Decisions, Not Feeds: Building a Personal Information Diet

An information diet is a construction job. Write down the decisions you owe in the next ninety days. Then admit only the inputs that can move one of them, and cancel everything else.

The admission test is narrow. If no realistic version of a source would make you act differently this quarter, that source has a coefficient of zero on your company. Reading it is a hobby, and hobbies belong in the evening.

Cutting volume is the wrong first move. Volume is the symptom. Composition is the problem, and composition is something you build.

Why comparative statics, cascades and aggregation

This piece runs the Wire Model: score the features of the decision, route to a small ensemble of formal models, force the ensemble to produce dated actions. The scores that carried the routing:

• Cognitive and affective distortion (0.8). Extra input reliably raises confidence. Whether it raises accuracy is a separate, measurable question with a discouraging answer.
• Deep uncertainty (0.75). You cannot know in advance which input will turn out to matter, so the diet has to be built on the decisions rather than on the topics.
• Contagion and spread (0.7). Almost nothing reaches you because it is true. It reaches you because it travelled.
• Regime-break risk (0.7). Producing plausible written content now costs close to nothing, which removes the last cheap quality filter volume ever provided.
• Exposure topology (0.65). Who feeds whom decides what looks like consensus.

Three models, one per job. Comparative statics answers what to admit. Informational cascades answers how much of what you already read is genuinely separate evidence. Crowd aggregation answers how to combine what survives. They sit in three different outcome families, equilibrium, complex and random, so their errors point in different directions.

The method usually adds a behavioural layer and a calibration layer as separate members. Both are folded here. The confidence-accuracy gap is the entire content of the first model, so a behavioural card would restate it. Calibration lives inside the third model, where scoring your own sources is the mechanic rather than a footnote. Neither fold gives up a lever.

The framework: your input stream is a portfolio priced in decisions

1. The derivative test: an input earns a slot only when it can move an action

Comparative statics does one thing. Hold everything else still, move one variable, watch where the equilibrium goes. Applied to your reading list, the question becomes brutally short: move this input across its realistic range, and does the action you take change?

The evidence on what extra input actually buys is old and it has held up. Thirty-two judges, eight of them clinical psychologists, worked through a case in four stages, answering the same 25-item test after each stage as more background arrived. Accuracy went 26.0%, 23.0%, 28.4%, 27.8%, against 20% for guessing. Confidence went 33.2%, 39.2%, 46.0%, 52.8%.¹ Information bought certainty at a steady rate and bought accuracy at none.

Three later studies found the same shape in modern conditions: as judges receive more information, confidence rises faster than accuracy, opening a gap the judges do not correct for.²

Now the sharper version. Forty-three loan officers were asked which 30 of 60 firms would go bankrupt within three years. They requested five financial ratios to work from and averaged 44.4 correct calls, 74%. The ratio of assets to liabilities, used alone, got 48 right, 80%. Across five separate judgment studies, ten thousand linear models per study with weights drawn at random, correct only in sign, performed about as well as models built from the judges themselves. Equal weighting did better than both.³

The instruction inside that result sits badly with anyone who reads for a living. Most of the work is done by a small number of variables pointed the right way. The rest of the input stream is decoration you experience as diligence.

A source you cannot attach to a named decision is not information. It is weather.

So write the decision ledger. One page. Five to nine live decisions, each with a date and the two or three parameters it actually turns on. “Do I extend 30-day terms to this distributor” turns on their last twelve settlement cycles, your current exposure to them, and whether the buyer behind their LPO has paid you before. It does not turn on a funding announcement in Lagos, a thread about pricing psychology, or a conference agenda.

Comparative statics, the admission lens

Assumes: you can name the decision and the few parameters it turns on.

Fits because: cognitive distortion scored 0.8, and the confidence-accuracy gap is measured, not asserted.

Breaks when: the decision itself is missing from your ledger. A zero derivative on an unnamed decision is not evidence of irrelevance.

Counteracts: reading as a substitute for deciding.

May reinforce: narrowness, and a ledger frozen around last quarter’s problems.

2. The independence count: a feed reports mentions, you need observations

Observational learning explains why crowds converge on the same behaviour and why that behaviour is prone to fads and to error. Once enough people have acted, the rational move for the next person is to copy the sequence and suppress their own private signal, which means their action transmits no new information to anyone behind them.⁴ The cascade grows while the evidence stays where it was.

Someone ran the experiment properly. In an artificial music market, 14,341 participants downloaded unknown songs, some seeing what others had downloaded and some not. Social influence raised both the inequality and the unpredictability of which songs won.⁵ Same songs, same quality, wildly different winners depending on which copy of the world you landed in.

Velocity is worse than neutral. Across roughly 126,000 verified true and false stories spread by about three million people on Twitter between 2006 and 2017, truth took about six times as long as falsehood to reach 1,500 people, and false stories were about 70% more likely to be retweeted.⁶ Speed in a feed measures novelty and shareability. It carries no information about whether the thing happened.

The operator move is the independence count. Before you act on a claim, count the distinct primary observations underneath it, not the number of times you have seen it. Thirty forwards of the same voice note in a distributor WhatsApp group is one observation from one person at one depot. In Kenya, 51% of the online news users surveyed for the Digital News Report say they share news through social, messaging or email.¹¹ That is a re-transmission layer, and it is where most founders now do their reading.

Count observations, not mentions. Most of the consensus in your feed is one observation, forwarded.

Practically: for any claim that would change a decision, find the primary artefact. The gazette notice, not the summary of it. The circular on the regulator’s own site, not the screenshot. The signed LPO, not the buyer’s assurance that it is coming. If the primary artefact does not exist, you have found a rumour with good distribution.

Informational cascades, the independence lens

Assumes: people act on what they observe others doing and stop reporting their own signal.

Fits because: contagion scored 0.7 and exposure topology 0.65.

Breaks when: sources are genuinely independent and publish their reasoning, which makes the mention count informative again.

Counteracts: treating repetition as corroboration.

May reinforce: contrarianism for its own sake, and dismissing a true consensus because it is popular.

3. The error-diversity rule: build the ensemble, not the best single source

The aggregation result is the one that makes a diet designable. Collective error equals average individual error minus the diversity of the individual estimates. Diversity is subtracted, which means it does real arithmetic work. In the formal version, groups of randomly selected agents outperformed groups made of the individually best-performing agents.⁷

So the way to improve your read is to add sources whose mistakes are uncorrelated with the ones you already carry. Adding a fifth analyst who reads the same three primary sources as your other four adds cost and no accuracy.

Exposure destroys the property that was doing the work. In a controlled estimation experiment, 144 subjects answered six factual questions five times each, some seeing what others had estimated. Social influence narrowed the diversity of opinions without improving collective error, pushed the truth toward the edges of the range where an outside observer would no longer find it, and raised individual confidence anyway.⁸ A group that reads each other converges, feels better, and gets no closer.

A group that reads each other converges, feels better, and gets no closer. That is your feed, described precisely.

The discipline is trainable, which is the useful part. In a two-year geopolitical forecasting tournament, probability training, working in teams, and promoting the top performers into elite teams all raised accuracy. The training pushed forecasters toward reference classes and toward averaging multiple estimates.⁹ Structure beat talent again.

So score your list on error correlation rather than on reputation. Five newsletters reading the same wire are one source. A churned customer, an ops manager at your competitor’s distributor, a field agent in a county you do not visit, and the regulator’s own published register are four. Nobody in that second list is impressive. Their errors have nothing to do with each other, and that is the whole point.

Crowd aggregation, the combination lens

Assumes: estimates are combined, and their errors are at least partly independent.

Fits because: deep uncertainty scored 0.75, so no single source can be trusted to carry the read.

Breaks when: your sources share an upstream, which collapses diversity to zero while the source count still looks healthy.

Counteracts: guru dependence and prestige-weighted reading.

May reinforce: false balance, and paying for noise on the theory that it is diversity.

GEER: what to pull, in cost order

Six channels carry the exposure: your decision inventory, what you admit, how independent it is, how correlated its errors are, how fast you can retrieve an answer when the decision is live, and what interruption costs you. Pull the cheap and reversible ones first.

1. Write the decision ledger. One page, five to nine decisions, each with a date and its two or three parameters. One hour, and it reprices everything below it.
2. Kill push. Leave notifications on for customers, payments and three named people. Interrupted work gets finished faster and the cost is paid in stress, time pressure, frustration and effort.¹⁰ Twenty minutes.
3. Build the instrument panel. Your own numbers, refreshed weekly, in one place: settlement file, collections ageing, agent or account churn, gross margin by SKU, cash days remaining. Half a day, and it answers more ledger questions than anything you subscribe to.
4. Convert subscriptions into questions. Replace “read this daily” with “answer this question every Monday”. Pull is indexed to the ledger. Push is indexed to someone else’s revenue.
5. Audit independence on your top ten sources. Mark which ones read each other. Cancel the duplicates. An afternoon.
6. Recruit three uncorrelated inputs. One churned-customer call a week, one field visit a month, one primary document channel such as the regulator’s register or the tender portal. Weeks, and it is the highest-yield lever on the list.
7. Buy latency where a decision waits on it. Paid data that answers a ledger question the same day beats free data that answers nothing.

No-lever flag: an empty decision ledger means you do not have an information problem. You have an unnamed strategy, and no filter fixes that. Write the ledger before you touch the stream.

RADAR: a dated portfolio for your inputs

DO NOW, by T+3 days. Reversible, and worth doing under every scenario.

1. Write the ledger. Date every decision.
2. Cull notifications to customers, payments and three named people.
3. Stand up the instrument panel, even as a spreadsheet with five numbers.
4. Cancel every subscription that failed the derivative test. Do it in one sitting so it cannot be negotiated.

HEDGE, by T+14. Cheap insurance against a ledger that is missing something.

1. Add one human source whose errors have no shared upstream with yours.
2. Book a standing thirty-minute weekly review on one question: what changed that moves a dated decision.
3. Write down your current answer to your two largest open questions, with a probability and a date. It costs nothing now and it is the only way you will ever score a source, including yourself.

DEFER AND TRIGGER. Irreversible or expensive, so wait, and set the observable trigger today.

1. Defer: an expensive data subscription, a hired analyst, a rebuilt reporting stack.
2. Trigger to spend: two months in which a dated decision slipped because the answer was not available in time. That is a latency problem and money fixes latency.
3. Counter-trigger: by T+28 the panel has answered every ledger question on schedule. Cut the budget instead of raising it, and put the hours into the uncorrelated sources.

Sitting on the other side of the stream. If you are the one writing the investor update, the distributor note or the board memo, run the same rule outward. Lead with the number that changes their next action, state what it changes, and stop. Anything they cannot act on is you spending their attention budget on your own comfort.

CHAIN: the base rate, and what moves downstream

The comparison set here is any system where someone must act on a deadline while a channel supplies more input than they can weigh. Clinical case judgment qualifies. Bank underwriting qualifies. Two-year forecasting tournaments qualify. Across all three, the pattern repeats with unusual consistency: accuracy plateaus early, confidence keeps climbing, and a written structure beats the well-read individual.^{1, 3, 9}

Present-state modifiers push in one direction. Generated text has driven the marginal cost of plausible input to roughly zero, and retrieval is migrating from feeds into assistants, which moves the filter upstream and out of sight.

Strip out what would have happened anyway before you credit the diet. A large share of the improvement founders report after cleaning up their inputs comes from the act of writing the decisions down, not from the reading that followed. Give the ledger its credit and the reading list only what is left.

Second order: a founder working from a ledger asks narrower questions, buys fewer inputs, and becomes a worse audience for the ecosystem. Third order: their peer group’s consensus stops being available as a crutch, so their errors decorrelate from the cohort’s. That decorrelation is the return. It is also why the discipline feels lonely for the first month.

Matrix-break flag. Two rules the models assume are moving. Volume once carried weak evidence about effort, and it no longer does, which leaves the independence count as the only filter that still works. And if a single assistant becomes your main retrieval path, your error diversity collapses to that one system’s, silently, while your source count looks unchanged. Keep one structurally different channel alive: a human who tells you unwelcome things, or a primary register you read yourself.

What this ensemble cannot see

Three gaps, and a better filter closes none of them.

The decision you have not named. The ledger prices what is on it. The input that changes your company most is usually about something not yet written down, which is exactly the input the derivative test discards.

Whether a source is skilled or lucky. A writer with four good calls has a sample of four. In a domain this noisy, a track record that short carries almost no information, and treating it as a signal is the failure mode of everyone who takes information diets seriously.

Your own state. None of these models knows what sleep, cash pressure or a bad week does to your judgment. Every one of them assumes a stable reader.

Here is the action that survives all three. Keep the ledger to one page and rewrite it on the first Monday of every month, from scratch, so a decision has to earn its place again. At each rewrite, add one input whose errors have nothing in common with anything already on the list, and retire one that has not changed an action in ninety days. That single monthly hour is the whole diet. Book the first one before you close this page.

Sources and notes

1. Oskamp, S. “Overconfidence in Case-Study Judgments.” Journal of Consulting Psychology 29(3), 1965, 261-265. Thirty-two judges, 25-item case test, four cumulative information stages. Accuracy and confidence figures are from Table 2. Full text.
2. Tsai, C. I., Klayman, J., and Hastie, R. “Effects of Amount of Information on Judgment Accuracy and Confidence.” Organizational Behavior and Human Decision Processes 107(2), 2008, 97-105. The publisher page bot-blocks a browser user agent, so the abstract carrying the confidence-accuracy claim is linked at the RePEc record.
3. Dawes, R. M. “The Robust Beauty of Improper Linear Models in Decision Making.” American Psychologist 34(7), 1979, 571-582. Loan officer figures are Dawes’s account of Libby (1976); the random-weight and equal-weight results are from Dawes and Corrigan (1974), summarised in his Table 1. Full text.
4. Bikhchandani, S., Hirshleifer, D., and Welch, I. “Learning from the Behavior of Others: Conformity, Fads, and Informational Cascades.” Journal of Economic Perspectives 12(3), 1998, 151-170. Record and abstract.
5. Salganik, M. J., Dodds, P. S., and Watts, D. J. “Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market.” Science 311(5762), 2006, 854-856. 14,341 participants. Author full text and supporting material.
6. Vosoughi, S., Roy, D., and Aral, S. “The Spread of True and False News Online.” Science 359(6380), 2018, 1146-1151. Approximately 126,000 cascades, three million people, 4.5 million tweets, 2006 to 2017. University of Vermont mirror.
7. Hong, L., and Page, S. E. “Groups of Diverse Problem Solvers Can Outperform Groups of High-Ability Problem Solvers.” PNAS 101(46), 2004, 16385-16389. The PNAS site returns 403 to curl; open text at PubMed Central.
8. Lorenz, J., Rauhut, H., Schweitzer, F., and Helbing, D. “How Social Influence Can Undermine the Wisdom of Crowd Effect.” PNAS 108(22), 2011, 9020-9025. N = 144, six factual questions, five consecutive estimation rounds. Open text at PubMed Central.
9. Mellers, B., Ungar, L., Baron, J., Ramos, J., Gurcay, B., Fincher, K., Scott, S. E., Moore, D., Atanasov, P., Swift, S. A., Murray, T., Stone, E., and Tetlock, P. E. “Psychological Strategies for Winning a Geopolitical Forecasting Tournament.” Psychological Science 25(5), 2014, 1106-1115. Author-hosted full text.
10. Mark, G., Gudith, D., and Klocke, U. “The Cost of Interrupted Work: More Speed and Stress.” CHI 2008, 107-110. Author full text.
11. Reuters Institute for the Study of Journalism, Digital News Report 2025, Kenya country page. Sharing figure of 51% is for a survey sample of mainly English-speaking online news users and is not nationally representative, as the report states. Country page.