HN Debrief

Suspicious Discontinuities (2020)

  • Economics
  • Regulation
  • Public Policy
  • Data
  • Infrastructure

Dan Luu’s post is a tour of “suspicious discontinuities,” places where a smooth-looking human process suddenly develops a spike because an arbitrary threshold changes the stakes. The examples range from a giant pileup of Polish exam scores right above the passing mark, to marathon finish times clustering just under round-number goals, to income cliffs where earning slightly more can wipe out health subsidies or other benefits. The point is not just that people respond to incentives. It is that a simple histogram can expose where a system has turned a continuous reality into a cliff.

If a metric, benefit, or rule in your business or product has a hard cutoff, expect people to bunch around it and optimize for the line instead of the goal. Audit your own thresholds now, because the failure mode is often invisible until you graph the distribution.

Discussion mood

Engaged and positive. People liked the core idea because it matches lived experience, and many immediately supplied concrete examples of cliffs in taxes, benefits, testing, and service metrics that create perverse behavior.

Key insights

  1. 01

    Benefit cliffs change real decisions

    Benefit and tax thresholds do more than look ugly on a chart. They push people and firms into worse choices because stacked phase-outs across programs can create crushing effective marginal rates. The UK examples stood out. Frozen thresholds, the personal allowance taper, childcare cutoffs, and the VAT registration line can all make extra income barely worthwhile or actively harmful, and one commenter pointed to The Economist’s chart suggesting firms stay small to avoid the compliance jump.

    Do not evaluate a tax credit, subsidy, or internal compensation rule in isolation. Model the combined marginal effect across all adjacent programs and thresholds, because that is what people actually optimize against.

      Attribution:
    • mnahkies #1
    • rwmj #1
    • Georgelemental #1
    • entrope #1
    • zipy124 #1
  2. 02

    Cliffs reshape family life, not just paychecks

    The damage from hard cutoffs is often personal and long-lived. One example described a widow losing more than 100 euros of support because her pension rose by 2 euros after inheriting part of her husband’s pension. Another pointed out that household-based thresholds can delay young adults moving out or starting work because a small change in combined income can strip benefits from parents or siblings. These systems do not just tax income. They steer life paths.

    If your product or policy interacts with household eligibility, test for edge cases around bereavement, cohabitation, and dependent status. Those are exactly where small arithmetic changes can cause outsized harm.

      Attribution:
    • benjiro29 #1
    • TeMPOraL #1
    • ajsnigrutin #1
  3. 03

    Targets create bunching even in good faith

    The AWS latency example made the post’s point feel operational instead of academic. Engineers clustered service latency just under P50 and P90 targets, which produced visible fence-post effects. The important claim was not simple metric gaming. It was that a percentile target turns an arbitrary boundary into the thing teams optimize, even when they are trying to help. Once that happens, performance just beyond the line gets neglected.

    When you set SLOs or KPI thresholds, inspect the full distribution, not just the headline percentile. If you see bunching just under the target, redesign the metric before teams learn to ship to the line.

      Attribution:
    • dadkins #1
    • fc417fc802 #1
    • TeMPOraL #1
  4. 04

    Exam score spikes expose hidden grading slack

    The pass-mark spike in the Polish exam example looks less like random noise and more like discretionary rescue around a life-changing threshold. Commenters argued that if a few generosity points can move many students from 28 or 29 to 30, then the grading precision is already coarser than the exact numeric score suggests. That matters because passing the matura functions as a basic credential in the labor market, so teachers have a strong reason to avoid failing students who are close.

    Be skeptical of fine-grained scores when the process includes subjective grading and a high-stakes cutoff. If your system acts on those numbers, treat them as bands or confidence ranges unless you can prove tighter measurement.

      Attribution:
    • irishcoffee #1
    • jameshart #1
    • rvba #1
    • TeMPOraL #1 #2
  5. 05

    Round numbers pull behavior everywhere

    The marathon and chess examples broadened the post beyond public policy. People naturally fixate on clean goals like 2:30, 3:30, or rating milestones divisible by 100, then make a late push to cross them or protect them. In races, pace groups amplify that effect by making the target visible and social. In competitions and age-binned events, threshold structure also changes who shows up in the first place. The pattern is not rare. It is a default human response to salient lines.

    Any visible milestone in a product, game, or marketplace will attract optimization pressure. If the milestone is arbitrary, expect user behavior to become arbitrary around it too.

      Attribution:
    • ziofill #1
    • fwipsy #1
    • chantepierre #1
    • brightbeige #1
    • jtolmar #1
  6. 06

    Undated posts invite false objections

    One of the liveliest corrections had nothing to do with discontinuities. A reader rejected the opening health-insurance example based on 2025 subsidy rules, then withdrew the criticism after others established the article was published in February 2020. That was a useful reminder that an undated post forces readers to guess the policy context, which is risky for anything tied to changing law or regulation.

    If you publish analysis that depends on current rules, put the date where nobody can miss it. If you consume that analysis, verify the publication date before arguing from present-day policy.

      Attribution:
    • XRG #1
    • forbiddenlake #1
    • PaulDavisThe1st #1
    • dvh #1

Against the grain

  1. 01

    Universal subsidies may beat phase-outs

    Instead of smoothing cliffs with slower tapers, one comment argued for removing the tapers entirely. The claim is that higher earners already contribute more through taxes, so recapturing subsidy costs through the tax system is cleaner than building eligibility machinery that creates cliffs and complexity. Universal programs also keep wealthier users exposed to the system’s failures, which can improve political pressure to fix them.

    When you see a means-tested benefit with ugly edge effects, do not limit the redesign space to gentler taper curves. Compare it against a universal version plus tax-side recapture.

      Attribution:
    • amluto #1
  2. 02

    People often misunderstand tax bracket effects

    Not every reported cliff is real. Several comments pushed back on the common claim that crossing a tax bracket leaves you with less money overall, noting that people routinely confuse marginal rates with average rates. Another useful nuance was that some people saying they would "lose money" really mean the extra work is not worth the modest net gain after reduced benefits and lost time, which is different from a literal drop in take-home pay.

    Separate actual discontinuities from bad mental models before redesigning a system. You need distribution data and net-income calculations, not anecdotes about tax brackets.

      Attribution:
    • bostik #1
    • degamad #1
    • procaryote #1
    • encoderer #1

In plain english

matura ↩
A secondary-school leaving exam used in several European countries as a credential for graduation, higher education, or employment.
P50 ↩
The 50th percentile, meaning half of observed values are below this point.
P90 ↩
The 90th percentile, meaning 90 percent of observed values are below this point.
VAT ↩
Value-added tax, a consumption tax charged at each stage of production and sale.

Reference links

Tax and benefit cliff examples

Metrics and performance targets

Behavior around visible thresholds

Policy design and legislation references