TattvaDocs
Reference

Glossary

Every term used in Tattva and in these docs, A to Z.

Terms in bold inside a definition have their own entry. Your project's own business terms (what "net revenue" means for you) live in Ontology › Make, not here.

A

Absolute threshold — A tracker rule. Every run of the question is marked Needs attention; the upper and lower values you enter are saved but not checked. See Tracker rule.

Accept / Defer / Reject — The three actions on a candidate in the review queue. Accept stores it, Defer parks it in the Deferred list, Reject drops it. From Deferred, Back to pending returns it undecided. In the Workspace Library these need the owner, admin or ontology admin role; in other projects any member can use them.

Agent's own SQL — A figure the agent computed with a query it wrote itself, because no stored definition covers it. It is labelled as such in the answer and the Trail, so you know it was not checked against a definition. When your analytics team limits figures to stored definitions (answer scope), the agent writes no such queries: the figure shows as "—" with a note that the metric is not defined yet.

All clear — A tracker status: the last run was not flagged. Questions with no rule, or with an LLM judge rule, also show All clear, so check the rule before reading it as healthy.

Answer scope — What your analytics team has told Tattva to answer in a project: the What you can ask examples, the Scope line, topics not answered yet, the date each metric is available from, whether figures may come only from stored definitions, and which kinds of question are answered. See Answer scope.

Answer tab — The first tab on an answer card, with the written answer. The second tab is the Trail.

Answered by — In the configuration document's word meanings, the stored metric that a word meaning's text or target names, so a reader sees which metric answers the word.

Archetype — The label on a Decision Sign that sums up how a team decides.

Available from — The first date a metric's figures are valid from, for example 1 Apr 2025 for a P&L that starts then. The part of a period before it is answered with "not available before …"; the rest of the period is answered.

B

BigQuery — Google's data warehouse. Tattva queries your BigQuery tables in place; the rows stay in your warehouse. Stored definitions run on BigQuery.

Board — A page where you pin whole answers with Pin to board in the answer toolbar.

Build — A run of the ontology pipeline, started with Regenerate on the Ontology page or Rebuild on a banner. It reads your sources and documents, proposes candidates, and re-checks that stored definitions still compile.

C

Calculate — The tool the model uses for arithmetic inside an answer, so sums and ratios are computed, not guessed. It accepts only figures the answer's own queries returned (or numbers in the question); a typed figure that is not one of them is refused.

Candidate — Something the build proposes and a person must review before it takes effect: a metric definition, a family, a filter, a settlement rule, a column meaning. Candidates wait in the review queue.

Changes by product — Part of Comparisons (computed in code) for a comparison broken down by product (or another key): each row's prior and current figure, change and percentage, marked new or gone when it is in only one period, and the biggest falls and rises. The answer copies them.

Check — One test run against a proposed metric definition (V1–V12) or family (F1–F5). Each ends as pass, fail, or not run ("skipped: …" with the reason). V1–V4 always block Accept when they fail; the checklist on the card marks every blocking check. See Review what Tattva proposes.

Clarification — A question Tattva asks you when one input is missing and can't be worked out. It appears as a Needs clarification card with clickable choices and a box to type your own answer. It asks for one input at a time.

Clear stored answer — A button under an answer that was stored for Repeat questions, or one that reused an earlier answer. It stops that answer being handed out to anyone; the next time the question is asked, it is answered in full. Owners, admins and ontology admins can use it.

Column profile — What Tattva records about a column: type, an approximate distinct count, and for text columns with up to 1,000 distinct values, the 20 most common values. The Knowledge Graph source drawer shows it as "≈N distinct" and "common: …". Common values are never recorded for columns whose names look like personal data or that are marked personal or restricted, and policy-tagged columns aren't profiled at all. BigQuery profiles are refreshed after 7 days.

Claim check — How the prose of an answer is checked. A model writes the prose; a second model, the reader, lists every claim each sentence makes (a number, a rise or fall, better or worse, a size word, a rank, "all", a comparison, a share, a cause); code checks each claim against the figures. A sentence that fails is revised once, then dropped; when the first sentence drops, code's verdict line replaces it. Answers recorded before it had up to three sentences of commentary instead.

Comparisons (computed in code) — The block in a Compute card that shows period-over-period changes, and for a breakdown the changes by product (or other key). Code works them out, not the model.

Compute card — The part of the Trail that shows each metric Tattva computed: the value (or a table), Provisional or Incomplete notes, Filters, a Parameter / Value / Source table, Joins, the SQL, and Comparisons (computed in code).

Concept — A business idea your team uses that is not a number, for example "new customer".

Conflict — A disagreement Tattva found, for example two definitions of one metric or two sources that disagree. Listed under Ontology › Maintain › Conflicts. LOW conflicts are advisories: hidden on that tab until you click Show N low-severity advisories, and never mentioned in answers.

Convention — A rule your business runs on, such as the financial year or the reporting currency.

Correction — What you leave with Not satisfied under an answer: the reasons you pick (figure looks wrong, definition is not what we use, wrong period, breakdown missing, wording unclear, other), what was wrong and what is correct. It changes no answer, not even yours, until the analytics team settles it in Maintain › Feedback (New, In review, Fixed, Not changed) — by making a convention or correction card in the review queue, or marking it Applied (fixed in the ontology, naming what was changed), Already covered (naming the existing rule) or Not applied (with a reason). You see its status under the answer and in Review › My corrections, and get a notification when it is settled. Satisfied is saved too.

Critical — The most severe tracker status. Set by a Relative delta rule when the change is larger than its threshold.

Curation pending — A banner on Ontology › Make saying edits (such as column meanings) are saved but not yet in effect. They take effect after the next build.

D

Dashboard — A report your team keeps (for example in Looker Studio), registered in Make › Dashboards and linked to the metrics it shows. When Repeat questions is on, a question it fully covers gets its link instead of a full answer, with Answer here instead.

Data analyst (DA) — The person who connects data, writes metric documents, reviews candidates and keeps the ontology healthy. See For data analysts.

Decision — A logged call with its evidence, hypotheses and actions. Statuses: proposed, accepted, invalidated, resolved, archived. Opened at /reasoning/decision/[id] with the tabs Summarised, Deep dive and Trail.

Decision Sign — A profile of how a team decides, drawn from its decision history.

Derived date — An input worked out from the question's period instead of being asked, for example "30 days after the period ends" or "today minus the settlement days".

Direction (of a metric) — Which way is good for a metric: higher is better, lower is better, closer to a target, or not set. A movement (rose, fell, up 4%) is a fact any answer can state. A judgement (better, worse, improved, deteriorated, strong, weak) is said only when it agrees with the metric's direction. With no direction set, answers use movement words only. Set it on the metric's Definition tab or on the Metric directions review card.

F

Family — A group of related metrics that people call by one name, for example several versions of "revenue". Each member is a variant, and the family has a variant policy.

G

Goal — A measurable target on Overview › Goals, created with Add goal or Make a goal on a triage card. A goal can link to trackers.

H

Hypothesis invalidated — A tracker rule. Every run of the question is marked Needs attention.

I

Incomplete — A figure for a period whose data doesn't all exist yet, for example a cohort whose outcome window hasn't closed. It stays incomplete until the window's end date plus the settlement days. The Compute card says so.

Input (also parameter) — A value a stored definition needs to run, such as the start and end date or a channel. Tattva fills inputs from the question's period, derived dates, values you name in the question ("from X", "to Y"), or a clarification.

K

Knowledge Graph — The map of your sources and how they join, under Ontology › Manage. Click a source to open its drawer.

Known limits — Free text an analyst writes in Settings › Project › Known limits about what answers cannot do yet. It is printed as its own page in the configuration document; answers do not read it.

L

LLM judge — A tracker rule whose instruction is saved but not used to judge runs. Runs show All clear.

Link role — How a tracker relates to a goal: primary, leading indicator, lagging indicator or guardrail.

Line of — A link saying a metric is one row of a stored statement (for example mrp is a line of d2c_pnl). When the statement runs for a question, the metric is not run again: its figure is that row. Tattva proposes the link on Maintain › Conflicts when the two give equal values over the same period; an analyst accepts it. See Knowledge graph.

M

Make / Manage / Maintain — The three groups of tabs on the Ontology page. Make is what exists (terms, metrics, definitions, the Workspace Library). Manage is how it connects (the Knowledge Graph, References). Maintain is what needs fixing (Health, Conflicts, Review, Promotion, Versions).

Metric — A number your business tracks, such as net revenue or RTO rate. A metric may have a stored definition or a formula.

N

Needs attention — A tracker status. Every run of a question with an Absolute threshold or Hypothesis invalidated rule gets it.

Not answered yet — A topic your analytics team does not answer yet, for example a metric whose data is not loaded. Asking about it gets the team's reason and questions you can ask instead. The list is under Not available yet on an empty Reasoning page.

Not run yet — A tracker status: the question has never run.

P

Parameter — See Input.

Phase — One step of answering a question, as the Trace and Trail name it: Analyze, Select, Execute, Reflect. An answer code lays out from your metrics shows only Analyze and Reflect.

PII clearance — A member's level of access to personal data, set on Settings › Workspace by an owner or admin. See Sensitive data.

Project — One working area with its own sources, ontology, answers, trackers and decisions. Create one from the workspace switcher with New project.

Promotion — Making a project's source available to the whole workspace. Request it with ↑ Promote to workspace on the source; a reviewer handles it in the Promotion Inbox.

Provisional — A figure that includes days that have not settled yet. Settled days use actual figures and the unsettled tail is estimated. The Compute card says "Provisional: days from … are not settled yet."

R

Reading the numbers — The first page of the configuration document: scope, currency, timezone, settlement, available-from dates, and every convention tagged Reading the numbers on Make › Data › Conventions.

Reflect — A phase on answers recorded before the claim check: the commentary and its check, or the self-check of an answer a model wrote. New answers do not run it.

Relative delta — A tracker rule. A run is Critical when the measured change, in either direction, is larger than your critical if Δ > percentage; otherwise it is All clear. The critical if Δ < value is not used.

Report — A saved, structured write-up of an answer, created with Save as report. Listed on the Reports page.

Repeat questions — A project setting (Settings › Project). When on, a question a registered dashboard covers gets the dashboard's link, and a question answered recently with the same metrics, dates and filters, and no data changed since, gets that earlier answer ("Answered on …; the data has not changed since") with Ask again. Only answers that passed their checks cleanly and were not marked Not satisfied are reused; see Clear stored answer.

Request check — The step after Analyze where code checks the model's reading of a question against your project: which stored definitions it names, whether a breakdown and the values it names exist, which question kinds are switched off, the family policy, and the layout the answer gets. A reading that failed is asked again, never guessed.

Review queue — Ontology › Maintain › Review, where candidates wait for Accept / Defer / Reject and correction cards the team made from answer feedback wait for a decision. It has four tabs: Pending, Deferred, Resolved (the history of every decision, with who decided and when) and My corrections (the comments you left under answers, with their status).

runMetric — The single piece of code that computes a metric. It runs the stored query, or builds the SQL from a formula with its filters and checked joins. Answers, trackers and Try it all use it, so they give the same number.

S

Scope line — The Scope: line under every answer, saying what the data covers (for example "D2C website orders only"), as your analytics team wrote it. Figures outside it are not in the data. When the question narrowed it ("Store RTO"), it ends with what this answer covers, for example "This answer: Store only."

Session — A conversation on the Reasoning page. Reopen past ones from the Sessions drawer.

Settlement — The number of days after which a metric's figures stop changing, for example once returns are known (settles_after_days). Periods that cross that line are Provisional.

Skill — A reusable, versioned analysis the agent can run, listed on the Skills page. A question it fits is answered by the planner following it as a playbook: the Trail's Steps header names the skill and its version ("Followed: rto-drivers v18"), the Source line names the stored definitions the answer uses, and the answer is not reused once a new version is uploaded.

Skipped — A tracker status: the run did not compute the question. The question shows "Skipped —" followed by the reason.

Source — A BigQuery table or an uploaded file or document.

Stored definition — A metric's SQL, written by a data analyst and kept whole, with its declared inputs. It is checked by running it (V1–V12) before it can be accepted, and it is what runMetric runs.

T

Total row — The last row of a table code lays out, shown only when the rows can be added: amounts are summed, and a ratio (such as a margin) is never added up. The same one rule check decides every total, share and "rest" on every path, planned answers included. For a figure by day, week or month that does not add up, the Total row is the definition's own figure for the whole period, with a note when the rows add to something else.

Trace — The right-hand rail on the Reasoning page that shows the live progress of a run.

Tracker — A saved question that re-runs on a schedule (hourly, daily or weekly, or paused) and flags changes. Found under Trackers in the sidebar.

Tracker rule — How a tracker question is judged on each run: No rule, Relative delta, Absolute threshold, Hypothesis invalidated or LLM judge.

Tracker status — The state of a tracker question after its last run: Critical, Needs attention, All clear, Skipped or Not run yet.

Trail — The second tab on an answer card. It keeps every step and every attempt, including withdrawn attempts, plus the Compute card and any Metric family block.

Try it — A panel on a metric's Definition tab that runs the stored definition for dates you pick, so you can compare it with your own number.

U

Unconfirmed notes — Audit notes disputing a figure that code could not confirm from the results. They are listed in the Trail under the Reflection, never under the answer.

V

Value grouping — A named group of raw column values, such as several spellings of one warehouse grouped as "Warehouse 1". Made in Knowledge Graph › source drawer › Schema › Values › Group as. Inputs on that column accept the group name.

Variant — One member of a family.

Variant policy — What Tattva does when you ask for a family by name: Show all variants, Ask which one, Use one by default and mention the others with their figures, or Use one by default, don't mention the others.

W

Withdrawn attempt — On answers recorded before the claim check, an answer draft the self-check rejected. The Trail keeps it and marks it "Withdrawn by the self-audit", with the reason.

Workspace — Your company's space in Tattva. It holds members and projects.

Workspace Library — The workspace-wide set of metrics, concepts and conventions that every project inherits. Found under Ontology › Make › Workspace Library.

Glossary — Tattva Docs