How to Create an Assumption Register in Eight Steps

An assumption register is one table listing every input nobody filed, observed or measured. Each row carries six fields: the assumption, who made it, what it rests on, how far the answer moves if it is wrong, what would falsify it, and when it was last revisited. Build it in eight steps while the work runs, never afterwards from memory.

Here is why the table has to exist at all. Assumptions are invisible from the inside. A person who has decided that 60 per cent of a cost base moves with revenue stops seeing that as a decision within about an hour, and from then on it reads as part of the situation rather than part of the working. By the time the work is finished most assumptions have stopped looking like choices, so the person doing the work is the one person who cannot see their own. The register is the only instrument that catches them at the moment they are still visible, and that moment is the moment they are made. Which is why the eight steps below are as much about when as about what.

Evidence in finance work sets out what an assumption is and how it differs from an inference. The contents of a source log and the contents of a workpaper are set out under the decision record and the workpaper behind it. Both are assumed below. The eight steps have a fixed order, a fixed table shape, and a definable point at which building stops. The work used throughout is a single assignment: the Kavery research desk, four people inside Kavery Capital Services Private Limited, looking at a borrower the desk already lends to, Meenakshi Tubes Private Limited. Sharada Iyer produced the work, Prakash Nadar reviewed it, and Latha Menon decided on it. The subject of the assignment does not matter: a bond, a fund or a factory in its place leaves all eight steps reading the same.

When is the table opened, and what goes into it first?

Step one is to open the table before anything has been computed. Six columns, no rows, on the first day. On this assignment that day is 3 March, the day the question stopped being whether the desk should be worried about this borrower and became something answerable. An empty assumption registerOne table holding every input in an assignment that was chosen by the person doing the work rather than found in a document. on day one is a healthy document. A register created on day nine is a different document with the same name, and the difference is the subject of the failure block below.

Step two is the sweep. The work done so far is gone through, and every input that nobody filed, observed or measured is pulled out. The test is one question with a yes or no answer: is there a document, a reading or a count that produced this figure? If yes, it belongs in the source log with an item number. If no, it belongs here, in the register, with a row. Found things go in the log, chosen things go in the register, and no input belongs in both. A log with assumptions mixed into it stops being a list of where things came from, so the split is what keeps either table scannable.

The household version is the same move at a smaller scale. Take that one first. Somebody is working out whether a wedding fits inside Rs 6,00,000. The hall quotation is a document, so it goes in the log. The caterer quoted a per plate rate, and that is a document too. But the number of plates is 400 because that is what the guest list looked like in March, and nobody has counted anybody. The per plate rate is treated as holding at 400 plates even though the quotation was for 250. Neither of those two is written anywhere. The plate count and the rate are the two numbers the whole budget is standing on, and they are the two that will be forgotten first. By the time the arithmetic is done they have stopped feeling like decisions and started feeling like the size of the wedding.

On this assignment the sweep produced nine rows by the end. Six of them came out of the sweep itself. Three more arrived only at step three, and one of those three turned out to be load-bearing. Step three exists as a separate step rather than as a second look at step two for exactly that reason.

How does an assumption nobody noticed making get caught?

Step three is where the register earns most of what it is worth. The sweep at step two finds the assumptions the analyst remembers making. The sweep cannot find the ones that were never registered as decisions, and those are the ones that hurt. An assumption nobody can see is an assumption nobody can test. So the third step is to go over the same work a second time, asking three fixed questions in this order, out loud if that helps.

The three questions are: what number did I choose, what did I carry forward unchanged, and what would somebody else have done differently at this point. The three are not variations of one another. The first catches figures picked because a cell needed filling. The second catches figures inherited, from last year, from the previous assignment, from the template, and never re-examined, and carrying something forward is the one that hides best because it does not feel like a decision at all. The third catches the choices that felt like the only available option at the time. Looking like the only option is the disguise a genuine judgement wears from the inside.

Which prompt found which row is worth noting. The first found the cost split, a figure somebody typed. The second found the input cost assumption, a figure nobody typed at all. The assumption is the absence of a change, and an absence is invisible to a sweep looking for entries. The third found the December order, the row that turns out to move the answer furthest. Three prompts, three different kinds, and running only the first one keeps the cheapest row and loses the two expensive ones.

Which six fields does every row carry?

Step four fills the row. Six fields, the same six on every row, in the same order, whether the row is the biggest thing in the work or the smallest. The shape is fixed so that a half-filled row is visible without anybody reading it. A rigid layout earns itself on that alone. A column of nine rows can be scanned in four seconds and three blank cells show up at once. Nine paragraphs of prose can be scanned to no effect at all.

Two of the six fields do work that most people expect from a register, and four do work that most people never ask of one. Naming the assumption and naming who made it is the part everybody writes. The third field, what it rests on, either points at log items or admits it points at nothing, and a row resting on nothing is not a defect, it is a finding. The last field, the revisit date, looks like housekeeping and is not: a row nobody has touched since the first day is a row that has been carried unexamined through the entire assignment, and the date column is the only place that shows.

How is the field that says how far the answer moves written?

Step five is the field that earns the whole table, and it is the one most registers get wrong. Write the answer worked out again with the row wrong, as a figure, never as the word high, medium or low. A word in this cell tells a later reader nothing they can act on. A pair of figures tells them exactly how much of the conclusion is resting on this one line, and it does it without them having to trust anybody's judgement about what counts as high.

The procedure for the sensitivityHow far the answer moves if an assumption turns out to be wrong, written as a distance between two figures rather than as a word. field is mechanical. The row is taken, assumed wrong in the direction that matters, and the answer worked out again. Both numbers go in the cell, the one that was published and the one that would have been published, with the units on both. The field is those two figures and nothing else. Filling it takes four minutes on most rows and about twenty on the worst one, and the twenty minute row is always the row that most needed the time.

The same field works on the wedding. Back to it for a moment. The row says the caterer rate holds at 400 plates when the quotation was written for 250. Sensitivity written as a word: moderate. Sensitivity written as a number: the hall is Rs 2,40,000, and at the quoted Rs 900 a plate the food for 400 comes to Rs 3,60,000, so the budget lands exactly on the Rs 6,00,000 limit. If the rate on the extra 150 plates is a fifth higher, at Rs 1,080, the food is Rs 3,87,000 and the budget is Rs 6,27,000. One row moves the answer from sitting on the limit to Rs 27,000 outside it. The second version tells somebody what to do with their afternoon. The first version tells them somebody had a feeling.

Now the row on this assignment. Row two says the December order repeats. The December order produced 14 per cent of the revenue of the year just reported. The desk had already recorded, as row four, an assumption that 60 per cent of total cost moves with revenue and 40 per cent does not, so the recomputation is fully specified. The desk holds money in whole rupees in this workpaper. Here is the whole build.

All figures are in whole rupees. The 10.8 per cent is the margin as filed, being Rs 2,30,00,000 over Rs 21,20,00,000. The 5.0 per cent is the scenario, and the cell in the register carries both, in that order, with the word scenario attached to the second so that nobody reading it later can mistake a constructed case for something the accounts said.

Because a reading that lives only inside an interactive is invisible to anybody who cannot run it, here are the figures worth writing down. At zero, nothing is taken out and the margin is the reported 10.8 per cent, already 0.2 points under the floor. At 5 per cent it is 9.0 per cent. At 10 per cent it is 6.9 per cent. At the 14 per cent that row two actually assumes, it is 5.0 per cent. At 20 per cent it is 1.9 per cent. Now read the same five numbers as steps rather than levels: the first five points of revenue cost 1.9 points of margin, the next five cost 2.1, the five after that cost 2.3, and the last five cost 2.6. Each step costs more than the one before it, and the reason is the pine block in the bar: the fixed Rs 7,56,00,000 is the same width at every setting of the slider, so every rupee of revenue that leaves has less and less company on the way out. The rising cost of each step is what a sensitivity field is really recording, and a cell reading medium records none of it.

What would show this row is wrong, and when is it written?

Step six fills the fifth field. The single observation that would show the row is wrong is named, and named before the answer is known. A falsifierThe specific thing that could be observed that would show an assumption is wrong. Naming one turns a belief into something that can be checked. written after the conclusion is settled is always one the conclusion already survives, and nobody does that on purpose. A late falsifier gets chosen through an answer already believed, and the test that comes to mind first is the gentle one.

A falsifier is usable when somebody else could go and look for it without asking the author what was meant. On row two, the falsifier is the order failing to appear in the next volume series, or the finance head describing it as one-off. Both are things a person can go and check. Set against that, a falsifier reading if demand conditions deteriorate cannot be checked by anybody, cannot be dated, and quietly means whatever the reader wants it to mean six months later.

There is a second reason to write this field early, and it is the reason the field survives arguments about whether it is worth the time. The falsifier is what a revisit triggerThe event that means a row has to be looked at again, such as a new filing, a date passing, or something said that contradicts the row. is built out of at step eight. Once what would show the row is wrong has been written down, most of what to watch for has been written down too. Skipping step six leaves step eight with nothing to attach itself to. Registers end up with a revisit column reading as required all the way down.

Which rows is the answer actually standing on?

Step seven sorts the table. The first row is removed, and the conclusion is tested for whether it still holds without it. The row goes back, and the second row gets the same treatment. Nine rows, nine passes, and at the end of it the rows the answer is standing on are separated from the rows that are furniture. Removing each row in turn is the drop-one testRemoving each assumption in turn, one at a time, to see which removals change the conclusion and which do not., and it takes about half an hour on a table this size.

On this assignment three of the nine changed the conclusion when they were removed. The first is that the filed accounts are final: remove it and the 10.8 per cent is provisional, so there is no settled answer to the first half of the question. The second is that the December order repeats: remove it and the second half of the question, what could move the margin most, has no answer at all. The third is that input cost stays inside the range of the two reported years: same effect, and this is the row the sweep missed and the second prompt caught. The three survivors are the load-bearing rowsThe rows whose removal changes the conclusion. Everything else in a register shapes the context without carrying the answer. of this register. The other six stay in the table, and staying is not a consolation prize.

One detail is the part people argue about, and it is worth slowing down for. Row four, the cost split, is not load-bearing. The reported 10.8 per cent comes from a filed revenue and a filed operating profit and needs no split, so removing row four does not move it at all. But row four is the row that row two borrows to compute its own sensitivity figure. A row can be necessary to another row and still not be carrying the conclusion, and the drop-one test is asking only the second question. Keeping the two apart matters. Otherwise half the table ends up marked load-bearing and the signal the test exists to give is lost.

When does a row get looked at again, and who does the looking?

Step eight closes the table. Every row gets a trigger, meaning the event that says this row has to be read again, and every trigger gets a name attached, meaning the person who will notice it. Three kinds of trigger cover almost everything. The first is a date, and a date goes in a diary. The second is a filing or a document arriving, and somebody has to be watching for it. The third is a conversation that contradicts the row. Nobody can schedule one of those, and it is caught only because somebody remembers the row exists.

On this assignment the dated trigger is the review that falls 90 days after 12 March, when the decision was recorded and the line was reduced from Rs 40,00,000 to Rs 25,00,000. Latha Menon is named against it because she made the decision. The filing trigger sits on row one: a revised set of accounts would land straight on the assumption that the filed accounts are final, and that row is load-bearing, so the trigger reaches past the row and into the conclusion. A trigger on a load-bearing row is not a request to update a table, it is a request to reopen the answer.

What does the finished register hand over?

The register is not a filing cabinet. The register produces one visible thing, and that thing is the last sentence of the work. The closing lineThe last line of a written note, saying plainly what would change the answer it just gave. is read off the load-bearing rows rather than composed, and reading takes minutes where composing takes an afternoon. Three rows survived the drop-one test on this assignment, so the closing line names three things, in the order the rows sit in the table, and adds the one sensitivity figure that is worth carrying out of the working and into the note.

Naming the rows in table order does something particular for a reader who arrives with none of the author's context. The reader does not have to work out which parts of the answer are solid and which are propped up. The last sentence tells them. One of the three rows carries a figure, so nobody has to guess how much is at stake in the propping. Each row behind the three names what would falsify it, so all three can be checked. A closing line built any other way is a summary. Built this way it is an instruction for what to watch.

Who reads the register apart from the person who wrote it?

Three people read this one, and each of them did something different with it. The reviewer, Prakash Nadar, used it as the shortest route into work he had not done: nine rows told him in two minutes what had been chosen rather than found, and his whole review pass took 40 minutes against 9 hours of production. The person who made the decision, Latha Menon, used the three load-bearing rows to set the shape of the decision rather than just its answer. The further review in 90 days came from there. A decision that names the assumptions it rests on can be given a review date. A decision that does not name them has nothing to review, so it cannot.

The third reader arrived six months later and had never seen the work. Rebuilding the 10.8 per cent from the workpaper took that reader 25 minutes, and nobody was asked anything. A rebuild by a stranger is the only real test of a register, and the test is worth naming plainly. A table that a stranger can follow with nobody from the desk in the room is a table that works, and everything else is a table its author believes in. When a lender, an analyst or a committee reads the note, the register is what lets them argue with the conclusion instead of arguing with the person who wrote it.

The household version is smaller and just as useful. Somebody deciding whether to take a home loan writes down four things nobody filed: that the second income continues, that the rent they are paying now would have risen anyway, that the maintenance charge holds, and that nobody in the household has a large medical expense in the first two years. Then the sensitivity field: if the second income stops, the payment goes from comfortable to Rs 9,000 a month more than the household earns. The second income row is load-bearing. Everything after that is a conversation about one row instead of a vague worry about the whole decision, and that narrowing is the entire product of an assumption register at any scale.

The Kavery research desk, Kavery Capital Services Private Limited, Meenakshi Tubes Private Limited, Sharada Iyer, Prakash Nadar and Latha Menon are invented.
Educational material. Not advice on any investment, tax, budget or market position.