---
title: "Rates and Forecasts"
url: "https://toddpaulbrownjr.com/corrigibility/rates-and-forecasts/"
author: "Todd Paul Brown Jr."
description: "Swap \"will it happen?\" for \"which rate is faster?\", state the method's bounds, pin forecasts with a six-slot register, and keep the framework honest."
kind: "guide-chapter"
updated: "2026-09-26T02:50:19+00:00"
---

# Rates and Forecasts

A regional newspaper chain is losing subscribers, laying off reporters, and folding three newsrooms into one. Ask ten people who follow the industry whether it survives the decade, and you get ten confident answers split roughly down the middle. The optimists point to a loyal core readership, a new digital subscription push, a masthead people still trust. The pessimists point to the layoffs, the empty seats at city council meetings nobody covers anymore, the investment firm that just bought a stake. Both sides are certain. Both are reasoning from real evidence. And both are answering a question that can't be answered the way they're answering it.

Neither side is wrong about its facts. The loyal readership is real. So are the empty seats. What's missing is a way to combine the facts into a forecast instead of a mood. The optimist is describing a force that's still building something. The pessimist is describing a force that's tearing something down. Neither has said which force is currently winning, because neither asked the question that way.

"Will this institution collapse or muddle through?" sounds like a factual question with a yes-or-no answer waiting in the future. It isn't. It's a request for a single verdict about a process still running, made from a snapshot taken inside it. You can't read a verdict off a snapshot. What you can read, if you know where to look, is which of two things is currently moving faster.

That's the move this chapter is built around. Stop asking whether something will happen. Start asking which of two rates is winning.

## The rate-comparison method

Take a fishery. A coastal town pulls a certain tonnage of fish out of the water every season: call it the catch rate. The fish population replaces itself at a certain rate through spawning, growth, and survival to maturity: call it the spawn rate. Nobody needs a prophecy to forecast this fishery. They need two numbers and a comparison. If catch exceeds spawn, the stock declines season over season until it collapses or the town stops fishing. If spawn exceeds catch, the stock holds or grows. Most real cases sit close enough to even that you have to measure. But the question has changed from a guess about a distant future into an observable contest happening now.

Take a company that is burning trust. Every quarter it does something that damages its relationship with customers: a price hike buried in fine print, a support line that takes three weeks to answer, a promise walked back. Call that the damage rate. Every quarter it also does something that rebuilds the relationship: a genuine fix, an apology that changes a policy, a product that works better than expected. Call that the repair rate. A company can survive a lot of damage if repair keeps pace. It can look fine right up until repair falls behind, and then it doesn't look fine for long. The question isn't "is this company in trouble?" It's "which rate is winning, and what's the trend on each?"

Take a captured institution, one whose correction machinery has been narrowed, redirected, or turned against the people trying to use it. Here the race is between the rate at which internal correction is suppressed and the rate at which reality forces correction anyway, from outside. An oversight body that keeps quietly closing the complaints it was set up to investigate is winning the suppression race, for now. When something it waved through fails in public, the second race catches up with the first, all at once.

Three cases, three kinds of machinery: a biological population, a customer relationship, an institutional feedback loop. What they share is a shape: two processes running against each other, one building and one consuming, with an outcome that depends on which is faster now, not on some fixed destiny. Name the shared shape without confusing it with the mechanics underneath. Fish know nothing about trust, and no customer has ever spawned. It's the same structural pattern in different machinery, and once you've seen it in a fishery you stop needing a fishery to see it.

Why reframe the question at all, instead of picking a side and arguing harder? Three practical reasons. First, "which rate is faster" can be answered with things you can observe: trend lines, not vibes. You can watch subscriber counts and reporter headcounts move over eighteen months; you can't watch a newspaper's fate. Second, the reframe forces you to name both forces. A common way forecasts go wrong is looking at only one side of the race: all doom and no repair, or all confidence and no damage. Naming both is half the discipline. Third, the reframe tells you when you're bluffing. If you try to fill in the catch rate and spawn rate and can't get a number for either, you've learned something true: you don't have a forecast. You have a hunch wearing a forecast's clothes. That isn't the method failing. That's the method working.

## A fence for the persistence argument

A related claim needs its own fence before you rely on it: a steady, standing pull wins over the long run against resistance that isn't adapting, even when any single week looks like noise.

Think of a river cutting a bank. On any given afternoon, nothing visible happens. Ask "will the river erode this bank by Thursday?" and the honest answer is: who knows. But the river pushes the same way every day, the bank doesn't move out of the way, and over ten years the outcome isn't in doubt. You didn't need to predict any Thursday. You needed to know the direction was fixed and the resistance was static, and then let time work. The technical version: you're not forecasting an event, you're forecasting an integral — the sum of a steady push over a long stretch, not any one moment inside it. That's why you can be confident about the ten-year trend of something whose week-to-week behavior is close to random.

This is a real and useful argument, and exactly the kind that gets abused once nobody states its limits. So state them. The persistence argument holds only when two things are roughly true: the target of the push is fixed, and the thing resisting is static — not learning, adapting, or reorganizing under pressure. Change either condition and the argument stops being safe to rely on. It degrades as the resistance adapts, and you won't always see how fast.

Go back to the fishery. If fishing pressure stays constant and the fish population is a static resistance, forecast the integral: sustained overfishing collapses the stock, and you don't need to predict any single season. But fish populations aren't always static. In the years before the northern cod fishery off Newfoundland and Labrador collapsed in the late 1980s and early 1990s, researchers later found the fish were maturing at steadily younger ages and smaller sizes, a shift they judged consistent with fishing-driven evolution. The population was changing its own resistance in response to what was being done to it. When that happens, the old integral is wrong, because the thing being pushed on has changed what it is. The same goes for the captured institution. Steady suppression, against employees, whistleblowers, and journalists who never change tactics, would win in the end. But institutions under sustained pressure sometimes provoke the very coalition they were suppressing: a new overseer is appointed, a union forms, a competitor exploits the opening. The opposition adapted, and the fence closed on the argument.

Naming the bound isn't a hedge tacked on to sound careful. It's the difference between a forecasting tool and a rhetorical trick that sounds like one. "This will happen because the trend is inexorable" is a real claim only if you can say what would make the target shift or the resistance adapt, and then check whether either has.

## Direction without a director

A worry shows up the moment you point at a revealed direction inside an institution, a market, or a culture and call it a *telon*: the direction a system's actual dynamics pull toward, whatever it says its goals are. The worry goes: if there's a clear direction, doesn't someone have to be steering? Isn't "the institution's real telon is X" a polite way of alleging a conspiracy?

No, and walking through why removes a whole category of bad argument.

Traffic makes the point cleanly. In 2008 a team of Japanese physicists put 22 cars on a circular track about 230 meters around and asked the drivers to cruise at a steady 30 kilometers per hour while keeping a safe distance. There was no bottleneck: no merge, no lane closure, no accident. At first, traffic flowed smoothly. Then small, ordinary differences in how each driver held speed began to amplify, and a jam formed and rolled backward around the track like a wave. Nobody caused it. Every driver was doing something sensible about the car in front. The jam was still real: it had a shape, a speed, a direction of travel.

That's a telon with no purposer behind it. The direction is real — you can point to it, measure it, forecast its rate — and entirely undesigned. Markets show the same thing. No one is in charge of a price, yet prices tend to move toward the level where buyers and sellers match, because each trader is independently doing the locally sensible thing with the information in front of them. In both cases the aggregate pattern is steadier and more predictable than any single participant's behavior, which is the reverse of what the conspiracy framing expects. A conspiracy predicts that the plan is the stable thing and individual actions are its obedient shadow. An emergent direction predicts the opposite: the individual actions are the stable thing, each locally sensible and reliably repeated, and the aggregate pattern is what falls out when you add enough of them together.

This matters for forecasting because it tells you which claims you're allowed to make. You can say: this institution's revealed direction is toward optimizing its proxies, because each actor inside it — the manager protecting this quarter's numbers, the employee protecting this year's review, the department protecting its budget — is rationally responding to the incentives in front of them, and the sum of those choices is an institution drifting somewhere none of them chose. That claim needs no boardroom, no memo, no architect. What you cannot say, without separate and much stronger evidence, is that someone designed it that way. The rate-comparison method doesn't need a designer. It needs incentives and a place for them to add up. That's a more accurate account of how most institutional drift happens, and a much less paranoid one.

## What this instrument reads well, and what it doesn't

Before handing over the sharpest version of the tool, be honest about what it's rated for, the way you'd read the operating range on any instrument before trusting a number off it.

This kind of analysis is good at structure and failure modes. It can say, with real confidence, that the coupling between a system and what it depends on is severing, and that the severing, left running, will force a crisis. It can say what shape that crisis will likely take: a fishery collapse, a customer exodus, an oversight failure visible only after something breaks. It is not good at contingent, sequential detail: which quarter, which triggering event, which particular decision lights the fuse. "This coupling is severing and will force a crisis" is a claim the method earns. "It happens in March, triggered by the compliance officer's resignation" is not. That rests on facts about individual people and timing that a structural method was never built to see.

State that as the tool's rated range, not as an apology. A thermometer is a fine instrument for telling you water is heating toward a boil, and a poor one for telling you which bubble forms first. Nobody apologizes for that. The honest version of this instrument says the same about itself. Many bad calls made with frameworks like this one don't come from getting the structure wrong. They come from spending structural confidence on a detail it was never rated to support: taking "this will force a crisis" and quietly upgrading it, somewhere between the analysis and the headline, into "this will happen on this date." Keep the two apart and the tool stays honest. Blur them and it starts to sound like a fortune teller.

This cuts against a habit much forecasting commentary encourages. A forecaster who gets the big structural call right is often credited as though every detail had been right too. One who gets a single detail wrong is often dismissed as though the whole structural read had failed. Neither response tracks what happened. Grade the two claims separately, every time: was the structure right, and, separately, was the detail right?

The next few pages are for readers who want the machinery of how a forecast gets pinned down precisely enough to be wrong. If you'd rather keep moving, skip ahead to the drill near the end of the chapter.

## The prediction register

Here's a habit that is nearly universal and almost entirely useless: the retrospective "I called it." Someone points back at a remark from months or years ago and says, see, I predicted this. The trouble is that memory isn't a recording device. It edits. The vague hunch you actually had gets sharpened, after the fact, into a crisp prediction that fits what happened: the threshold moves, the timeline moves, the disconfirming case you'd have accepted moves, and nobody notices the edit. Usually that isn't dishonesty. It's what memory does under the pull of a story that already has an ending. The only defense is writing the prediction down, in detail, before you know the answer. That's what pre-registration is for in fields that take forecasting seriously, and this framework borrows the habit.

A usable forecast, one that someone who wasn't in the room can check later, fills six slots. Skip a slot and you don't have a bet. You have an impression.

1. *Mechanism or branch.* What process is doing the work: which rate is racing against which?
2. *Present-tense discriminator.* What can be observed right now, before the outcome is known, that separates this branch from the alternative?
3. *Metric, threshold, and window.* What number, crossing what line, by what date?
4. *Disconfirmer.* What outcome would prove this wrong? The strongest version names an outcome the framework forbids, not just one it didn't expect.
5. *Base rate.* What would a forecaster with no access to this framework have said? Score only the divergence from that baseline. Credit is for what the framework added, not for predicting the obvious.
6. *Confound list.* What else could produce the same result, for reasons that have nothing to do with this framework?

Here is the register worked through on an invented case with nothing riding on it. Take a mid-size professional association whose membership has been falling for three years while its leadership keeps raising dues.

- *Mechanism:* proxy displacement. Leadership is optimizing dues revenue and headline membership over the service the association exists to provide, and the correction channel, a member advisory board, is increasingly bypassed.
- *Present-tense discriminator:* the advisory board's public minutes show recommendations logged and not acted on, three cycles running. That is visible now, before any collapse.
- *Metric, threshold, window:* the renewal rate falls below 70 percent of the prior year's total within the next two renewal cycles.
- *Disconfirmer:* if the advisory board's authority is genuinely restored (recommendations acted on, not just logged) and the decline continues anyway, or if renewals hold without any such restoration, the claimed mechanism was wrong.
- *Base rate:* suppose associations of this type lose, say, 3 percent of members a year regardless of how they're governed. Only a decline steeper than that counts as evidence for this mechanism. (In a real forecast, you'd look this number up; here it's a placeholder to show where it goes.)
- *Confounds:* a recession hitting members' employers, a rival association launching in the same period, a pricing change unrelated to governance.

That's the shape. Six slots, each closing off a way for a prediction to quietly become unfalsifiable after the fact.

## The absorption reflex

There's a move to watch for in this kind of framework more than in most, because the framework's own architecture makes it tempting. A counterexample shows up, a case that seems to break one of the core distinctions. The system looks captured but isn't, or looks healthy but clearly isn't. The tempting response is to widen the concept a little, just enough that the awkward case slides inside after all. In the moment it feels like the theory getting stronger: look, it explains this too. That feeling is the trap. A concept that can be stretched to fit anything explains nothing, because it has stopped making any claim a real case could fail. This is how a framework goes unfalsifiable: not in one dramatic move, but by accretion, one reasonable-sounding widening at a time.

Two firewalls hold this off, and they apply to this book as rigorously as to anything it examines.

First, a genuine unification has to produce a new testable prediction, not just a new label for something already known. If stretching a concept to cover a fresh case doesn't generate any claim that wasn't available before — a forecast, a discriminator, anything that could turn out false — it isn't a unification. It's relabeling, and should be called that.

Second, every branch of the framework needs a discriminator that doesn't depend on the outcome being explained. If the only evidence for "this was really proxy displacement" is that things went badly, the diagnosis is doing no work. It's redescribing the bad outcome in the framework's vocabulary. A real discriminator has to be checkable before the outcome is known, the same requirement the register makes of slot two.

A third check is less formal but just as important: could a competent, honest person look at the same evidence and reach a different conclusion? If yes, then somewhere in the analysis a choice got made — a boundary drawn one way, a substrate named this way rather than that. That choice needs to be shown, not buried inside a conclusion dressed as an observation. Naming the selection is not a weakness in the argument. Failing to name it is.

This belongs in the forecasting chapter on purpose. Every other chapter asks you to diagnose something in front of you. This one asks you to say what hasn't happened yet, and the pull to stretch a favorite concept until it "predicts" whatever does happen is strongest where being wrong feels most costly. The chapter that teaches you to forecast is the chapter most likely to cheat, if nobody's watching for it.

A skeptical reader is right to ask whether a framework willing to see one pattern across a fishery, a company, and an institution could ever be wrong about anything. The honest answer is that broad structural claims are harder to falsify than narrow ones. That's a real limitation. What keeps them accountable is the machinery above: every branch carrying its own discriminator, every unification paying for itself with a new prediction.

## The drill

Put the chapter to work in four steps, on whatever standing question you actually have: an institution, a relationship, a company, a movement.

1. Name the two rates that are actually racing: what's building, what's consuming, and what shared substrate both act on.
2. Say which one is currently faster, and exactly what observable trend tells you so. Not a feeling: a number or a trajectory someone else could check.
3. Fill all six slots of the register for the one claim you'd actually be willing to stake something on, not the safest claim available.
4. Run the absorption check on your own answer. Could someone equally competent, looking at the same facts, have drawn the boundary differently and landed somewhere else? If so, say where you drew it, and why.

None of the four steps requires certainty about the future. All of them require honesty about the present.

## Close

The instruments of Part III are now built: the audit, the markers, and the register that keeps a forecast honest. What's left is taking them somewhere they'll cost something to use. Part IV applies them in four settings: AI systems, public institutions, corporations and the workplace, and finally the individual reader. In each, the question is the same one this chapter taught you to ask: which rate is winning, right now?

## Sources

- Esben M. Olsen et al., "Maturation trends indicative of rapid evolution preceded the collapse of northern cod," *Nature* 428: 932–935, 2004. https://www.nature.com/articles/nature02430
- Yuki Sugiyama, Minoru Fukui, Macoto Kikuchi, Katsuya Hasebe, Akihiro Nakayama, Katsuhiro Nishinari, Shin-ichi Tadaki, and Satoshi Yukawa, "Traffic jams without bottlenecks — experimental evidence for the physical mechanism of the formation of a jam," *New Journal of Physics* 10: 033001, March 2008. https://iopscience.iop.org/article/10.1088/1367-2630/10/3/033001
- Institute of Physics (news release via Phys.org), "An accident? Construction work? A bottleneck? No, just too much traffic," March 4, 2008. https://phys.org/news/2008-03-accident-bottleneck-traffic.html