How to Build a Source Log for Financial Research

A source log is one row per item with seven columns: number, what it is, where it came from, when it arrived, its rank, whether a second item corroborates it, and its status. Each row is written when the item arrives, never afterwards. The rows are numbered. The number is what ties a sentence in the finished writing to the item behind it.

Here is what sits underneath that. A source logOne row for every item of evidence touched, kept while the work is still moving. assembled after the work is finished is not a record of the evidence at all. A log assembled that way is a summary of the argument, built out of whatever can still be called to mind on the last afternoon, and what can still be called to mind leans, with no dishonesty anywhere in it, towards the items that pointed at the conclusion reached. Timing is the only part of the procedure that cannot be repaired afterwards, so the timing rule matters more than any column in the table. A column can be added to a log on Friday. Tuesday cannot.

Start with something too small to feel like paperwork. A household has a roof that leaks and asks three people to come and quote for the repair. The quotes arrive over ten days, one on the phone, one on a sheet of paper left in the door, one from a cousin who came to look and talked for an hour. If somebody writes each one down the evening it arrives, with who gave it, what it covered and the date, the household can compare three things on a Sunday. If nobody writes anything down until Sunday, the quote that survives clearest is the one from the person everybody liked, and the household is now comparing a document with a feeling. Nothing about that failure is careless. Memory simply does that with ten days.

Sharada Iyer, an invented analyst seven months into a job at the Kavery research desk, part of Kavery Capital Services Private Limited, had nine days to answer one question about a borrower named Meenakshi Tubes Private Limited. The file was opened on 3 March and the note went out on 11 March, and across those nine days fourteen items of evidence arrived by six different routes. The seven columns behave the same way whether the subject of the work is a borrower, a bond, a warehouse or a household budget, so what Meenakshi Tubes Private Limited manufactures never once matters below.

The seven steps below cover opening a log, setting its seven columns, writing a row in under thirty seconds while the document is still on screen, marking a second independent item against the first, retiring an item that has been beaten without erasing it, and pointing from any finished sentence back to the row it stands on.

What are the seven steps, in order?

Seven moves, and only two of them happen once. Steps one and two are set-up, done before any evidence exists. Steps three and four repeat for every single item that arrives. On this assignment that meant fourteen times. Steps five and six are triggered by an event rather than by the clock: a second item turning up, or a better item beating one already held. Step seven happens once, at the end, during the writing. The order is fixed, and the two set-up steps are the ones people skip. Skipping the set-up is exactly why the other five stop working.

1. Open the log before the first document is openedThe sheet exists before any evidence does. Seven headings across the top and nothing underneath them.
   Checking: is the log open and empty before the first document is?
2. Set the seven columns, left to right, onceNumber, what it is, where it came from, when it arrived, rank, corroborated, status. Set them at the start and never add an eighth mid-assignment.
   Checking: can every column of a row be filled without leaving the document on screen?
3. Write the row while the item is still on screenUnder thirty seconds, in the words that would be used out loud. Each new row goes underneath the last one, never above it and never in a gap.
   Checking: is the item still open on screen as the row is typed?
4. Write the rank into the same row before moving onWhatever hierarchy the desk or the teacher works to, the label goes into the rank column inside the same thirty seconds as the rest of the row.
   Checking: does every row that exists carry something in the rank column?
5. Mark the second item against the first when one turns upThe mark names the number of the item doing the agreeing. Where nothing agrees yet, the cell says that in as many letters.
   Checking: does every row carry either a corroborating item number or a visible mark saying there is none?
6. Mark a beaten row superseded and leave it exactly where it isIts number is never reused and its content is never edited. Only the status cell changes, and it changes to the date and the number of the item that beat it.
   Checking: is the older figure still readable in the log after the better one arrived?
7. Put the row number into the sentence as it is writtenEvery sentence resting on evidence carries the number of its row. Then read the finished note once, looking only for sentences that carry none.
   Checking: does every sentence lead back to a row, and does every row that carries the conclusion lead to a sentence?

Nothing in the seven steps requires permission, a team, a system or seniority, so a person in their first month can run one from their second day. A shared sheet works. A single file on one laptop works. A ruled notebook with seven columns pencilled across two facing sheets works, and has the advantage that nobody can sort it.

When is the log opened, and what is the first thing written in it?

The log is opened before the first document, and the first thing written in it is the seven headings. Writing the seven headings is the entire step, and it takes about a minute. Sharada Iyer opened the sheet on 3 March at nine in the morning and pulled the first filing at twenty past, so the log existed for twenty minutes before it had anything in it. Nobody stops to build a table once a document is already open and interesting. Opening the log first is not tidiness, it is the only way the first item ever gets recorded.

The pull of the interesting document is the thing to watch for. The filing loads, there is an unexpected number on the third printed sheet, and forty minutes disappear. Forty minutes on an unexpected number is good work. Those forty minutes are also how an assignment reaches its fourth item before its first row, and four items behind means writing from memory rather than from the screen.

What columns does a source log need?

Seven, and the reason there are exactly seven is that each one answers a question somebody will actually ask, and nothing else does. Number, what it is, where it came from, when it arrived, rank, corroborated, status. The seven are set left to right at the start and then left alone. An eighth column added in the middle of an assignment means every row written before it is now incomplete, and incomplete rows are the ones that get quietly abandoned.

Two of those columns look like clerical detail and are not. The rank columnThe cell holding where an item sits in the hierarchy a desk has fixed in advance, so a clash between two items is settled without an argument. is what lets somebody settle a clash later without reopening the argument, and where that hierarchy comes from is set out under source hierarchy. The status column is the difference between a log that records what happened and a log that records what survived. The seven columns are chosen so that a stranger holding the log can find every item, weigh it, and tell which ones were eventually no longer relied on.

The columns that are missing matter as much. No column for whether the item was useful, no column for a view of it, no column for what it means. Judgements of that kind belong to the writing, not to the record. A log with an opinion column becomes an argument, and what is wanted on a bad Thursday is a record that is not arguing with anybody.

When is each row written?

The row is written on arrivalAt the moment the item arrives, while the document is still open, rather than at any point afterwards., with the document still open on screen, and it takes under thirty seconds. Not at the end of the day, not at the end of the assignment, and not in a spare moment. On a nine day assignment no spare moment arrives before the deadline does. Two of the seven cells cannot be recovered afterwards at all: the route the item came by, and the date it arrived. The figure itself is still sitting in the document next week. How it arrived and when is gone within days, and it is gone without feeling gone. A loss nobody notices is the dangerous kind.

Consider the last time a colleague sent something useful. The number itself is probably still findable. Whether it came by email, on a call, or in a shared folder somebody pointed to, and which day it came, usually is not. The route and the date felt too obvious to write down at the time, and feeling too obvious to write down is precisely the property of a fact that will not survive the week.

Why is the rank written on arrival rather than at the end?

Because by the end the answer is known, and the answer changes what a source looks like. Ranking an item at the moment it arrives is a judgement about the item: who produced it, how it arrived, whether it was filed or said. Ranking it on the last afternoon is a judgement made by somebody who already knows which items agreed. Nobody promotes a weak source on purpose, and that is exactly why the drift is invisible: the items that supported the conclusion simply start to feel more solid than they did on Tuesday.

The effect is familiar from ordinary life. Somebody who has just bought a car, asked which reviews they trust, will call the reviews that praised the car the careful ones and the reviews that criticised it the work of people who did not understand the car. The car owner is not being dishonest. Judgement of source quality simply moves once an outcome exists. On the assignment here, the rank on every one of the fourteen rows was written inside the same thirty seconds as the rest of the row, so no rank on this log was ever set by somebody who knew how the work ended.

How is corroboration recorded?

With a number, not a tick. The corroboration markThe entry in the log recording that a second, independent item agreed, written as that item's number rather than as a tick. in row 7 reads item 10, and the mark in row 10 reads item 7, so the agreement is recorded from both ends and a reader can walk it in either direction. A tick tells a reader that somebody once felt satisfied. A number tells them what to go and open. The cell is filled at the moment the second item arrives, so a row is revisited exactly once and never re-read wholesale.

Where nothing has agreed yet, the cell is not left empty. The cell says so, in words, and on this log it said none found. The distinction sounds small, and it decides how the log reads six months later. A blank cell has two meanings that cannot be told apart: nothing agreed, or nobody looked. One of those is a finding about the evidence and the other is a finding about the person who kept the log, and a reader who cannot tell which is being told will assume the worse one.

What happens to an item that turns out to be wrong?

The item stays in the log and its status cell changes. Nothing is ever deleted, nothing is edited, and the number it was given is never handed to anything else. An item marked supersededBeaten by a better item, and kept in the log with a note of the date and of which item beat it. has not left the record; it has been retired inside it, with a date and with the number of the item that beat it.

Row 8 on this assignment is the whole lesson. On 4 March at half past two the relationship manager read out a revenue figure of Rs 20,90,00,000 for the year just reported and forwarded the spreadsheet extract it came from that afternoon. Row 1, written the previous morning, already carried Rs 21,20,00,000 from the filed accounts. The two disagree by Rs 30,00,000. Which of them wins is settled by rank, and the hierarchy the desk fixed in advance is set out under source hierarchy. The losing row was logged on arrival like every other item, and five minutes later its status cell read superseded 4 Mar, by item 1.

Consider what the alternative would have produced. Deleting row 8 leaves the log carrying one revenue figure. A single figure is a true statement about the conclusion and a false statement about the work. Editing row 8 to say Rs 21,20,00,000 manufactures a row claiming an extract carried a figure it never carried, and that is worse than deleting it. The only move that leaves the record honest is the one that costs nothing: change one cell, keep the rest.

How does the log connect to the finished writing?

Through the item numberThe short label, usually a plain number, that gets a reader from one sentence back to the row standing behind it., typed into the sentence at the moment the sentence is written. Not attached afterwards, for the same reason the row itself is not written afterwards: matching forty sentences to fourteen rows on the last afternoon is an hour of work that produces guesses. Typing a number as the sentence is written costs nothing. The number is the thread, and a thread only exists if it was tied at both ends.

Then comes the one sweep that makes the whole apparatus pay. Read the finished note once, looking for nothing except sentences that carry a figure or a fact and no number. On this note there was exactly one. Tracing it landed on row 14, the remark from the borrower's finance head on 10 March, whose corroborated cell read none found. The sentence was not deleted and the remark was not suppressed. The number went in, and the sentence now says plainly what it rests on.

What did the finished log look like on one assignment?

Fourteen rows, numbered 1 to 14 in the order the items arrived, written between 3 March and 10 March. Six of them were statutory or regulatory, four were market and industry items, two were published works and two were conversations. Eleven carried the number of an independent item that agreed and three carried none found. One row, number 8, was retired inside the log rather than removed from it. Four of the fourteen rows carry the conclusion, and all four of those are statutory. The log makes that visible without anybody writing a sentence about it.

The whole log follows. The number column, read downwards, gives the order in which things were learned. The corroborated column gives how solid the evidence is. The status column gives what stopped being used.

The four shaded rows at the top are the ones the conclusion actually stands on, and they are rows 1, 2, 3 and 4, all of them filed and all of them written inside the first seven minutes of the assignment. The pattern is not a coincidence. Opening the filing first is what produces it. The other ten rows are not padding. Without those ten nobody could say the four load-bearingHolding the answer up, so that the answer itself moves if the item turns out to be wrong. rows were checked rather than merely trusted.

How long does keeping a source log actually take?

Under fifteen minutes for the whole assignment. Fourteen rows at under thirty seconds each is under seven minutes of typing. The eleven corroboration marks went in as the agreeing items arrived, a few seconds each, and the single status change on row 8 took about as long as reading this sentence. The last read of the finished note, looking only for sentences carrying no number, took the rest. Set against nine hours of production, the entire record cost under three per cent of the time, less than the assignment lost to one interruption.

Most people guess an hour, and the guess is worth understanding because it is what postpones the log in the first place. An hour is roughly right for the other version, the one where somebody sits down on the last afternoon with a folder of documents and tries to work out what came from where. The last-afternoon version genuinely does cost an hour. So the choice is never between fifteen minutes and nothing. The choice is between fifteen minutes spread across nine days, and an hour on the one afternoon when there is least of it to spare, in exchange for a worse record.

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