that sentence exists, every later join is a guess.
The NIST Privacy Framework belongs here when the sentence would otherwise pull person-level rows into an operations pack. Segment analysis on this pillar is an illustrative complex-dependency example. Do not write identifiable people. Aggregate to account, plan, region, or queue. A sentence that needs a person’s name to be true is the wrong grain.
This page is not a rewrite of how a semantic layer compiles a contract. That pillar owns who publishes the certified measure and how it compiles. This pillar owns the order of work before the number: the sentence must be spoken, bound, and visible even when no certified contract exists yet.
After the run, open the trail the way explainable AI data analysis describes. That is P55. P70 binds the sentence before SQL is interesting.
If the organization already stored last quarter’s sentence, reuse it from organizational analysis memory rather than inventing a new sentence in chat.
A four-layer harness framework
Score the pack with one table. A metric definition that cannot sit in the third row is not ready.
| Layer | Object you must see | Pass signal | Fail signal |
|---|---|---|---|
| Direction | The business question in one sentence | A reviewer can restate the decision | A vibe, a dashboard name, or “just explore” |
| Steps | Locate, filter, join, calculate, check | Each step is named and reopenable | One giant query with no names |
| Metric definition | Grain, filter, window in one sentence | A colleague can edit the sentence | A nickname with no grain |
| Number | The computed value plus the trail | The figure matches the bound sentence | A confident digit with no sentence |
The model can plan. The harness decides whether that plan looks like a careful analyst or like a chat that races to a digit. The third layer is how the harness forces the sentence to exist.
State five facts before any calculate step: the data source, why this table, the filters, the sort, and the juxtaposition rule. Those facts make the sentence operational.
Microsoft Azure architecture guidance is a later industrialization map, not a reason to skip today’s lock. Apache Airflow documentation describes scheduled graphs. A harness outline is not a DAG. Do not confuse a scheduler with a bound sentence.
Three layers stay distinct: business-question mapping; locate / filter / join / calculate / check; inspectable, editable, reusable steps. The sentence lives in the first layer and is consumed by the second. Start in the second and you will compute a number no one can defend.
How this work differs from a semantic-layer purchase
Teams hear “metric definition” and open a buying page for a metrics store. That is the wrong first move for an agent session.
Choose a compiled contract if
You already have a published owner, a compile path, and many products that must reuse the same certified measure. Then a semantic layer is the industrial form of the same sentence. Buy or build that surface when the sentence is stable and the audience is wide.
Choose a harness-bound sentence if
You are answering one decision this week on authorized sources, and the sentence is not yet certified. Bind the sentence in the outline. Let operations read it. Compile later if it survives review. P9 is how contracts compile. P70 is the order of work before the number.
Choose an open-SQL review if
The number already exists and someone distrusts it. Open the trail. That is not a new bound sentence; that is P55. Do not pretend a post-hoc SQL dump is the same as binding the sentence first.
A Code Agent can hide steps in a notebook. A Data Agent cannot. Operations and compliance will ask who is in the filter. If the sentence is not readable, the path is not a professional analyst path.
Named views in SQL AI are how the language carries the bound sentence across steps. The language is not the sentence. The sentence still has to be written.
Tool landscape for outlines and bound sentences
You do not need a new aisle of “AI metrics” logos. You need objects a business reviewer can read.
ChatBI returns a digit and a paragraph. Spreadsheets hide the sentence in a comment no one opens. BI tiles hide it in a model the asker never sees. A harness that shows direction, steps, and the bound sentence before compute is the missing object.
InfiniSynapse’s published pattern matches that object: authorize a source, stop at the outline, bind the sentence, then accept SQL. It will not invent a certified warehouse of every KPI, and it will not write the sentence into production systems. Schema recall proposes tables. The reviewer still confirms why this table.
What the outline must expose
Source name. Why this table, not the adjacent one. Filters. Sort. Juxtaposition (for example, this month beside last month on the same grain). The metric definition sentence. If any of those are missing, refuse the number.
What still belongs in a later contract
Nightly certification, cross-product reuse, and an owner of record. Those are P9 jobs. Do not stall this week’s metric definition while a metrics platform program writes a RACI.
Refuse a tool that hides the sentence, a tool that wants write access to the system of record, and a tool that “explains” a KPI without naming grain. Accept on day one: a read-only source, a one-sentence metric definition, and a reviewer who can say no.
How to bind the sentence, then compute
Keep the sequence boring. Boring is how a metric definition survives a compliance question. The metric definition is the object a reviewer can quote without opening the model.
Write the direction
Input: one decision (“which account segment drove ticket load last month”). Acceptance: a reviewer can repeat the decision without naming a dashboard. If the ask is “show me activity,” you do not yet have a metric definition problem—you have a direction problem.
Name the mechanical steps
Input: candidate tables and the locate / filter / join / calculate / check list. Acceptance: each step has a name a colleague can reopen. If the plan is one unnamed blob, stop. A metric definition cannot attach to a blob.
Bind grain, filter, and window
Input: one sentence. Acceptance: grain, inclusion, exclusion, and window are all spoken. This is the metric definition. Bind it before compute. If two reviewers write two sentences, you do not have a metric definition—you have a conflict. Resolve it here, not after the digit.
Compute last, then check
Input: the bound metric definition plus authorized reads. Acceptance: the figure matches the sentence, the trail is open, and a check (row count, nulls, reconciliation) exists. If the figure is pretty and the sentence drifted, reject the pack.
Numbered work:
- State the decision in one line.
- List source, table reason, filters, sort, juxtaposition.
- Write the metric definition.
- Require a reviewer continue-signal.
- Run locate / filter / join / calculate / check.
- Open the trail and keep the sentence next to the number.
A metric definition that cannot survive step 3 should never reach step 5.
Desk sample: an illustrative account-segment pack
This desk is an illustrative composite, not a customer result. No person is identifiable.
A B2B operations desk asked which account segment drove support load. The harness stopped before “tickets per account.” Direction: enterprise versus mid-market after a billing change. Steps: locate invoices, seats, tickets, events; filter paying tenants; join on account id; calculate rate; check a monthly control. Bound sentence: one row is a billing account; paid invoice in 90 days; exclude internal and partner sandboxes; August 2026 UTC; same grain on both sides.
Illustrative object counts: 8 extracts, 42 named views, 88 dependencies. The first draft used “active user” and would have required person-level rows. The desk rewrote grain to account. When the illustrative 42 versus 18 appeared, a reviewer could edit “90 days” to “60 days” and reuse the steps.
Figure. Illustrative desk composite, not a customer result.
Rushed packs spent the clock on the number. The chart is illustrative, not a customer uplift.
Scorecard: ready to compute or still drafting
| Signal | Ready | Not ready |
|---|---|---|
| Direction | One decision sentence | A dashboard nickname |
| Source | Named, authorized, read-only | “The warehouse, I guess” |
| Table reason | Why this table, said out loud | First table the model saw |
| Metric definition | Grain, filter, window written | A KPI nickname |
| Juxtaposition | Compared objects share grain | Two grains on one slide |
| Path | Readable by operations | Only an engineer can parse it |
| Number | Accepted after the sentence | Accepted to end the meeting |
If that column is empty, the pack is a draft. Do not paste the digit.
Failure modes
Most failures are order failures, not model failures. This page did not run a live customer ask.
Digit first, sentence never
The agent returns 18,442. Someone pastes it. A week later two teams dispute who is in. There was never a metric definition. Bind the sentence, then compute. If a vendor boasts about model size, ignore it. This pillar does not score parameter counts.
Hallway grain
“Active” means seats in product, invoices in finance, and logged-in users in success. Each team thinks it has a metric definition. None wrote grain. Write one sentence or accept three truths.
Juxtaposition on mixed grains
Enterprise rate sits beside a person-level mid-market rate. The slide looks comparative. The metric definition was never shared. Force the same grain before you place two numbers next to each other.
When the missing object is the sentence itself, continue in the definition of metrics, which locks grain, filter, and window before any digit. If you need the brand object around the model, read data agent harness for direction, steps, metric, then number. If the gap is how an analyst actually works, use the data analysis workflow page to map, locate, filter, join, calculate, and check.
When you need an exit test, the definition of analysis writes done as an auditable pack rather than a fluent paragraph. If two tiles disagree, metrics meaning opens both sentences instead of averaging the digits. If reviewers only see a number, keep the four layers visible in the analysis workflow.
A bound sentence that cannot survive those six reads is not ready for a memo.
When the object itself is still a column nickname, start with what is a metric. When the method is write-then-bind, define metrics before the agent runs SQL.
When finance and product collapse a field into a KPI, split metric vs measure before you quote either digit.
Bind the metric definition, then ask the number
Write the metric sentence, pick the table and filter, then ask. Open the outline before you accept the figure. This check uses only sources you authorize.
Commercial association: You do not need the workspace to complete the educational diagnosis on this page.
Open InfiniSynapseHow this page is sourced. William Zhu is cofounder of InfiniSynapse (GitHub @allwefantasy); InfiniSynapse on GitHub. Company self-description, not independent authority. No personal LinkedIn is published. Evaluation basis: We evaluate (hands-on) by designing and reviewing analysis-pack methods—definition locks, read-only source binds, and downloadable
/tasksartifacts. Reviewed internally by analytics engineering · data platform · LLM security · editor. Editorial standards · corrections · publishing principles · About · Privacy · Terms · Contact zhuhl@infinisynapse.com. COI: InfiniSynapse sells an AI-native Data Agent; the banner is a commercial association. Fact-check: NIST AI Risk Management Framework · UK NCSC secure AI guidelines · NIST Privacy Framework · Apache Airflow documentation · Wikipedia SQL · Microsoft Azure architecture. No external organization audited this page. This page is not third-party recognition.
Frequently Asked Questions
Is a metric definition the same as a semantic layer?
Bottom line: No. A metric definition is the sentence you bind before this number. A semantic layer is how a stable contract compiles for many products. Do this week’s sentence even when no layer exists yet.
Can I skip the sentence if the SQL looks clean?
Bottom line: No. Clean SQL can encode the wrong grain. The metric definition is the human-readable test. Open the trail after, but bind the sentence first.
Why must operations be able to read the path?
Bottom line: A Data Agent is closer to operations and compliance than a Code Agent. If only an engineer can parse the plan, you do not have a metric definition that business will sign.
What three fields must the sentence name?
Bottom line: Grain, filter, and window. A metric definition that omits one of those three will invent a second truth the first time another team repeats the ask.
Conclusion
The model decides whether it can think like a person. The harness is what decides which role the model is allowed to play. Bind the metric definition before the number. Map the question, name the steps, write grain and filter and window, then compute. Keep the path readable by operations. A correct digit with no sentence is not a decision input.
InfiniSynapse describes itself on About. Privacy and Terms apply. If you later use the workspace, open InfiniSynapse only with authorized, sanitized inputs.