Accuracy Before Speed

A familiar process already runs near the fastest pace the quality bar still allows. Real speed comes from changing the process, and from running the current process at a pace the quality survives. An efficient technique run accurately produces that pace as a by-product.

Core takeaways

• Familiar work already runs at roughly 80–90% of the fastest pace the quality bar allows. The unused band above that pace is maybe 5–10%, and taking it makes more mistakes.
• Repairing one error runs near four times the time of doing the step right once. The four costs are catching the error, restoring context, doing the step again, and unpicking whatever was stacked on top.
• Add a tenth of speed and a tenth of errors and the total cost is about four tenths extra overall.
• Switching to a fundamentally better technique can be worth several times. Hurrying the old technique never reaches that gain.
• A new technique is practised for accuracy alone. Speed that arrives without being forced is the mark that the method has settled.
• Trying hard all the time leaves no pause for noticing whether the method still aims at the goal.

What one error costs

The competing view is that effort converts to output the way input converts to product, so pushing harder produces more finished work. That view is what makes leftover minutes look like unused capacity. Work that fails the quality bar is not finished work. Pages turned that later have to be done again add nothing.

Familiar work already sits at roughly 80–90% of the fastest pace that still survives that bar. That band is a practitioner estimate, not a measured constant. There is usually an internal sense of where mistakes start to climb, though that sense is noisy. The unused band is maybe 5–10%. Taking it makes more mistakes.

Repair, in this system’s teaching picture, runs near four times the time of doing the step right once. That 4x is a heuristic, and the speaker who taught it disclaimed it as research. It is not a laboratory result. The four costs are catching the error, restoring context, doing the step again, and unpicking whatever was stacked on top. Errors cascade. Treat 1+1 as 3 and the next line becomes 1+1+2=5.

Foundation errors hide longest, for weeks or months. Then they show up everywhere and cannot be traced cleanly to their source. After a rush the person often feels more sure rather than less, because the knowledge that would catch the failure is the knowledge that was never built. Nothing can be reverted to the state before the error. Later material sits on the error, so diagnosis stops being which of four to six clean mistakes and becomes which of a dozen to thirty interacting ones. Rebuilding costs more than first-pass care and is more confusing, which is why people walk away from it. A mid-arc repair can take a season that a careful fortnight at the start would have spared.

In this arithmetic, add a tenth of speed and a tenth of errors and the cost is about four tenths extra overall. To break even, the rush would have to be four times the leftover that exists. A rush with zero errors still returns only that leftover band, and it is rarely worth the exposure. A stretch of rushing tends to finish with the same finished work as a stretch at a pace that can be held, and with more tiredness, which then makes more mistakes.

Where real speed comes from

Switching to a fundamentally better technique can be worth several times. No amount of hurrying the old technique reaches that gain. The comparison only holds at a comfortable pace: the same process, calm against rushed. Done correctly, the efficient process produces fewer errors to review, so the advantage it has by design is realised in full. The gain only disappears if the pace is slowed on purpose.

An efficient process, rushed, loses to the same process run calmly. Leftover minutes on the current method barely exist, and claiming them raises the error rate the quality bar was there to stop. The change with real range is a better process at a pace the quality survives. Marginal Gains is that change as a practice: one named change to the method, rather than a harder push on the old one.

Speed that arrives on its own

Hands get faster once the method is known. Speed that arrives without being forced is the mark that the method has settled, the readiness mark in Rapid Skill Acquisition. Careful work can feel slow in the sitting. It covers fewer pages, and more of what it covers is held. The fast and shallow path pays those hours back later as relearning and re-testing. Time efficiency is judged across the whole arc of mastery, not across the single sitting.

A new technique is practised for accuracy alone, and pace is left to arrive once the motions are right. Forcing speed while the method is still wrong makes the mistakes part of the habit. Relearning a mislearned skill then runs much slower, around three times slower, with the same heuristic status as the 4x. Each move has to be watched for the old version firing.

Time on the calendar is not practice

Elapsed time is not practice time. Two months in which the method was not actually run is the same as no time having passed. Two weeks of heavy, accurate practice can be enough to move on. The unit is what was practised, plus an honest check that the work now runs correctly without strain. How long the topic has been on the calendar is not the unit. Pacing Skill Development owns the same load logic for how many processes to run at once.

Trying hard all the time

Trying hard all the time leaves no pause for aiming the work. Someone fully occupied by the push has no mental space left to notice whether the method still aims at the goal.

Work that actually aims at the goal seldom has to feel like a grind. Productive is whatever moves the goal. A rest, a walk, or an hour with people is the productive move when focus or mood is what is binding. The same hour changes value as the state changes. Priorities get recalibrated weekly, and daily on rough days.

If a level can only be held by grinding, the method is what is wrong. Take the short-term cost and build the method that holds the level without the grind. Priority 0+1 System is the companion: ranking work only pays when each unit of work is done right.

Working with agents

Accepting agent work without checking saves minutes today and costs the fourfold repair when the fault appears later in the chain. This is the cleaner 4x, because later commits sit on the bad one, so the cascade is steeper. Throughput lives in the harness, the prompts, checks, and reusable patterns around the agent. It does not live in running more jobs. Watching more jobs per day is the same push applied to a new kind of work. It spends the exact hours that improving the prompts and checks would have used.

The two mistakes are not equal. Going too slow takes deliberate effort. Going too fast happens on its own. If the right level is unclear, stay on the current level another week or two, and actually practise it. Sitting on a calendar without running the method is not holding the level.

What is known and what is not

The speed–accuracy tradeoff says that faster work produces more errors. That direction holds. Late-found defects cost more than early ones, and schedule pressure raises defect count. Those are different 4xes from the claim that one study error costs four times the time of doing it right. The study-repair 4x remains this system’s teaching number. The 80–90 / 5–10 / several-times figures are the same family. The checkable expectation is the unforced speed-up arriving, or a first rush that already produced work that had to be done again.

Leftover ambition goes into the process, or into the harness. It does not go into pushing the old process harder. What replaces the push is a better method at a calm pace.

How to practice this

1. Run a familiar method at the fastest pace the quality bar still allows. Notice where mistakes start to climb. That point is the top of the band.
2. When you want more speed, make one named change to the method. Compare the new method calm against the old method calm. Notice whether the change is worth several times.
3. Practise a new technique for accuracy alone. Leave the pace alone. Notice when speed starts arriving without being forced.
4. Before moving on from a level, check that the work runs correctly without strain. Count what you actually practised, not the weeks on the calendar.
5. When the work feels like a grind, stop and check whether the method still aims at the goal. Recalibrate priorities weekly, and daily on rough days.
6. Check the agent’s work before starting another job. Put the hours into the prompts, checks, and reusable patterns instead. Notice whether later work is sitting on an unchecked result.

Related pages

• Priority 0+1 System: ranking work only pays when each unit of work is done right.
• Rapid Skill Acquisition: unforced speed as the mark that a method has settled.
• Pacing Skill Development: same load logic for how many processes to run at once. Elapsed time is not the unit.
• Marginal Gains: process change as the one speed lever with range.

Open questions

When is a 5–10% rush actually worth the exposure?

How is habituation-speed told from rushed-speed from inside the session?

Sources

• Fitts 1954, speed–accuracy tradeoff: faster responses produce more errors. Support for the direction of C8, not for a study-repair multiplier.
• Wickens, Engineering Psychology and Human Performance; Pachella 1974: same SAT family. Bound: the finding is error rate under speed, not repair-time arithmetic.
• McConnell, “Software Quality at Top Speed”: excessive schedule pressure associated with up to 4× defects. Different claim: defect count, not time-to-repair a learning error.
• IBM Systems Sciences Institute (widely cited); Boehm defect-cost curve; Labovitz & Chang 1992: late defects cost more than early ones (design-to-release ~4–5×; earlier phases up to 100×). Different 4x: phase of discovery. Fits agent-output better than study errors.
• Newell & Rosenbloom 1981, power law of practice; Fitts & Posner 1967, autonomous stage: hands get faster once the method is known. Support for unforced speed as a habituation mark.
• Bouton 2004: extinction is not unlearning. Support for early-automated errors being expensive to reverse.