How it works
Nothing to hand? Load the , where the activity notes reach every flagged line and the commentary comes back complete. Or the , where the notes cover three lines of nine and six honest gaps is the right answer. Or the , where one column of one subtotal is out by eleven thousand and a single line turns out to be decisive on its own. to watch the export refuse itself. All three replay a saved run for free.
1. The pack is read, not assumed
Comma, tab, semicolon, pipe and markdown tables all parse, and the column roles are matched by name so a sheet headed Actual / Prior / Plan works untouched. Parenthesised values are negative, currency symbols and thousands separators are stripped, and a magnitude suffix counts only when it is welded to the digits — a space in front of it is how $1,125,000 b becomes a quadrillion. When no header can be recognised the columns are read positionally and the assumption is printed rather than hidden.
2. Polarity comes from the caption
Favourable and unfavourable are computed, not typed in, which means the caption has to be read. Unambiguous cost markers win outright, so cost of revenue is a cost; balance-sheet captions come next, so deferred revenue is neither; and what is left is resolved head-finally, because English puts the head last — subscription revenue is revenue and income tax expense is an expense. A plain expense-first rule gets one of those wrong and an income-first rule gets the other.
3. A percentage is refused where it would lie
Against a zero comparative, or across a sign change, a percentage movement is not meaningful and is printed as n/m with the reason rather than as a number nobody can use. The amount test still runs, so those lines are still flagged when they matter — which is the point: the inconclusive case is resolved, not skipped.
4. The subtotals are footed
Every stated subtotal is added back against the block above it, in all three columns, under both readings — the plain sum a total caption means and the polarity-signed sum a gross profit caption means — and the reading that footed is named. Where the pack carries a balance sheet, the identity is checked per column too. Commentary written over a table that does not add up is commentary on the wrong numbers.
5. The decisive line is found by pulling it
Not by ranking magnitudes. Each contributor is genuinely removed and the net earnings movement re-measured, and a line counts as decisive only if pulling it alone brings the movement inside the floor. A large favourable line and a large unfavourable one can each be individually decisive while a ranking shows only one of them.
6. The pass may only explain what the notes cover
Before the model runs, the engine measures which flagged captions the activity notes actually mention. The rest are named to the model as unexplainable and its only permitted answer for them is the skill's own wording. Afterwards the same measurement checks it: exactly one driver per flagged line, counted rather than sampled; any driver sourced to the notes for a line the notes never mention reported as an invention by name; any driver that merely restates the movement flagged as a restatement; and every currency and percentage figure traced back to something measured, at the precision it was written to.
Method derived from @anthropics/variance-commentary (anthropics/financial-services-plugins). Flux Desk is a derived work, not a republication of that skill.