Cell details, reviews and history

An AI answer is only useful if you can check it, correct it and show where it came from. Every cell of a table has a details panel that shows its value, how sure the AI was, the pages and web sources it used, and its latest runs and actions (up to 20). From there you approve or reject the value, lock it so that nothing overwrites it, ask for a better answer, or bring back an earlier one. Rows that a column flags for a person collect in the review queue.

Before you start

• The details panel is part of every plan. Approving, rejecting, locking and restoring need edit access to the table; with view-only access you can read everything.
• Rerun and Refine run the AI again and cost credits like any run (see Columns and AI).
• The history keeps the latest 20 runs and actions of a cell, from when the cell was first computed. A cell with no entries yet says "No changes recorded for this cell yet."

Open the details of a cell

• Right-click a cell and choose View details.
• Or select an AI or computed cell and press Enter.

The panel opens on the right of the grid. While it is open it follows your selection: click another cell and it shows that cell. Drag its left edge to make it wider, use the expand button to give it more room, or close it with ×.

The header shows the column's name and the cell's status:

What the panel shows

From top to bottom, depending on the cell:

When an answer holds a list ("This answer contains 6 items"), the panel offers Insert as new page (the items become rows of a new page) or Ask AI to split (the chat builds a sub-sheet or separate columns from it).

Approve or reject

Every AI cell has a review state: pending, approved or rejected.

• Approve marks the value as checked.
• Reject marks it as wrong.

Whenever a new value is written into the cell (by a run, a refine or a restore), the state goes back to pending, so an approval never refers to a value nobody looked at. Approving and rejecting are recorded in the cell's history ("Approved by reviewer", "Rejected by reviewer").

Correct a value

Edit value (in the panel, or in the cell menu) lets you type the right value, then Save. The value is kept as "Manually edited by user" and recorded in the history as "Manually corrected". It stays until the cell is computed again: Rerun, Recompute all cells, a changed prompt or model, or Run AI and Update after one of its inputs changed. To protect it for good, lock it.

Ground truth

Lock freezes the cell's value as ground truth: the value you checked by hand and want no automation to touch.

• Runs and refines still compute, but the new result is not written into the cell. The panel shows it instead: "Rerun produced 3 450, locked value kept".
• Formulas skip a locked cell when they are applied.
• You cannot type into a locked cell, refine it ("locked · unlock to refine") or restore an earlier value into it.
• The record window shows the field as "Locked as ground truth".

Unlock returns the cell to normal.

Refine an answer

Refine runs the AI again for this cell with a follow-up instruction that applies to this run only, without changing the column's prompt. For example "exclude one-off setup fees".

1. Type the instruction in Refine (up to 20,000 characters).
2. Press Enter (Refine (Enter)).

The new value replaces the old one (unless the cell is locked) and the instruction is kept under History. A column-wide refine is in the column's settings: "Re-runs every cell of this column with this follow-up instruction. Applies to this run only, the prompt above stays unchanged. Ground-truth cells record drift instead of being overwritten." (see Columns and AI).

Rerun runs the cell again with the column's prompt as it is.

The history of a cell

How this was decided lists the latest 20 runs and actions of the cell, newest first: Run 3 with current on the one shown now, what started it, who started it, when, and the value it produced. A line says what is kept: "Only the latest 20 changes to this cell are kept, approvals included."

What started a run:

A run can also be marked failed, → agent (handed to the research agent) or ground truth kept.

Open a run to see the full record:

• Instruction used: the column's prompt exactly as it was then, with any follow-up ("Plus follow-up:").
• Sources used: every passage, value and page it read, each marked Document, Row text, Column value, Knowledge base, External data, Web page or Previous value, with how well it matched ("match 87%") and Open source document. A line says how the search went: how many files were searched, how many passages were considered, how many were dropped as irrelevant.
• Model's explanation, the Exact prompt sent and the Raw model response.

Restore an earlier value

1. Open a run in How this was decided.
2. Click Restore this value (Restore the value from run 2).
3. Confirm: "Set the cell back to this run's value? The current value is replaced (the change is recorded in this history)."

The restore is added to the history as a new run; the earlier values stay in the list. It is not possible while the cell is locked.

The review queue

A Single choice column can act as a gate: in its settings, Route to human review when picks the values that send a row to a person: "Rows landing on these values queue for human review, the end-of-pipeline gate." (see Columns and AI).

When a table has such a gate, the toolbar shows the review button ("Answers waiting for your approval") with the number of rows waiting. It opens Human review, "routed by validation gate":

1. Each card names the row and lists its problem columns with the value the AI found, or "not found" or "failed", the page chip of its source and "Model reasoning: …".
2. Type the right value into the box ("enter Kwota…") and press Enter or Resolve. "resolving writes the value + marks the cell approved · downstream reasoning then unblocks for the row": the columns that depend on it are marked out of date and computed again.
3. Open opens the whole row; Delete row deletes the row from the table, after the same confirmation as deleting rows in the grid ("The rows are deleted from the table, with their values and comments. Source files stay in Files. You can undo this from Undo history."); the table's Undo history brings it back (see Undo).

With nothing waiting: "No rows need review". Without AI on your plan, resolving still writes the value and approves the cell, but the columns after it are not computed again until AI is on.

Good to know

• Approval is per value. A new value always needs a new approval.
• Locked is stronger than edited. An edited value can be recomputed by Rerun; a locked one cannot be changed by any run or formula.
• Comparing two models. Compare models (A/B)… in the column menu runs a column with two models and flags the rows where they disagree for review (see Columns and AI).
• The pipeline map sums up runs, failures, cost and average confidence per column (see Columns and AI).

Related pages

• Columns and AI: prompts, runs, branches and review gates.
• The grid: the cell menu and the toolbar.
• Web research: web sources and research runs.
• Formulas and totals: formulas skip locked cells.