Cap Medicare drug prices at 120% of OECD international average. Saves ~$100B. **Illustrative**: an unfitted reconstruction a long way from its published figure, shown for the shape of the policy rather than as an estimate of it.
Ten-year deficit effect, conventional-$801.0Breduces the deficit
With macroeconomic feedback and debt service-$898.0BFRB/US-Lite (Federal Reserve calibrated)
Checked against
Published score -$100.0B; this estimate is 701.0% away (unfitted reconstruction). Unfitted reconstruction, 701.0% from -$100B (RAND price-comparison study) — no constant is fitted to this figure. The target itself is this model's own estimate, not a published score.
No out-of-sample band: no pre-registered row scores a drug-pricing reform. The 44 pre-registered Tier 1 rows are the only tier that measures this model's accuracy, and none of them scores a policy of this kind.
This is a spending or sector policy, so there is no tax table by income group.
Economic effects
difference from the CBO baseline
GDP, summed over the window-1.13percent-years
Jobs at the largest effect-149K
Deficit effect of the GDP change+$79.3Bless tax collected
Deficit effect of interest on the debt-$176.2Bless debt, less interest
Fiscal year
GDP level
Jobs
Unemployment
10-year rate
GDP change, deficit
Interest, deficit
2026
-0.203%
-130K
+0.10 pt
-0.01 pt
+$14.2B
-$3.2B
2027
-0.233%
-149K
+0.12 pt
-0.02 pt
+$16.3B
-$6.4B
2028
-0.203%
-130K
+0.10 pt
-0.03 pt
+$14.2B
-$9.6B
2029
-0.158%
-101K
+0.08 pt
-0.03 pt
+$11.0B
-$12.8B
2030
-0.116%
-74K
+0.06 pt
-0.04 pt
+$8.1B
-$16.0B
2031
-0.082%
-52K
+0.04 pt
-0.05 pt
+$5.7B
-$19.2B
2032
-0.057%
-36K
+0.03 pt
-0.06 pt
+$4.0B
-$22.4B
2033
-0.039%
-25K
+0.02 pt
-0.07 pt
+$2.7B
-$25.6B
2034
-0.026%
-17K
+0.01 pt
-0.08 pt
+$1.8B
-$28.8B
2035
-0.018%
-11K
+0.01 pt
-0.09 pt
+$1.2B
-$32.0B
FRB/US-Lite (Federal Reserve calibrated). Demand effects fade over five to seven years as monetary policy responds; there is no supply-side channel. The dynamic score is the conventional score, plus the deficit effects of the GDP change and of interest on the debt: -$801.0B + $79.3B - $176.2B = -$898.0B.
Other scoring models
exploratory
Model
Ten-year deficit effect
Method
CBO-Style
-$801.0B
Static + ETI behavioral response + optional FRB/US dynamic feedback
TPC-Microsim Pilot
Cannot score this
International Reference Pricing (DrugPricingPolicy) does not map onto current microsim pilot reforms (income-tax rates, CTC, EITC, SALT, std deduction, AMT exemption).
No pilot model returned distributional output; the comparison is aggregate-only.
CBO-Style is the model behind the headline above. The TPC-style microsimulation is a pilot: where the two disagree, the gap tells you how much the answer depends on method. It is not an official range.
Limitations
holdout labels follow the locked post-2026-05-02 regression protocol; they are not retroactive historical out-of-sample claims.
No pre-registered out-of-sample row scores a policy of this class, so this model's accuracy on one has not been measured.
Scored on all four of the thresholds the Green Book prints (report p. 78): $450,000 joint and surviving spouses, $425,000 head of household, $400,000 unmarried, $225,000 married filing separately. This is the one row in the battery whose base GREW under the filing-status split - the separate-return floor is $175,000 BELOW the unmarried one, so those returns had been under-counted, not over-counted - and the row moved 12.0% to 9.2%.
The thresholds are on taxable income and SOI's classes are on AGI. The thresholds are C-CPI-U indexed after 2024 by the proposal's own text and the model's do not move; the base now does, so the two no longer push the same way. This is a registered regression of the base-growth lane: the row was 9.2% UNDER and is now 18.0% OVER, because it had been under-predicting by less than a decade of nominal growth is worth (planning/lanes/HSB_h2_base_growth.md section 3.2).
The target is Treasury's single combined row - the rate change and the realization-at-death change together - and the model's two channels split it $359.0B of rate against $7.4B of death. The rate channel alone therefore exceeds the whole published figure by 24.4%, which is where this row's residual now lives and is a floor no change to the death channel can get under. It over-predicts by 27%; before Wave 5 projected the realizations base it under-predicted by 17%, on a base two years stale and never grown.
The death channel is integrated over a fitted size distribution of estates at death rather than five class means (Wave 7), which moved it $20.2B to $24.2B: with dispersion the top of the 99th-99.9th percentile band clears the $5,000,000 per-donor exclusion that its class average of $1.9M after reliefs never did. Wave C then took it $24.2B -> $7.4B by counting decedents at an NCHS life-table rate instead of at a flow of estate dollars, 3.38 million a year against 408,532: the same flow over 8.3 times as many decedents leaves far less of each gain above a $5,000,000 exclusion. That improved this row 32.8% -> 27.0% and worsened cbo_opt51_gains_at_death, which is the trade a level change to this channel makes. Neither move is where the residual lives.
At +19.6pp the semi-log response is evaluated well up its own curve: eliminating step-up divides the coefficient by the lock-in wedge, so b = 2.27 and the revenue-maximizing rate is 44.1% against a reform rate of 43.4%. Treasury's own estimate implies a much stronger response than the frozen literature value gives at that rate, and owner Decision 3 freezes the value.
The $1,000,000 threshold is not indexed, so the projected base is the above-$1M slice measured in tax-year-2023 AGI. Real bracket creep would push more gains across a fixed nominal threshold, which understates the base growth on a row that already over-predicts.
The same shape as biden_capital_gains_39 - +19.6pp above $1M plus realization at death - against a COMBINED FY2022 Green Book row, and since PR #126 scored on the ten fiscal years that document covers (scoring_window_first_year=2022, the treasury_capgains_39_plus_stepup_elim.v2 manifest row; the target is unchanged at -$322.0B and only the shape input moved). On FY2022-2031 the rate channel is $303.1B and the death channel $13.0B under this design's $1M per-donor exclusion. The row read 0.2% before Wave 5 projected the realizations base, and that agreement was already documented as two errors cancelling (planning/lanes/W4_gains_at_death.md section 8.4); one of the two was a tax-year-2023 base priced unchanged in every year of the window.
The window is worth 14.9 points, not the 'about 17' an early version of this note claimed and not the 27.0 the version before Wave C did - the offset shrank with the death channel, because the channel it discounts is smaller once the decedent count is a life-table rate. Re-scoring the same shape on the default FY2025-2034 decade returns -$375.8B and 16.7%, against -$316.1B and 1.8% on FY2022-2031; scripts/window_offset_capgains.py prints both. The old figure discounted the RATE channel only (359.02 x 0.844354 + 102.45 = 405.6) and missed that the death channel grows with the same 5.80% net-worth CAGR and that it does not grow proportionally, because a fixed nominal per-donor exclusion is a step function whose bite moves faster than the stock it is subtracted from. PR #126 measured the offset at 28.7 of 43.3 points and PR #132's decedent ladder took the merged figure to 27.0 of 45.4; Wave C's headcount took it to 14.9 of 16.7. The control is biden_capital_gains_39 - same module, same frozen elasticities, an FY2025 target on the FY2025 window - whose offset is exactly zero to the cent. See planning/memos/FY2022_TARGET_WINDOW.md.
1.8% is not accuracy, and the cancellation is larger than it was. On its own decade the row is a net: the model books $0.0B of revenue across FY2022-2024 where Treasury books $66.0B - the model's enactment year is a $61.4B transitory revenue LOSS against Treasury's $7.7B gain, because the module has no receipts lag - against $316.2B across FY2025-2031 where Treasury books $256.5B. A 23% over-prediction on the later seven years and a $66B hole in the first three net to 1.8%; read the annual path, not the total. Treasury's own answer for this row is not stable either: $322,485M, $174,488M, $213,855M and $288,583M across four consecutive Green Books, spanning 85%.
The gap between this row and biden_capital_gains_39 is the per-donor exclusion alone - $1M against $5M - and the model's price for that step on the common FY2025-2034 decade has now crossed the published one. Wave 4 blamed the five-class ladder for an $82.3B step against the two published rows' $33.4B; Wave 7 replaced the ladder with a fitted size distribution and the step went to $85.0B, the wrong way, so dispersion was not the cause (planning/lanes/W7_decedent_ladder.md section 8). Wave C corrected the headcount, which W7 named as the real lever, and the step is now $9.4B - through $33.4B and out the other side, a third of it rather than two and a half times it. The two published rows are also on different windows, which makes $33.4B an understatement of the step Treasury itself paid, so the overshoot is larger than $9.4B against $33.4B makes it look. See planning/lanes/HSC_h5_decedent_headcount.md section 8.
The earlier note here cited a -$456B companion target that Phase E retired; it appears in no Treasury volume, so there is no published-disagreement floor under this row.
Lane R2 moved this target onto the document its own record already described. -$960.0B was a rule of thumb ('1pp ~ $85-100B/year') in no JCT publication; the figure is now CBO publication 58164, Options for Reducing the Deficit: 2023 to 2032, Volume I, Option 13 alternative 1, 'Raise all tax rates on ordinary income by 1 percentage point', -$1,081.3B over FY2023-2032 (report p. 72), 'Data source: Staff of the Joint Committee on Taxation'. The prediction did not move - the model scores the same -$1,195.3B it did - so 24.5% to 10.5% is the target finding its document and nothing else.
MOST OF WHAT IS LEFT IS THE WINDOW, and the size is published rather than argued. The target covers FY2023-2032 and the runner opens its window in FY2025; CBO's own 2022 and 2024 editions price the identical reform at -$1,081.3B and -$1,185.3B, $104.0B apart (9.6%) for nothing but two years of base. FY2022_TARGET_WINDOW_RULE would close it with scoring_window_first_year=2023, which moves a model output and is therefore a .v3 the owner registers, not a provenance lane's to take.
Read this row beside cbo_opt45_all_rates_1pp rather than as an independent observation. They are the same reform in two Options volumes, scored on one window with one shape, so the pair measures the model's insensitivity to the decade (-$1,195.3B here against -$1,207.3B there, 1.0% apart, where CBO's two figures are 9.6% apart) rather than two separate predictions. A third edition exists and gives a third figure: CBO 2020 (pub. 56783) Option 1 alternative 1 prices it at -$884.0B over FY2021-2030.
Real bracket creep is still not modelled and pushes the other way, so part of the over-prediction against this target is a term the model is missing rather than one it added.
Scored on the SOI ordinary-income base with a single ETI (0.25). The base is now projected from its SOI tax year onto each year being scored, on the February 2024 vintage's own nominal path (1.0962x in FY2025 to 1.5438x in FY2034, 1.3118x on the window average), which took this row from 22.4% under to 1.9% over. This is the control for the base-growth lane: it is the row whose target is CBO's own published option on the very window being scored, and the growth term alone closes it.
What is left is real bracket creep, which pushes the other way and is not modelled: JCT's estimate rises through the window partly because rising income moves into higher rates, and a uniform index cannot reproduce that. Note also that illustrative_1pp_all scores the SAME reform against a different published figure, -$960.0B against -$1,185.3B, 23.5% apart; the model cannot agree with both.
This row declares bracket 1 of the statutory schedule and scores to the cent what it scored before, because bracket 1's floor is $0 in every year of every vintage. That is not filler: the declaration takes the whole schedule path end to end - four per-status floors read per year, deflated onto the SOI base year, scored through the filing-status split - and a $0 floor must come back out the other side unchanged. If this row ever moves by a cent, the schedule is perturbing something nobody asked it to touch.
The filing-status boundary is now the option's own: joint returns and surviving spouses face the 2025 24%-bracket floor of $206,700 and every other status $103,350 (IRS Rev. Proc. 2024-40 section 2.01, tables 1-4). That took the row from a 17.9% OVER-prediction to a 12.4% under-prediction - it removed 25.7% of the base, which was more than the gap.
The threshold now moves in 2026, which the option says it should. CBO's own text: 'Under both alternatives, the scheduled changes to the underlying tax brackets and rates would still take effect in 2026', after which the four highest brackets are 28/33/35/39.6 percent and the boundary is the 28% floor. The boundary is read per year and per filing status from CBO's own published schedule (publication 53724, June 2024 edition on this record's February 2024 vintage) - the table this note used to say does not exist. The reversion is not a uniform shift: bracket 4's joint floor falls 3.5% in CY2026 while its head-of-household floor rises 65.5% and its single floor 15.9%, so a scalar threshold could not have expressed it and neither could a scalar plus one joint amount.
The floors are statutory boundaries on TAXABLE income and SOI's size classes are by AGI, so the base is 'returns whose AGI clears the bracket floor' rather than 'taxable income above it'. That predates the filing-status lane and is unchanged by it.
The base is now projected onto the years being scored rather than held at its SOI tax year, which took the row from 12.4% under to 14.9% over - a registered regression, because it was under-predicting by less than a decade of nominal growth is worth.
Reading the schedule was a second registered regression and the plan's stated direction for it was backwards: the row over-predicts, so it went FURTHER OVER rather than further under. It decomposes into two terms pointing opposite ways, both of which the base projection created and neither of which it could resolve. Deflating the boundary into the SOI base year's dollars - the unit conversion that projection implies, since scaling the base by g is the same as indexing the threshold by g - is worth -$68.4B, because the statute indexes on chained CPI and the base grows on nominal GDP. The 2026 reversion is worth +$48.1B. The net is -$20.2B, 3.6 points of error. Applying the year's nominal boundary to a TY2023 income instead - comparing a 2031 threshold to a 2023 return - would score this row at 3.6%, and that figure is recorded here because it is what two errors cancelling looks like, not because it is available.
The $20,000 single / $40,000 joint threshold is now the option's own, applied per filing status against an SOI base split the same four ways (Table 1.2, TY2023). It removed the 46.1M joint returns' $20,000 of wrongly-taxed income - $839.8B, 9.2% of the base - and the row got WORSE, 44.7% to 49.8%, because that error was cancelling two larger ones. It is kept at the honest number rather than reverted.
The base is now projected from its SOI tax year onto each year being scored, on the February 2024 vintage's own nominal path (1.3118x on the window average), which took the row 49.8% to 34.1%. It is the largest single move this row has had and it is not enough on its own.
The base is now AGI, which is what the option says: 'a surtax of 1 percentage point would be imposed on AGI above $20,000 for single filers and $40,000 for joint filers'. SOI Table 1.1's rows are AGI size classes and it publishes both columns, so this branch had been subtracting an AGI threshold from an average of TAXABLE income - $92,658 minus $20,000, where the single-filer AGI average above that floor is $120,414. Reading the column the option states took the row 34.1% to 7.4% (AGI_BASE_RULE). It is applied INSIDE the filing-status split rather than to the pooled aggregate: the split ratio here is 1.4052 where the pooled one is 1.3820, and on the 2pp alternative the two move opposite ways.
That was the third of the three terms W7 finding 1 measured, and the three together land at 7.4% rather than the '9.1%' planning/HIGH_STAKES_ACCURACY.md section 1.3(c) quotes. W7 computed its endpoints by hand on a pre-projection tree; both the growth term and the AGI term were rebuilt as mechanisms afterwards, and a mechanism does not have to reproduce a hand calculation to be right.
The projection indexes the threshold with the base. Scaling the aggregate above a fixed nominal floor by a factor f is arithmetically the same as indexing that floor by f, so the term is a lower bound on the growth of an unindexed-threshold base and this row stays under for that reason too.
No behavioural distinction between a broad low-threshold surtax and a narrow high-income one: both erode by ETI x 0.5.
Same three residuals as the 1pp alternative, and this row is where they are visible: its 16.1% before the filing-status split was TWO ERRORS CANCELLING A THIRD. Starting 11.1M joint returns $100,000 below their own floor added 34% of base; a taxable-income base rather than AGI and a decade of no growth each removed about 22%. Splitting the floors alone therefore takes the row 16.1% to 37.4%, and the old number measured the cancellation rather than the fit - the same finding fra_2023_discretionary_caps produced when spend-out landed.
Both remaining terms are now in. The base is projected from its SOI tax year onto each scored year at 1.3118x on the window average (37.4% to 17.9%), and it is then read from SOI's AGI column, which is what the option says (17.9% to 2.9% OVER, crossing the target). All three terms were measured before any of them was built and each landed where it was pre-registered, so the 16.1% this row showed before the filing-status split is now decomposed rather than argued about.
At 2.9% this row is the closest non-spending case in the out-of-sample tier, and that is not evidence the AGI-inclusive family is solved: three of the six rows carrying agi_inclusive_base keep the taxable-income column because their own sources state it, or state a base that is neither SOI column, and they sit at 31.8%, 20.2% and 18.3%.
The thresholds are indexed after 2025 by the option's own text and the model's are not. The projection indexes them implicitly, at its own rate rather than the option's, which is a lower bound on the base of an unindexed floor and not a match for an indexed one.
The realizations base is projected across the window from its IRS SOI tax year at the accrued-gains stock's own growth rate (5.80%/yr, Federal Reserve DFA 1998-2024), on the identity that realizations are a flow off that stock at the observed hazard. Until Wave 5 it was held flat, which was most of a 44.8% under-prediction. The rate is a net-worth CAGR carried forward, not a realizations projection - CBO's own baseline has realizations returning toward a historical share of GDP - and tax year 2023 is a trough: the vendored aggregate reads $1,283.6B (2022), $943.4B (2023), $1,368.1B (2024). So the level anchor is probably low and the growth possibly high, in opposite directions, and neither is corrected.
The enactment year is scored at full strength against a fiscal year CBO scores at a fraction of one. The module's FY2025 rate channel is negative, because the enactment-year coefficient (8.06, the persistent 3.27 plus the transitory term on the timing-margin share) takes realizations down 14.9% in a single year; CBO's FY2025 is +$2.4B against its own +$8.3B for FY2026, which is what a calendar-year tax paid at filing looks like in fiscal-year receipts. This module has no receipts lag, and building one is a scoring-engine change that would move every row in the battery.
Inverting CBO's own published annual path is the check on the frozen elasticity rather than a limitation of it: on a projected base its out-years imply a semi-log coefficient of 3.0-3.2, against the frozen 3.2727 (Dowd, McClelland & Muthitacharoen 2015 at CRS R48562's 22% reference rate) and JCT's own working 3.1. On a flat base the same inversion gives a coefficient falling from 4.17 to 1.81 across the window, which no scorekeeper's method produces.
The +2pp does apply to the 0% bracket ($80.7B, 7.29% of the base), and that is the option's own design - it raises all three preferential rates. An earlier note here claimed the opposite and used it to explain an over-prediction; dropping the bracket moves the row further under, not closer.
The whole score is the death channel: the option changes no rate and states no per-donor exclusion, so it is the one row that tests the level of gains transferred at death rather than a design. The $54B constant an earlier note here described was deleted in Wave 2; the base is now Poterba & Weisbenner's flow carried as a share of household net worth and grown with the Financial Accounts stock, $196.2B in 2025.
It under-predicts by 35%, and every point of that was bought back deliberately. Wave 4 registered a regression from 8.4% when it stopped taxing charitable bequests and small decedents' housing gains, which no realization-at-death regime reaches; Wave 7 added a further point by integrating over a fitted size distribution, which shifts gains toward estates whose charitable share is 36% rather than 4%; and Wave C added fifteen more, pre-registered, by correcting the decedent count (below).
The decedent headcount was the coarsest thing left in the channel and is now a life-table rate: 3.38 million a year against roughly 3.09 million NCHS deaths, where it was 408,532 from Poterba & Weisbenner's *dollar* flow of estates over net worth used as a headcount rate. That took this row 20.3% -> 35.5% and is a registered regression, because the count and the LEVEL come from the same PW ratio and only the count was authorised to move: spreading an unchanged $196.2B flow over 8.3 times as many decedents shrinks the gain each one carries, so the carve-outs and the statutory rate ladder reach less of it. The level is the other half of the correction and would move this row back: held against the module's own death_exit_rate, the flow implies 0.372% of the accrued-gains stock where the stock's death exit is priced at 2.647%. How much of that factor of 7.1 is PW's inter-spousal exclusion and how much is the 1998 vintage is unmeasured - planning/lanes/HSC_h5_decedent_headcount.md SS7.
Grading the death rate by estate size was measured and refused rather than left unexamined. The wealthy are older, so they die at a HIGHER rate: the size-graded rate at the top of the wealth distribution is 2.84% against the 2.65% now used, so grading would raise the top count rather than lower it toward SOI's own return count. Both figures are emitted as CHECK-ONLY rows by scripts/build_capital_gains_data.py.
No lock-in unwind: constructive realization at death removes the incentive to hold appreciated assets, which raises lifetime realizations. The module models that channel only through an elasticity multiplier that a zero rate change leaves inert.
A third of the residual is the first fiscal year alone: the option takes effect in January 2025, so nine of FY2025's twelve months are inside it and the model books 0.75 of a year, where CBO's own FY2025 row is 0.48 of its FY2026 row. The volume's income-tax options run 0.77-0.85 and its other payroll options 0.29-0.31, so there is no stated convention to read off and 0.75 is the calendar rather than a fitted lag.
The rest is base growth. The model prices covered earnings off CBO's own February 2024 wage path, which grows 3.9%/yr; the base implied by CBO's published revenue row grows 3.45%/yr, so the model drifts from +3.2% in FY2026 to +7.2% in FY2034. Nothing in the option text says why CBO's base grows more slowly than CBO's wages.
The compensation-shifting response uses the repository's frozen ETI of 0.25 against CBO's own 31% economywide marginal rate on labor income. Neither was chosen against this target; an ETI of 0.40 would land the row at 0.9%.
Same base, incidence and shifting response as the 1% alternative, so the same three residuals apply. The model is very slightly convex in the rate (the shift grows with it) where CBO's row is very slightly concave, which is why the two errors differ by half a point rather than not at all.
Scored in the corporate module's derived mode, so the base is not the fitted profits aggregate. Since PR #121 it is CBO's own projected corporate receipts path (publication 59710 Table 1-1) converted to a statutory base at one ratio measured on completed history - IRS SOI Table 11's credit-realized TY2022 base over Treasury's actual FY2022 receipts, 4.80133, which is 1.0083 / 0.21 - and settled on IRC section 6655's estimated-payment calendar. That base is about a quarter larger than the fitted one in the window's first year, so the derived path over-predicts this row by MORE than the fitted path did, while reproducing the 21%->28% benchmark to 4.04%.
The residual is an ESTIMATOR difference and a SCOPE difference, not the 42% document gap an earlier version of this note claimed. Both halves are corrected by `planning/memos/CORPORATE_PER_POINT_YIELD.md`. (1) This option is a JCT estimate that CBO publishes - every corporate-rate option in every Options volume carries 'Data source: Staff of the Joint Committee on Taxation' - so the comparison with Treasury's Green Book row is JCT against Treasury OTA, not CBO against Treasury. (2) Treasury's row is not rate-only: since the FY2023 edition it carries the GILTI effective rate up with the statutory rate, and its own chapter says so. (3) Per-point dollars are not comparable across rate levels or scopes anyway; on the implied marginal base (|total| / step / 10, as a share of the vintage's own average base) the record is four estimators clustered and one high - Tax Foundation 55.1%, JCT 55.9%, PWBM 64.4%, Treasury (rate + GILTI) 79.5% - and this model reads 80.8% at +1pp and 76.1% at +7pp since PR #121 projected the base off the vintage's own receipts path, above all four at the small step and above three of the four at the large one. There is also NOTHING in the record supporting a per-point yield that rises with the step: where one estimator prices a big step and a small one, JCT's 14-point cut from 35% and its 1-point increase from 21% imply marginal bases of $963.2B and $963.0B.
Option 64 carries NO income-and-payroll-tax offset footnote, though the facing Option 63 does. An earlier version of this note called the individual-side interaction 'the largest single unmodelled channel'; JCT does not apply one here either, so that channel is refuted as an explanation of this row rather than merely unmeasured. Two channels that would push the derived score down are still named and not built, because each needs a number nobody publishes. Credit CARRYFORWARDS: section 38(c) and section 904(c) limits rise with the rate, so marginal absorption exceeds the average ratio the module applies - and CBO's 2018 Option 24, the only volume with a narrative, names exactly this channel as included in JCT's estimate ('An increase in the corporate tax rate would increase corporations' ability to use tax credits ... That use of credits would reduce revenues'). Lane R5 DID the acquisition the previous note called for and the stocks exist: IRS Publication 5108 (TY2022, PDF pp. 164 and 166) prints Form 3800's carryforward into 2022 at $91.16B (Part I line 4) plus $33.31B (Part II line 34), $124.47B against $72.17B of claims, and SOI's Corporate Foreign Tax Credit Table 1 prints a $174.44B section 904 limitation - 20.7% of its own $842.41B of foreign taxable income, so the mechanism is confirmed on published columns. What is STILL missing is the share of the base held by taxpayers in the constrained position: SOI's excess-credit and excess-limitation tables were published for TY2010 only and the series was discontinued, and the aggregate columns cannot identify the split (column 19 is limitation less credit by construction). So the channel is bounded rather than priced - see `corporate.credit_absorption_bounds` and `fiscal_model/data_files/corporate/credit_carryforward_stocks.csv`. The bound is informative: the section 904 channel at its own upper bound absorbs 28.85% of a marginal pre-credit dollar against the 29.15% the derived path already books by substituting the average credit ratio for the marginal one, so the section 38(c) channel is ENTIRELY unbooked and points this row DOWN. CAMT: for a book-minimum payer a point of regular rate raises nothing until the regular tax clears the minimum; it began in TY2023, after the last SOI year on file, and no source publishes the taxable income of CAMT-liable corporations (Treasury's JY2574 gives ~100 payers and a 2.6% counterfactual effective rate; JCT's 2022 letter to Chairman Wyden gives 175-200 corporations averaging $8.2B of book income in 2019 - neither is a base).
CBO's own loss-firm haircut was examined by lane R5 and REFUSED. `business-investment-model/source_code/Create_Tax_Data.prg:32-33` publishes `dmyrevnfc = 0.85` and `dmyrevx = 0.80`, derived from SOI 2005, and applying 0.85 would take this row to 22.8% - inside that lane's own pre-registered band. It is not applied because it multiplies a statutory RATE in a user-cost expression while this path multiplies a BASE that already nets loss firms: SOI's income subject to tax is net of a net-operating-loss deduction running 8.7-11.6% of the pre-NOL base against CBO's 12.8%, and CBO itself DIVIDES by the factor at `:172-175` wherever the other input already carries it, 'to avoid double-counting'. The pair is also not financial versus nonfinancial, as this repository's own plan said: 0.80 is 0.85 further reduced by the nonprofit share of nonresidential investment, and there is no sector split in the source. `tests/test_corporate_loss_firm_haircut.py` is the gate.
SOI's TY2022 base is inflated by two timing items that reverse - section 174 R&D capitalisation, which began that year, and the bonus-depreciation phase-down - and the memo sizes it: SOI's income subject to tax averaged 71.8% of NIPA pre-tax corporate profits over TY2019-2021 and 79.8% in TY2022, so TY2022's base is 11.1% above what the prior ratio implies. That is an upper bound, since it charges the whole ratio move to the timing items; deflating by it takes this row to about 46%, not to CBO's figure. The module anchors on the latest published complete report and does not adjust, and an earlier anchor year would score this row better - which is exactly why the year is fixed as 'latest published' rather than chosen.
A fourth, smaller item: the module's section 6655 first-year factor is 0.757, where JCT's options run 0.591-0.732 first-to-second-year and Treasury's FY2022/FY2023 rows run 0.593/0.601. Treasury's is a statutory blend the module has no concept of; JCT's is unexplained, because no volume mentions payment timing anywhere.
Foreign assistance is classified 'grants_and_procurement', a profile fitted on defence procurement options this battery does not score. What is left is the gap between that generic profile and this account's own speed.
Same generic grants profile as Option 37. National-service grants spend out a little slower than the profile implies, so the model still books slightly more inside the window than CBO does.
Pell spends out faster than the generic grants profile - CBO's own path is essentially complete in two years - so the model now defers past the window savings CBO books inside it. Closing this needs an account-level rate rather than an account-class one.
The target is the discretionary outlay total only; CBO reports a separate -$9.2B mandatory effect that this shape cannot represent.
A broad nondefense reduction scored on one grants profile; the option's real composition spans several account types with different speeds.
What remains is the level, not the lag. The 2026 budget authority (-$12.0B) is inflated by IIJA advance funding and by the option's 25%-then-50% schedule, so anchoring a constant level on it over-states every later year; CBO's own path drops to -$9.3B in 2027.
Infrastructure and block grants have the slowest spend-out in the battery (CBO's 2026 outlay saving is -$0.4B against -$12.0B of authority), which is why this case gained most from the spend-out model.
The excess share is now year-indexed (Wave 4 lane 3a), which is what CBO's own text specifies: the limit grows with a price index while premiums grow with health costs, so a widening slice of every premium sits above it. That took the row from 24.0% to 13.1% and the model's revenue growth from 4.0%/yr to 8.3%/yr against CBO's 14.4%/yr. The remaining shape gap is the two items below; neither is a parameter.
The base is premiums only. CBO caps 'the total amount of contributions for a worker's premiums and health spending accounts' (pub. 60557, report p. 66), and the repository's premium distribution has no flexible-spending, health reimbursement or health-savings-account dimension, so the dollars above the limit are understated by whatever those contributions add.
No plan-switching channel. CBO's own text names enrolment in lower-premium plans as the dominant behavioural response, which converts excluded premium into taxable wages; the module carries only a flat 0.2 elasticity. The option states the direction and publishes no magnitude, so a value for it would have to be fitted to this row.
The limit's indexation uses the baseline's own price path in place of the chained CPI-U the option names, because the repository carries no chained-CPI-U series. CBO projects both near 2.0% over 2028-2034.
CBO's statutory parameter schedule (publication 53724) covers CY2021-CY2034 on this record's vintage, so the option's CY2019 and CY2020 bracket floors are clamped to CY2021 and the loader says so. The record's written amounts are CBO's CY2021 figures for the same reason: the amount written is the amount actually applied in every year of the window.
As the four-bracket alternative: CY2019 and CY2020 are clamped to CY2021, the first year CBO's transcribed schedule covers on this vintage.
CBO's 2018 option raises the *basic HI rate* and says the increase 'would be evenly split between employers and employees'; the runnable payroll shape sets employer_share=0.0, which is CBO Option 61's own words for a different design ('the new tax would be paid entirely by employees'). The base and the rate are identical; the statutory incidence is not, and the compensation-shifting offset is applied to the employee side only.
The model's base is CBO's February 2024 covered-earnings path, so an FY2019-2028 window is priced on a wage path projected back past its own vintage - the residual grows the further the target's decade is from the model's anchor (20.1% here, 33.7% on the 2020 volume, 12.1% on the 2022 volume, 7.5% on the 2024 volume).
As the 1pp alternative: CBO splits the rate increase evenly between employers and employees while the runnable shape books it entirely on employees, and the covered-earnings base is projected back past its February 2024 vintage.
The module answers a 1pp corporate rate increase with nearly the same figure on every decade: -$192.3B, -$191.8B, -$192.9B and -$196.1B across the four Options volumes, a 2.2% spread, where CBO's own estimates for the identical reform span 41.0% ($96.3B to $135.7B). The residual on any one of these rows is therefore mostly a level (the implied marginal share of the statutory base, 80.8% against a published 55.1-79.5%), not a decade. See planning/memos/CORPORATE_PER_POINT_YIELD.md and lane R5/H3b.
The residual on every edition of this reform is one number, not an edition-specific mechanism. The derived path reaches about 80.8% of the credit-realized statutory base its receipts path implies (CBO receipts x 4.80133, after the 17.6% profit-shifting offset and the IRC section 6655 phase), where JCT's own implied marginal share is 52.6% on the April 2018 baseline, 57.1% on May 2022 and 55.9% on February 2024 - and 66.1% on the COVID-depressed September 2020 one. `0.808 / share - 1` reproduces each row's error to within a tenth of a point, which is what makes the four residuals four readings of one gap. The channel that points it downward and is entirely unbooked is IRC section 38(c): a general-business-credit carryforward stock of $124.47B against $72.17B of claims, under a statutory cap that RISES with the rate. The share of the base held by taxpayers who are both capped and holding stock is published for no post-2010 year (SOI's excess-position tables stop at TY2010), so nothing here multiplies one. See planning/memos/CORPORATE_PER_POINT_YIELD.md and planning/lanes/CORP_class_accuracy.md section 2.2.
Two causes, both stated. CBO's 2020 edition prices this reform *below* its 2018 edition ($203.3B against $222.9B) on the pandemic-depressed September 2020 baseline, which a base growing monotonically with CBO's February 2024 nominal path cannot reproduce. And the boundary is read from CBO's own statutory schedule, which begins in CY2021 on this vintage, so the option's own CY2021 floors are applied for the whole window rather than the schedule's later re-indexation being anticipated.
The 2020 volume is scored on CBO's September 2020 baseline, whose wage path is depressed by the pandemic: CBO prices this reform *below* its own 2018 edition ($877.5B against $898.3B) though the decade is two years later, and five of the six options this battery repeats across those two volumes fall the same way. The model's base is CBO's February 2024 wage path times the Trustees' covered-earnings ratio and grows monotonically, so it cannot reproduce a baseline that went backwards.
As the 1pp alternative: the 2020 volume's pandemic-depressed wage path is below the 2018 volume's for the same reform two years later, and the model's monotonic February 2024 wage path cannot reproduce that.
This row's receipts path is CBO's September 2020 Outlook (publication 56517, Table 1), which is the baseline JCT's estimate was priced against and is also a COVID outlier: it projects FY2021 corporate receipts at $122.8B against a Treasury MTS actual of $371.8B. So this row's error is the module's usual marginal-share level measured against a depressed denominator, not a more accurate score, and it must not be quoted without that clause. The correct path is still this one - JCT scored a projection, not realized receipts.
The same reform as cbo_opt45_top4_brackets_2pp on the previous decade, and it carries that row's limitation: the option's own text says 'the scheduled changes to the underlying tax brackets and rates would still take effect in 2026', which a single statutory bracket index cannot express. The pair is a measurement of the model's sensitivity to the decade (23.8% on FY2023-2032 against 17.9% on FY2025-2034), not two independent predictions.
SpendingPolicy carries one annual level grown at 2%/yr; CBO's own path for the WEP/GPO repeal grows at about 1.1%/yr after the first full year, so the model drifts above the published path across the window.
NOT a spend-out miss, and the spend-out model confirms that rather than closing it: classified 'mandatory_benefit', 99.8% of the authority outlays inside the window and the residual is unchanged. Benefit payments are outlaid in the year they are owed, so the whole miss is the growth rate.
The FY2025 retroactive catch-up CBO describes ($25.0B, against a $19.7B steady state) is outside the level shape entirely; the model neither reproduces the spike nor is credited with it.
What remains is the level shape, not the spend-out. CBO's caps compound against a falling funding base and reach -$159.7B by 2033, while a level grown at 2%/yr reaches only about -$134B, so the model under-predicts for that reason alone.
This case got WORSE when spend-out was added, and that is the correct outcome. Its old ~6% total was a cancellation: the model over-predicted the early years (CBO's 2024 outlay saving is -$64.1B against -$112.3B of budget authority) and under-predicted the late ones. Spend-out removes the first error and leaves the second, so a truer path shows a larger total error. The old number measured the cancellation, not the fit.
Only the caps component is scored. The bill's -$1.5T headline also bundles the $45B Toxic Exposures Fund appropriation, student-loan payment resumption, an IRS rescission and debt service.
Neither a spend-out, a level-shape nor a window miss since iija_2021_discretionary.v3. The shape is the source's own authorization schedule ($163.0B of budget authority in FY2022, then $70.1B, $68.5B, $68.1B, $66.2B and $2.08B/yr, summing to CBO's stated $446.3B), spent out on the construction_and_capital profile, and scored on FY2022-2031 - the decade CBO's own estimate covers and the record's budget_window has stated since it was entered. It returns +$414.3B against +$415.4B, 0.3%.
What is left is a 4.5% over-statement of the TOTAL netting against the tail the window still clips, and neither term is behavioural. Outlays across every year the policy touches are $434.1B, which is the profile's 0.9727 spend-out sum applied to the full $446.3B of authority; $19.8B of them fall in FY2032 or later, outside even this window. The two nearly cancel, so 0.3% is smaller than either term and should not be read as evidence about the spend-out profile. The honest statement is that the authority path is CBO's own and the profile reproduces its total to 4.5%. (Earlier revisions of this note said $433.2B and 4.3%; the measured total is $434.1B.)
CBO's own table is headed FY2021-2031, eleven fiscal years, and a ten-year window cannot cover eleven. FY2021 is not a gap: the bill was signed on 15 November 2021, inside FY2022, and the record's own budget_window has read FY2022-2031 since it was entered - which is what FY2022_TARGET_WINDOW_RULE reads, so the window is not a per-case judgement about which eleventh year to drop.
It needed a WINDOW, not a VINTAGE, and an earlier version of this note named the wrong blocker. A discretionary SpendingPolicy scores its own source-stated authority and reads no baseline LEVEL, so the CBOScore.scoring_window_first_year mechanism PR #126 built for treasury_capgains_39_plus_stepup_elim scores this bill on FY2022-2031 with no 2021 vintage in the repository. On the runner's FY2025-2034 decade v2 read 18.2%, because $92.6B of the path's outlays fall in FY2022-2024 before that window opens. The number was published in planning/memos/FY2022_TARGET_WINDOW.md section 6 before the decision was taken, so that taking it would be a visible choice.
This row does NOT claim a 2021 information set. The outlay profile is the one fitted on CBO's own donor options (lane L2), and the years FY2022-2024 are priced with it today rather than forecast in 2021. What the window fixes is that the ten fiscal years scored are the ten the target covers.
The superseded rows are kept in preregistered.py: v1 (a level carried forward at 2%/yr) at +$1,894B / 356% and post-spend-out +$1,621B / 290%, and v2 (the authorization path on the runner's window) at +$340.0B / 18.2%. Between them the three rows separate the three defects this case surfaced - the missing spend-out model (L2), the missing authorization path (v2) and the window (v3).