retirement_assets

Risk-based spending guardrails

Preset names. Since v11.18b2 (2026-09-20) the page calls the presets High Safety, Normal, More Tolerant and More Risk; this report keeps the names it was run with: Tight = High Safety, Loose = More Tolerant (whose cut has since returned to 70% instead of its 80% target), Paper = More Risk. The keys in the code (tight, normal, loose, paper) never changed.

Whether the guardrail rule published by Derek Tharp and Justin Fitzpatrick - spend to a target probability of success, adjust when that probability crosses a rail - can be computed on this engine, what it costs, and where it disagrees with the Guardrails rule the tool already ships.

Sections 6 to 8 answer questions the comparison kept raising and the article did not ask: how far the shipped rule is from Guyton-Klinger as published. Four divergences, and the first one - the rate it tests - is what makes the hatchet invisible to it. Section 7 answers a second: how the rule composes with Spend Delta, which turns out to break a filter the Optimizer relies on, and which is where the gkShapeCeiling prototype came from. Section 8 re-scores every spending total on the page for the fact that early spending is worth more than late spending, which moves the rule and the ceiling in opposite directions.

Produced by .test_harnesses/rbg_harness.js, first on engine v11.1823 (1842270, 2026-09-15) and again on v11.1854 (2026-09-16). Every number below is printed by that script; nothing here is estimated by hand.

What the 2026-09-16 re-run changed, and why. Two things moved under this report, and every table was re-read against the re-run.

Reading guide

The source

Derek Tharp and Justin Fitzpatrick, “The Retirement Distribution ‘Hatchet’: Using Risk-Based Guardrails To Project Sustainable Cash Flows”, Kitces.com, 24 November 2021. Two claims matter here.

The hatchet is the shape of a real retiree’s portfolio draw: high in the years between retirement and a deferred Social Security benefit (the blade), a permanent drop when the benefit starts, then a slow decline as real spending falls with age (the handle). The article’s own example household draws 6.7% initially and $10,500/year at 90.

The rule replaces the withdrawal rate with the risk of the whole plan, and the article states it as a four-step recipe: solve for the spending that hits a target probability of success; choose an upper and a lower rail; find the portfolio values that would put the plan on those rails today, so the rails can be given to the household in dollars rather than percentages; solve for the spending that returns the plan to target at each rail. The article names two parameter sets and invents no others; both are used here, unchanged.

Codes

code meaning
PoS probability of success: the share of Monte Carlo paths in which totals.success is true, i.e. every year of the plan is funded
the shipped rule spendRule: 'gk', the Guyton-Klinger spend adjustment the tool labels Guardrails (resolveSpendTarget in optimizer_core.js). Trigger (as measured here, through 11.18b1): spendGoal / start-of-year portfolio against that same ratio in year 0; since 11.18b2 (P132j) the ratio is net of Social Security and pension and is compared with the plan’s own ratio for the year, so a plan on its own assumptions never adjusts, band 1 ± gkGuard (0.20), step gkAdjPct (10%). Since P127 (2026-09-16) it also never cuts in a plan’s final 8 years, caps the inflation raise at 6%, and freezes the raise on the portfolio’s return rather than the market’s
rails A the article’s implementation recipe: target 90% PoS, raise at 99%, cut at 70%
rails B the article’s income-risk framing: “target an initial income risk of 20% … increase at 0% … decrease at 60%”, read as PoS: target 80%, raise at 100%, cut at 40%
at its target the rail measured from the spending that rule would have the household on - the target-PoS number, not the plan’s own
at the plan’s own spend the same rail measured from the spending the household already has, which is what the shipped rule judges. The controlled comparison
first-year return the single market return in year 1 at which a rule first acts, found by bisection. One number that both rules can be asked for
published GK Guyton-Klinger as published - Guyton 2004, and Guyton and Klinger, “Decision Rules and Maximum Initial Withdrawal Rates”, Journal of Financial Planning, March 2006. Four rules: a portfolio-management rule for where withdrawals come from, an inflation rule that skips the CPI raise after a negative-return year while the current rate is above the initial one, a capital-preservation rule cutting 10% when the rate rises 20% above its initial value (suspended in the final 15 years), and a prosperity rule raising 10% when it falls 20% below
the published ratio the rate published GK tests: portfolio withdrawals divided by the portfolio, against that same ratio in year 0. Distinct from the shipped rule’s ratio, which divides total spending by the portfolio. The engine computes both; yr._wdRate (optimizer_core.js:4447), the wdRate% column, is the published one
the shape the plan’s own spending path: the year-0 goal carried forward by Spend Delta (spendChange, the household’s planned real drift) and CPI, and by nothing else. What spending would have been with no rule running
gkSpendStable the filter the Optimizer’s spend and conversion searches apply (optimizer_core.js): reject a candidate whose run lets real spending fall more than gkGuard below its year-0 value - below the shape, since P127b. Not a rule a household follows - a guard against a search “affording” a spend the rule only reaches by slashing
gkShapeCeiling the P127 prototype added by this report: an engine input, default off, that holds the rule’s goal at the shape. Since 2026-09-16 the Never above plan switch behind the nerdknob (P127a)
discounted real spend the same delivered real spending, with year y weighted 1 / (1 + d)^y. Section 8 reports d = 0%, 3% and 5%, because the rate is a statement of preference and not a measurement. Per-year real spending is (spendGoal + shortfall) / inflationFactor, which sums to totals.spendCurrentDollars to the dollar - the engine’s own accumulator (optimizer_core.js:4386) decomposed, not a second definition
re-plan a plan continued from the end of one of its own years with a different future: the run’s -resume record handed to a new run (resumeInputs), which carries the balances, the plan-year index, the inflation clocks and every other year-to-year state, and the spending the plan had. Before the 2026-09-16 re-run this was a hand-built restart - see the note at the top

The households

Four from the plan bank, all with the hatchet’s shape - retirement years before a deferred benefit.

household retires benefits at portfolio plan spend plan years draw state
bracket-filler-texas 64 70 / 67 $3,100,000 $110,000 33 bracket TX
mixed-portfolio-couple 64 70 / 67 $3,900,000 $234,000 33 propwd CA
long-widowhood 65 70 / 67 $3,620,000 $150,000 38 fixed, Cycle Brokerage on TX
age-gap-ira-heavy-ca 65 70 / 70 $4,414,000 $220,000 25 fixed, Cycle Brokerage on CA

The four differ in draw strategy and state as well as in shape, which is deliberate: both rules sit on top of whatever draw the plan already uses, and a result that held only under one strategy would be a result about that strategy.

age-gap-ira-heavy-ca is the only one carrying a planned real-spending decline (spendChange -1%/year), which is the handle of the hatchet. Every spending comparison below is scored against each household’s own planned real path, year by year, so that decline is never counted as a rule’s doing.

The grid

knob value
model synthetic GBM, mu 7%, sigma 12% - the page’s own #mc-mu / #mc-sigma defaults
paths 200 per probability estimate, one seed (42), common random numbers across every arm of a household
bisection 9 steps, which puts a spending answer inside ~0.4% of its bracket
the shock -22% then -13% in years 2-3 (the 2000-2002 shape), then the plan’s own growth

What it found

1. The recipe runs on this engine unchanged, at a measurable price

All four steps are already available: simulate() scores a path, buildBanks / buildPathInputs (montecarlo/mc_engine.js:198, :45) produce the paths, and bisection over spending or over starting balances does the rest - the same shape as optimizeSpend (optimizer_core.js:5436), against a probability instead of a pass/fail. Nothing new had to be added to the engine to produce every table in this report.

household plan years engine runs seconds ms/run
bracket-filler-texas 33 18,400 14.2 0.77
mixed-portfolio-couple 33 22,000 19.0 0.86
long-widowhood 38 20,200 21.7 1.07
age-gap-ira-heavy-ca 25 18,400 14.8 0.81

The run count is exact and the seconds are not: seed 42 reproduces every probability in this report run for run, while wall clock moved 0.99 to 2.15 ms/run across four repeats of the first run on the same machine (the table is the 2026-09-16 re-run; two households run 1,800 fewer engine runs than they did, because section 3 no longer bisects a spend the rails leave unchanged). Plan against the run count, not the clock.

One probability estimate is 200 engine runs; one spending or rail answer is 9 estimates. The harness runs about 100 estimates per household because it also measures things a product would not - two rail sets, five spending targets, a controlled column. The four-step recipe itself is 1 estimate plus 5 bisections = 46 estimates = 9,200 runs, 9-20 seconds per household at these speeds, which is the same order as the Monte Carlo runs the tab already performs (400 paths x ~123 strategies).

2. The shipped rule is BLIND to the hatchet, not tripped by it

This was predicted the other way round (G-P1), and the prediction was wrong. The article’s critique is about a rule that mistakes the hatchet’s high early withdrawal rate for overspending. The rule this tool ships does not measure that rate: its ratio is total spending over the portfolio, not the portfolio draw over the portfolio, so a benefit starting at 70 changes the numerator not at all and the denominator only slowly.

On a path with the market held perfectly still, the ratio’s peak in the pre-benefit years:

household peak ratio, years 1-6 band rule fired
bracket-filler-texas 1.000 1.20 no
mixed-portfolio-couple 1.179 1.20 no
long-widowhood 1.006 1.20 no
age-gap-ira-heavy-ca 1.098 1.20 no

So the defect here is not a false alarm during the blade. It is that the rule’s one reference number is set in year 0 and never learns anything - not that the benefit arrived, not that the horizon shortened, not that the plan is now funded from a different mix. Everything in section 3 follows from that, and section 6 has the reason: the rate it tests is not the rate Guyton-Klinger publishes, and the substituted one is the one the hatchet leaves alone.

3. It acts on smaller market moves than either rail, and cuts much harder than the plan’s risk calls for

The first-year market return at which each rule first acts:

household shipped rule rails A, at its target rails B, at its target rails A, at the plan’s own spend
bracket-filler-texas -14.4% -17.1% -33.0% -44.7%
mixed-portfolio-couple -11.7% -17.4% -34.4% +24.1%
long-widowhood -15.5% -22.1% -31.4% -25.1%
age-gap-ira-heavy-ca -14.1% -23.1% -30.4% -32.7%

The shipped rule cuts on a first-year drop of 12-16% in every household, and the article’s own rails want 17-23% (A) or 30-34% (B). Read as portfolio balances - the form the article asks for - the bracket-filler-texas household is told to cut at $2,584,564 by the shipped rule and at $2,470,162 (A) or $1,980,506 (B) by the rails.

What that early trigger costs, on the -22%/-13% shock, scored against each household’s own planned real spending path:

household trough real spend vs plan path years below plan lifetime real spend, rule on rule off did the un-adjusted plan fund every year?
bracket-filler-texas $76,326 -30.6% 24/33 $3,025,281 $3,630,000 yes
mixed-portfolio-couple $131,517 -43.8% 31/33 $4,779,073 $4,932,387 no
long-widowhood $103,073 -31.3% 35/38 $4,439,592 $5,700,000 yes
age-gap-ira-heavy-ca $118,198 -38.2% 23/25 $3,539,735 $4,887,930 yes

On three of the four households the shipped rule gives up 17% to 28% of lifetime real spending, on a path where the plan needed no protection at all: the un-adjusted run funds every year of the plan. The cut is not brief, either - real spending sits below the plan’s own path for 23 of 25, 24 of 33 and 35 of 38 years.

At the same moment, with spending never having been cut, the risk-based read:

household year portfolio PoS rails A says rails B says
bracket-filler-texas 2030 $1,999,135 92.0% no change, $110,000 no change, $110,000
mixed-portfolio-couple 2032 $1,677,238 2.0% cut to $125,741 cut to $134,539
long-widowhood 2030 $2,452,349 55.0% cut to $116,052 no change, $150,000
age-gap-ira-heavy-ca 2031 $2,539,995 74.0% no change, $207,126 no change, $207,126

(All in 2026 dollars, against the shipped rule’s $76,326 / $131,517 / $103,073 / $118,198. The year is the one the shipped rule stopped cutting; the probability is the plan’s from the end of that year on. age-gap-ira-heavy-ca’s “no change” is its own planned spending for 2032, which its -1%/year Spend Delta has already taken below $220,000.)

This table was wrong in the first run, and the correction strengthens the section rather than weakening it. The first run asked each household about 10-15% more real spending than it had (see the note at the top), which read 82%, 0%, 39.5% and 41%. At the spending they actually have, the rails leave two of the four households alone entirely - bracket-filler-texas at 92% and age-gap-ira-heavy-ca at 74% - where the shipped rule took 31% and 38% out of their spending, and rails B leaves long-widowhood alone too.

4. Risk-based is not simply GENTLER - it is aimed differently

The second half of G-P2 fails under rails A, and the exception is the interesting household. mixed-portfolio-couple spends $234,000 against a 55% probability of success: the risk-based rule says cut today, before any market move at all (its “cut at +24.1%” in section 3’s controlled column means the household would need a 24% first-year gain merely to get back above the 70% rail), and after the shock rails A cut it to $125,741 where the shipped rule stopped at $131,517 - deeper, on the one household whose un-adjusted plan genuinely fails (rails B stop at $134,539). Meanwhile both rail sets leave bracket-filler-texas alone entirely at 92% PoS where the shipped rule took a third of its spending away, and age-gap-ira-heavy-ca at 74% where it took 38%.

That is the whole difference in one line: the shipped rule reacts to a market move, the rails react to the plan.

5. Rails cluster only because the target normalizes the households

Prediction G-P3 expected the rails to vary household by household while the shipped trigger stayed put. Measured at each rule’s own starting point, both are flat: the shipped trigger spans 3.8 points across the four households and rails A span 6.0 (B: 4.0). The reason is in the recipe - resetting every household to the same probability of success is the normalization, so afterwards they all sit the same distance from the rail.

The variation reappears the moment the rails are asked about the spending a household actually has: -44.7%, +24.1%, -25.1%, -32.7% - a 69-point spread on the same rail. That is the column to quote when the question is “how different are these households”, and the “at its target” column is the one to quote when the question is “what would the rule do”.

6. The rule this tool ships is not Guyton-Klinger as published

Four divergences, measured on the rule as it stood on 2026-09-15. The first one decides sections 2 and 3, because it is the reason the hatchet is invisible to the rule. The other three were acted on the next day (P127, user): b and d now follow the paper, and c follows it with 8 years in place of

  1. Section a is the one that stays, by the user’s choice.

a. It tests a different rate. Published GK compares portfolio withdrawals to the portfolio. The shipped rule compares total spending to the portfolio (resolveSpendTarget in optimizer_core.js) - including the part Social Security and pensions pay for. The engine computes the published quantity already, yr._wdRate, and the rule does not use it.

Both ratios, read off one run with the rule off, at the plan’s own deterministic growth and CPI, as multiples of their own year-0 values against a 0.80-1.20 band:

household shipped ratio years outside published ratio years outside first divergence
bracket-filler-texas 0.43-1.00 18/33 0.21-1.02 27/33 2032 raise - published 0.77, shipped 0.96
mixed-portfolio-couple 1.00-3.59 26/33 0.91-2.30 15/33 2044 cut - published 1.20, shipped 1.57
long-widowhood 0.54-1.01 16/38 0.50-1.27 34/38 2027 cut - published 1.27, shipped 0.99
age-gap-ira-heavy-ca 0.74-1.10 3/25 0.45-1.24 22/25 2027 cut - published 1.24, shipped 1.00

Why the two move apart, in one household’s arithmetic. The rates are not independent - they are related by exactly one factor:

published rate = shipped rate x (portfolio draw / total spending)

and a deferred benefit is a thing that moves that factor. bracket-filler-texas, rule off so neither rate is being reacted to:

year age portfolio spend SS + pension portfolio draw draw ÷ spend shipped rate published rate
2026 64 $3,100,000 $110,000 $0 $145,710 132% 3.55% 4.70%
2030 68 $3,503,291 $121,419 $0 $166,862 137% 3.47% 4.76%
2031 69 $3,611,341 $124,455 $16,971 $154,203 124% 3.45% 4.27%
2032 70 $3,740,659 $127,566 $40,589 $135,011 106% 3.41% 3.61%
2033 71 $3,899,518 $130,755 $59,434 $120,710 92% 3.35% 3.10%
2037 75 $4,715,529 $144,330 $65,604 $133,961 93% 3.06% 2.84%

Before the benefits the portfolio funds everything, and more: the draw is 132-137% of spending, because it also funds the tax on its own withdrawal. Once both benefits are running it funds 92%. That collapse is the whole divergence - the published rate falls by a third, 4.70% to 3.10%, while the shipped rate drifts from 3.55% to 3.35%.

The reason the shipped rate barely moves is worth stating plainly, because it is the intuition the numerator breaks: a benefit does not reduce spending, it changes who pays for it. The shipped rate’s numerator is the spending, so the benefit never enters it; the denominator is the portfolio, which is only indirectly affected (it is drawn on more gently, so it grows a little faster). The published rate’s numerator IS the draw, so the benefit lands on it in full, the year it starts.

Which means the adjustment in the normalized table is not a near miss. On this household the published rate reads 0.768 of its year-0 value in 2032 and 0.659 in 2033 - through the 0.80 prosperity trigger - so published GK calls for a 10% raise, twice, on a household where nothing unexpected has happened at all: the benefit starting at 70 was in the plan from the first day. That is the article’s complaint stated as arithmetic rather than as an opinion, and it is what “a rule that judges a rate cannot tell a planned change in that rate from an unplanned one” means.

The two ratios disagree about the direction and the timing of the first adjustment on every household, and neither is uniformly the earlier one: the published ratio acts in year 2 on two households where the shipped rule sits at 0.99 and 1.00, and the shipped rule reaches 3.59 on mixed-portfolio-couple where the published ratio has not yet left the band.

That is the article’s critique, reproduced: under the published numerator these households behave exactly as the hatchet predicts - a cut in the blade years on two of them, and a raise on bracket-filler-texas in 2032, the year both benefits are running. The shipped rule escapes that by testing a rate the hatchet does not move, and pays for it with the year-0 anchor of section 2.

What the last column is not: the first year the published ratio leaves the band on a path where no adjustment is ever applied. A real published-GK run would adjust at that point and move its own ratio afterwards, so this is the first divergence, not a simulation of the published rule.

b. The inflation freeze read the wrong return - follows the paper since P127. The rule froze on yr.baseReturn, the scenario’s base (equity) return; published GK freezes on the portfolio’s total return, and since 2026-09-16 so does this one (portfolioReturnOf: the accounts’ returns weighted by their start-of-year balances, cash yield and brokerage dividends included). The two were identical where every account gets the base return, and not otherwise:

mode equity-negative path-years sign disagreement with the blended account return
Historical (bootstrap) 27.1% 7.7%
Synthetic (GBM) 29.9% 0.0%

So in Historical mode roughly one year in thirteen froze the CPI raise on an equity loss the portfolio did not actually take. (The blend here is the average of the per-account sequences the engine hands the accounts, which is the closest stand-in this harness has for a portfolio return. The engine’s own blend also counts cash yield and dividends, so it can disagree with the equity return under GBM too, in a year the market loses less than the cash earns.)

c. The capital-preservation cut was never suspended - suspended in the final 8 years since P127. Published GK stops applying it in the final 15 years of the plan; this one applied it to the last year, and now stops in the final 8 (the user’s number). Latent on the shock paths measured - every cut lands early, before either window:

household plan years cuts of those, in the final 15 held back in the final 8 raises
bracket-filler-texas 33 3 0 0 6
mixed-portfolio-couple 33 5 0 0 1
long-widowhood 38 3 0 0 4
age-gap-ira-heavy-ca 25 4 0 0 3

Early clustering is itself a consequence of (a) and of the year-0 anchor, not evidence that the suspension never matters. Where it does matter is a plan that spends more than its savings carry. One number from outside this harness, because it is the case: on the Optimize Spend fixture in optimizer_core.tests.js, the late cuts the rule now holds back were what stopped the search, and its answer rose from $64,829 to $78,687 with every year still funded (the test’s comment records it).

d. There was no 6% cap on the inflation raise - applied since P127. Published GK (2006) caps it. Inert on a deterministic 2.5% CPI run; under the Monte Carlo inflation model 11.1% of path-years draw inflation above 6%, worst 11.6%.

Three more things that are not GK, and are not defects. The portfolio-management rule - where withdrawals come from - is deliberately out of scope: the draw belongs to whichever strategy the plan selected, as simulate()’s strategy notes and the release that made Guardrails a switch rather than a strategy both say. spendDelta compounds on the goal every year (endYear), which the page discloses as “Spend Delta still applies on top”. And gkSpendStable rejects plans the rule only holds together by slashing - a guard the Optimizer’s searches apply, not a rule a retiree follows.

What to do about (a) was a decision, not a fix, and it has been made. The in-page copy describes the implemented mechanism accurately - “what you spend for each dollar you have saved” - so the overclaim was the label, not the explanation. Switching the numerator to yr._wdRate would move the numbers of every saved plan that has Guardrails on, and on the evidence above it would make the rule act in the blade years, which is the behavior the source article was written to argue against. The switch was relabelled GK-style in 11.1824, and on 2026-09-16 the user chose to keep the numerator.

7. How the rule composes with Spend Delta

The rule does not run alone. spendChange - Spend Delta on the page - shapes the goal the rule is adjusting, and the two compose in a way nobody chose: one defect in a filter, and one thing worth adding, prototyped here.

a. The search filter could not tell a planned decline from a rule-driven slash - fixed, P127b

gkSpendStable read the finished run’s minimum real spendGoal against its year-0 value, and spendGoal carries the delta. So the shape breached the floor on its own: at -1%/year, real spending is below 1 - gkGuard of year 0 from year 23, with no market move and no cut. Since 2026-09-16 it compares each year against the shape, spend0 x (1 + spendChange)^y, which at a flat delta is the same test it always was. Measured at each household’s own growth, so every run below is a plan doing exactly what it planned; the verdict is shown from the first run and from the re-run:

household delta GK cuts min real ÷ year 0 against year 0 (first run) against the shape (re-run)
bracket-filler-texas 0 0 1.000 accepted accepted
bracket-filler-texas -1%/yr 0 0.914 accepted accepted
mixed-portfolio-couple 0 2 0.810 accepted accepted
mixed-portfolio-couple -1%/yr 0 0.725 REJECTED accepted
long-widowhood 0 0 1.000 accepted accepted
long-widowhood -1%/yr 0 0.878 accepted accepted
age-gap-ira-heavy-ca 0 1 0.900 accepted accepted
age-gap-ira-heavy-ca -1%/yr 0 0.810 accepted accepted

Under the old baseline one household was rejected for its own planned shape with zero cuts on the path, and the other three survived for a reason no better than the failure: the rule’s prosperity raises happened to offset the decline. That made the verdict depend on how the portfolio performed, which is not something a filter over candidate spending levels should be sensitive to - and it was not scale-invariant either. On mixed-portfolio-couple at -1%/year the old filter accepted $93,600, $140,400 and $187,200 and rejected $234,000 and $280,800: the smaller goals earned raises, the larger ones did not.

What it did to the search it guards. optimizeSpend drives its binary search on passes(), which fails whenever this filter does:

household delta Guardrails on, first run Guardrails on, re-run Guardrails off
bracket-filler-texas 0 $188,311 $188,311 $195,723
bracket-filler-texas -1%/yr $168,652 $222,471 $223,760
mixed-portfolio-couple 0 $235,371 $240,855 $241,541
mixed-portfolio-couple -1%/yr no viable spend $281,303 $271,705
long-widowhood 0 $208,887 $217,676 $213,281
long-widowhood -1%/yr $195,703 $256,787 $242,725
age-gap-ira-heavy-ca 0 $269,629 $278,008 $287,031
age-gap-ira-heavy-ca -1%/yr $225,801 $317,324 $316,035

On the first run, a declining shape - a plan that is cheaper, and whose rule-off answer therefore rises by 10-17% - made the Guardrails answer fall by 8-16%, or disappear, widening the gap from about 4% to 25-29% for no reason the household would recognize.

The re-run carries all of P127, not only the filter: the flat-delta rows moved too, where the filter is unchanged, so those moves are the rule’s own - chiefly that it no longer cuts in a plan’s final 8 years. With a declining shape the Guardrails answer now follows the rule-off answer, and on four rows exceeds it: a plan that may still cut earlier on, within the floor, can start higher than one that never adjusts.

b. gkShapeCeiling: the prototype, default off - now a nerdknob switch, P127a

The rule can spend a household above its own plan: a prosperity raise lifts the goal whenever the portfolio grows faster than the plan assumed, and nothing stops it. If the shape is read as “what we would have spent”, those raises are spending the household never asked for. The prototype holds the goal at the shape:

What it costs and returns, on each household’s own deterministic path and on the -22%/-13% shock:

household path peak real ÷ shape clamped years lifetime real spend ending wealth
bracket-filler-texas benign 1.772 → 1.000 18 $4,340,533 → $3,630,000 $14,846,622 → $17,075,004
bracket-filler-texas shock 1.229 → 1.000 5 $3,025,281 → $2,956,046 $10,230,520 → $10,397,129
mixed-portfolio-couple benign 1.000 → 1.000 0 $6,797,700 → $6,797,700 $6,071,086 → $6,071,086
mixed-portfolio-couple shock 1.000 → 1.000 0 $4,779,073 → $4,779,073 $6,437,720 → $6,437,720
long-widowhood benign 1.464 → 1.000 16 $6,285,015 → $5,700,000 $18,471,164 → $20,812,089
long-widowhood shock 1.006 → 1.000 1 $4,439,592 → $4,438,683 $12,597,671 → $12,600,402
age-gap-ira-heavy-ca benign 1.100 → 1.000 3 $4,940,308 → $4,887,927 $8,962,648 → $9,080,064
age-gap-ira-heavy-ca shock 1.000 → 1.000 0 $3,539,735 → $3,539,735 $6,872,188 → $6,872,188

Three things to read off it. The whole effect is in good markets - on the shock path the ceiling takes 2.3% of lifetime spending at most, and on two households nothing at all. When it binds it binds hard: bracket-filler-texas gives up 16.4% of lifetime real spending for 15.0% more ending wealth, and long-widowhood 9.3% for 12.7%. And a household the rule never raises never notices it: mixed-portfolio-couple is identical to the cent on both paths, which is the same property the third test pins.

One side effect worth knowing. The rule’s anchor is still the year-0 ratio, so once the portfolio outgrows it the rule tries to raise every remaining year and is clamped every time: 18 of 33 years on bracket-filler-texas. The ruleAdj column reads +10%pros @shape for all of them. That is honest - the rule wanted more and the plan said no - but it means the ceiling is a permanent state rather than an occasional event; the column’s help text says what @shape means.

And it composed badly with (a), in the direction you would expect: with the ceiling on, real spending can never exceed the shape, so the minimum could only fall further below a year-0 baseline. That is why the filter’s baseline was fixed first (P127b), before the switch reached the page.

8. Early spending is worth more, and every total above prices it at par

Sections 3 and 7 add year 1 and year 33 at the same value. A household does not: a cut at 65 and a raise at 92 are not the same event (user, 2026-09-15). Re-scoring the same runs with the year’s real spending discounted at 0%, 3% and 5% moves the two rules in opposite directions, so this is not a presentational nicety.

household path comparison 0% 3% 5% first 10 years
bracket-filler-texas benign the rule vs no rule +19.6% +14.0% +11.0% 0.0%
bracket-filler-texas benign the ceiling vs not -16.4% -12.3% -9.9% 0.0%
bracket-filler-texas shock the rule vs no rule -16.7% -18.9% -19.8% -21.9%
bracket-filler-texas shock the ceiling vs not -2.3% -1.5% -1.1% 0.0%
mixed-portfolio-couple benign the rule vs no rule -12.0% -10.0% -8.8% -3.0%
mixed-portfolio-couple shock the rule vs no rule -3.1% -12.1% -16.1% -27.8%
long-widowhood benign the rule vs no rule +10.3% +6.7% +4.9% 0.0%
long-widowhood benign the ceiling vs not -9.3% -6.3% -4.6% 0.0%
long-widowhood shock the rule vs no rule -22.1% -23.1% -23.3% -22.4%
age-gap-ira-heavy-ca shock the rule vs no rule -27.6% -27.1% -26.6% -25.4%

The rule’s gains are late and its cuts are early. On a benign path the rule delivers bracket-filler-texas 19.6% more lifetime spending and none of it in the first ten years - every dollar of that gain is a prosperity raise that compounds from the middle of the plan onward, so at a 3% discount it is worth 14.0% and at 5% it is worth 11.0%. On the shock path the cuts land in the blade years and discounting makes them worse, not better: -16.7% becomes -19.8%.

mixed-portfolio-couple is the case to read twice. On the shock its undiscounted cost is -3.1%, which reads as almost free. The same run takes 27.8% out of the first ten years and -16.1% at a 5% discount. A number that says “the rule cost this household 3%” is describing a plan nobody lives: the household gets its money back in its eighties.

The ceiling is the mirror image, which is the argument for it that section 7 could not make: it gives up late raises and takes 0.0% out of the first ten years on every household measured. Its headline -16.4% on bracket-filler-texas is -9.9% to a household discounting at 5%, and nothing at all to one that spends most of what it values before 75.

Three limits on all of this. A discount rate is a preference, not a fact, and picking one is the household’s business - which is why three are reported and none is called correct. Real discounting is also not exponential: survival probability falls with age and would steepen every number here, while a late-life health shock works the other way. And a fixed discount says nothing about variability - age-gap-ira-heavy-ca loses about 27% at every rate because its cuts are spread evenly, which a utility measure with risk aversion would score quite differently from the same loss concentrated in two years.

What it means for the rest of this report. Every “lifetime real spend” column in sections 3 and 7 is the 0% column here. On the evidence above it understates what the shipped rule costs (its cuts cluster early) and overstates what the ceiling costs (its givebacks cluster late). Any future comparison of spending rules in this repo should carry at least the first-decade share beside the total; it is one line of arithmetic and it changed the sign of the argument twice on this page.

Predictions

Registered before the first run, against a 25% lower rail - the number the second-hand summaries available at the time gave, since kitces.com is unreachable from this environment. The source document arrived afterwards and carries two parameter sets, neither of them that one, so each prediction is scored against both of the article’s sets rather than re-aimed.

id prediction verdict
G-P1 the shipped rule fires in the pre-benefit years on a benign path REFUTED. Closest approach 1.179 against a 1.20 band; it fires on none of the four. Section 2 is what replaced it
G-P2 the shipped rule cuts earlier AND deeper than a risk rail, on every household PARTLY under rails A: earlier on all four, deeper on three - mixed-portfolio-couple reverses it. CONFIRMED under rails B on the re-run; the first run’s PARTLY there came from the re-plan defect noted at the top
G-P3 the shipped trigger is household-independent; the rail moves 20+ points REFUTED, both sets. 3.8 points against 6.0 (A) and 4.0 (B) on the re-run (5.3 and 4.3 on the first). Section 5 explains why, and where the variation actually lives
G-P4 a full guardrail set costs under 30s per household at 200 paths REFUTED as the first run measured it - 32 to 47s on the longest household across repeats - though the re-run took 21.7s, and the harness measures about twice the recipe. The recipe alone is 9,200 runs, 9-20s

What shipping it would take

Three scopes, smallest first. Only the first is priced by anything in this report.

  1. A guardrail readout. Target-PoS spending, the two rail balances, and the adjusted spending at each rail, for the plan as it is. Every piece exists; the work is a panel, a worker call and a progress bar, on the pattern the Monte Carlo tab already uses. ~9,200 engine runs per plan. Built as P128 (2026-09-16), and on the page for everyone since the same day: solved at a cadence along the whole plan rather than once, drawn on the live charts, and reporting its own cost with a projection for other settings. The first solver, measured in the page rather than by this harness on its default plan at 60 paths: every 5 years, 13,800 runs in 5 seconds; every year, 66,240 runs in 16 seconds, against 18 projected from the first. RISK_BASED_RAILS_PRECISION.md times the per-path solver that replaced it.
  2. A spendRule: 'rbg' beside 'gk'. A simulated run that follows the rails year by year. The rails cannot be re-solved inside every path-year - that is 200 estimates x 30 years x 200 paths - so it needs the article’s own device: solve the rails once as balances, compare the running balance against them each year, and re-solve only when one is crossed. Whether that approximation holds up under Monte Carlo is not measured here and would need its own harness.
  3. The hatchet itself. Nothing is missing: benefit ages and spendChange already produce the blade and the handle. What is missing is the reporting that would make the shape visible.
  4. A spending metric that weights early years (section 8). The totals this report and the Optimizer’s Maximum Spending objective use price year 1 and year 33 the same. Carrying the first-decade share, or a discounted total, beside the lifetime figure is arithmetic, not modelling, and it changes which rule looks better.
  5. The two things section 7 left on the table, neither of which needs the rails: the gkSpendStable baseline and gkShapeCeiling. Both done 2026-09-16 (P127b, P127a), the baseline first because it was the prerequisite.

Open questions this report does not settle

This site is open source. Improve this page.