TattvaDocs
Workflows

Review an answer

How to check an answer's figures on the Trail tab, what the Compute block tells you, and what to do when something looks off.

Tattva computes defined metrics in code, lays out most answers in code, and checks anything a model writes before you see it. Still, an answer you are going to act on deserves a minute of checking. Everything you need is on the Trail tab.

The one-minute check

  1. Does the headline answer the question you asked? Check the metric, the period and the segment.
  2. Open Trail · N. Read How this was answered first: it lists only what changed against what you asked, or says "Answered as asked." Open Details for every metric, split, filter and period the question was read with, and where each Total came from. Then, on the Analyze card, find the Compute entry for each figure in the headline. A planned answer has Steps instead of phase cards: each step's Details give its definition, filters, period and split, and How each figure was derived traces each figure to its steps.
  3. Do the figures match? The value in the Compute block should be the value in the answer.
  4. Is the period right? The card shows the dates it used, and the Period: line at the end of the answer names them; a week or shorter is written with weekday names, for example Mon 12 Oct – Sun 18 Oct 2026. Weeks run Monday to Sunday.
  5. Any notes? Provisional or Incomplete means the figure will still change.

Read the Compute block

Compute block

The Compute block sits on the Analyze card: code computes these metrics before the answer is written. Each entry is one metric run. Its heading names the metric, the period, the role of that period (the question's window, the prior period, the same period last year), and whether the figure came from the stored definition.

PartWhat to check
The valueThe figure the answer should quote. A breakdown shows as a table.
ProvisionalSome days in the period have not settled yet. Those days are estimated, so the total will move.
IncompleteThe data needed has not all happened yet, for example a cohort still inside its window.
FiltersWhich rows were included. Is anything missing that you expected, or included that you didn't?
Parameter / Value / SourceEach input and where its value came from: your question's period, a date worked out from it, a value you named, or your answer to a clarification.
JoinsWhich tables were combined. Tattva only uses joins your data analysts approved.
SQLThe exact query that ran. Expand it to read or copy it.

Below the entries, Comparisons (computed in code) lists period-over-period changes. These are worked out by code, not by the model. For a comparison by product (or channel, or another breakdown), such as "top 10 draggers and lifters last week vs the week before", it also lists each product's prior figure, current figure, change and percentage ("new" or "gone" when a product appears in only one period) and the biggest falls and rises. The answer copies these figures; check them here.

When a figure is the agent's own

If a metric has no stored definition, or the stored one cannot serve this question, the entry reads "Refused: … — the answer computes this with its own SQL, labelled as such." A table from that query is captioned Tattva's own calculation. Treat these figures with more care: check the SQL, or ask a data analyst to define the metric.

If your analytics team limits figures to stored definitions, the agent writes no SQL of its own: such a figure shows as "—", and a note under the answer says the metric is not defined yet.

Scope and dates

The Scope: line under the answer says what the data covers, for example "D2C website orders only"; figures outside it are not in the data. It appears under normal answers only. A note such as "not available before 1 Apr 2025" means the metric's figures start on that date, so only the later part of your period was answered.

When a comparison's current period leaves out today (a partial day), the comparison period is cut to the same number of days, so like is compared with like. If it cannot be, the Period: line states both day counts.

Totals and notes from code

A Total row is written by code, and only when the rows can be added: amounts are summed, and a margin or other ratio is never added up. 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 line under the table when the rows add to something else. A long list shows the first 10 rows, one Other N line and the last 5; Copy tables (TSV) copies every row.

The Layout row of How this was answered › Details says, for each such table, where its Total came from: the overall metric's own figure, the sum of the rows and why they may be added, the definition's own figure for the whole period, or no Total and why.

Check how the question was read

How this was answered, the first card on the Trail, lists each part of the answer that differs from your question, with one status:

  • Adjusted: code changed what was asked, for example a period moved back because its last day is still loading.
  • Assumed: the words could be read more than one way and Tattva picked one, for example "last week" read as the calendar week, or "DFJ" read as DIA FREE JUICE.
  • Not applied: part of the question is not in the answer, for example words nothing in the project matches, or a split the definition does not have.
  • Switched off: the project does not answer this kind of question, for example a "Why …" question answered as a comparison.
  • Asked you: Tattva asked which value or metric you meant instead of guessing.
  • Not what happened: the question says a figure dropped (or rose) and it moved the other way.
  • Reused: an earlier answer shown again, because nothing it read has changed.

When nothing changed, it says "Answered as asked." Details has one row per part of the question (metric, period, split, filter, skill, layout, commentary), with a badge only when that part was Adjusted, Not applied or Asked you.

If a row is not what you meant, ask again naming the metric, split or period.

Check the Analyze step

On the Analyze card:

  • Parameters for this question shows each input the metrics need and where its value came from. A "Missing" line means Tattva could not fill one in.
  • Metric family appears when your word covered several metrics, such as "revenue". It shows which variant was used or that all were shown. If it picked a variant you didn't mean, ask again and name the variant.

Metric family

Withdrawn attempts

If the self-audit rejected a draft of an answer a model wrote, the Trail's rejected Execute card is marked withdrawn, with "Withdrawn by the self-audit:" and the reason, and a grey note after the answer reads "An earlier draft was checked and revised before this answer was shown". Read the final answer; the draft is there to show what was changed and why.

Withdrawn attempt

A withdrawn attempt is the system working, not failing. Read the reason: it tells you what the first draft got wrong.

Check the definition itself

To see exactly how a metric is defined, open Ontology → Make → Data → Domain, find the metric, and click Definition →. The Definition tab shows the SQL, the Parameters, the Validation results and the Source text it was written from.

Definition tab

Under Try it, pick From and To (exclusive) dates and click Run to compute the metric yourself. It should match the Compute entry for the same dates.

Try it

What is not checked

Be aware of these limits:

  • A number without working. A figure in a table that has no Compute entry or SQL behind it is not verified by the self-audit.
  • Breakdowns of some metrics by a dimension, such as net revenue by channel, may be the agent's own SQL rather than the stored definition. Look for the label.
  • Data from different places. Tattva never joins a BigQuery table to an uploaded file. A question that needs both is answered in parts.

When something looks wrong

  1. Ask a follow-up in the same session: "Show me the SQL for net revenue", "Use only Warehouse 1", "Compare with the same month last year".
  2. Tell the team. Click Not satisfied under the answer, pick the reasons (for example Figure looks wrong or Wrong period), say what was wrong and what is correct, then Send feedback. The analytics team reviews it and fixes the ontology so everyone gets the right answer; until then it changes no answer, not even yours. To get a new answer now, click Re-ask with this correction: it asks again in the same conversation with your note. Track your comments under Ontology › Maintain › Review › My corrections. If the answer was right, click Satisfied: that is saved too.
  3. Tell your data analyst if the definition itself is wrong, or a metric you rely on has no definition. They fix it once for everyone.
  4. Check Ontology › Maintain › Conflicts if two sources seem to disagree.
Review an answer — Tattva Docs