Nothing from a workbook is used until an administrator has confirmed how Gable read it, sheet by sheet. When a sheet says its figures are in thousands, Gable quotes that text and asks a required yes or no before you can confirm; it asks about a stated scale and never guesses an unstated one.

A blank cell is never treated as zero, a dash that formats a zero is a real zero, “n/a” is recorded as not applicable, and unreadable text is flagged, not dropped. Your figures never enter the searchable text of your documents; a workbook’s entry carries only labels, and a guard blocks any digit of a stored figure. A re-uploaded file keeps its old version for history and answers only from the newest one you confirmed.

Every calculation runs in code with exact decimal arithmetic, so a tenth plus two tenths is exactly three tenths. Every computed answer lists each figure it used as written, the calculation in plain words, and one rounded result. A total is never added to the lines beneath it: a sum runs over line items or over one named total, never both, and two different totals are never added together.

If a needed cell is blank, not applicable or unreadable, the calculation stops and the reply names the cell. Workbooks are compared only when their periods line up, weekly is never mixed with monthly, and when confirmed workbooks disagree Gable names both figures instead of choosing. Every number in a reply is checked against what the engine produced; a flipped sign, a re-rounded figure or a second copy of the result replaces the reply with a fixed sentence from the engine’s output.

Forecasts use named methods from published forecasting texts, chosen by fixed rules from a line’s history, trend and seasonality; the language model never produces or adjusts a projected figure. Where the source text prints worked examples, Gable’s methods reproduce those printed figures in its tests; where a method’s parameters are not uniquely identified, the tests check that Gable’s fit is no worse than the text’s.

Every projection is a range around a central figure, never a bare point, and the reply says how the range was built. Each forecast carries a confidence label, high, medium or low, with every reason behind it: history, gaps, volatility, trend, distance ahead, and whether it beat the last-value benchmark. Gable may ask up to three questions first, such as whether an unusual period was a one-off, and applies only figures you type.

Every forecast restates each assumption it used, defaults included, and carries the same accuracy note: simple methods on monthly business data typically land within 11 to 15 percent one period out and worsen further ahead, an optimistic floor. Gable refuses to forecast for eight stated reasons: fewer than six periods, irregular columns, workbooks that disagree, gaps it cannot fill, a level shift in the last six periods, figures too volatile for any method to beat the last value, a non-administrator surface, and a ratio rather than a single line.

Administrators can ask for ratios in the forms lenders use, following the RMA Annual Statement Studies definitions, with days measures on one convention and the turnover form shown. A ratio is computed only when every line it needs is mapped; otherwise the reply names the missing one. When cash moved differently from profit, Gable builds the bridge from your mapped lines and either reconciles it to the actual change in cash within a dollar or names the unexplained amount; a term whose line is not mapped is named as unquantified, never estimated.

Guards refuse or caveat in plain words: no return on equity on zero or negative equity, no liquidity ratios without a current split, no gross margin without a cost of goods sold line, no percentages of zero or negative revenue, and a single period is called a snapshot. A benchmark is mentioned only with its source and vintage and the note that it is a guideline, not a norm; when Gable has none for a measure it says “no published benchmark on file”.

A business review opens with a one-word verdict chosen by fixed rules in code, and the reply must repeat that word and no other. Every finding is labelled observation, judgment or recommendation, and recommendations are options for the owner, never instructions. Every review ends with the informational-only line that closes this page.

• It will not answer forecast or analysis questions on the public assistant, customer links or text messages; those are for administrators.
• It will not treat a blank cell as zero, add a total to its own lines, or compare periods that do not line up.
• It will not give a forecast as a single number, forecast from under six periods, or apply a figure you did not type.
• It will not compute a ratio whose lines are missing, name a cash cause its bridge does not list, or quote a benchmark without source and vintage.
• It will not tell you what to do, state the future as certain, or give personal financing, tax or investment instructions.