v4.22

v4.22 (Sep 28, 2026) releases the policy-testbed round: sessions 1–11 of Next Round Plan: Simulation Testbed and Claims Ledger (Sep 26). Each session's notes follow as a "v4.22, session N" section, newest first. The engine in index.html is unchanged: the seed-42 regression and every preset's figures are bit-identical (validate passes). What a visitor sees does change: the page opens with a comparison of Compassionism against five other anti-poverty designs. unit passes (11 tests, 8 price-module tests, 4 labor-module tests, 6 session 5 tests and 16 testbed tests, 1 of them new); domtest passes all 94 checks (2 new). Hand-off: session-10-handoff.md.

What the release contains

Plan item	Where it lives	Sessions
A0 flow map: where program money enters, moves and leaves	A0-flow-map.md (project files)	1
A1 issuance ledger	node harness.js ledger	1
A2 money-creation and price module, breakeven additionality, price-neutral point	PRICE; node harness.js price	2, 3, 5
A3 labor supply: one income effect per unconditional dollar (ρ 0.16), one compensated elasticity (0.33)	LABOR; node harness.js labor	4, 5
A4 testbed: basic income, negative income tax, asset endowment, public grocery, X-Cents, one cost ledger, tax / money / hybrid financing	node harness.js testbed … match / frontier / alts / decomp	6, 7
A5 reporting: FGT₀–₂ at year 20 and over 20 years, spells, cost per point, prices, hours, income Gini, group losses in dollars	node harness.js testbed … a5	8
A6 front door on the page	index.html; node harness.js frontdoor	8, 9, 10
The replication page moves into this repository	replication.html; domtest Phase 10	7

Every harness addition sits behind switches whose defaults leave the engine bit-identical, and validate, unit and domtest run on every push.

Reading the comparison

• Same people, same rules. Every design runs on the same 500 simulated single adults for 20 years with paired random seeds, is paid for by a flat contribution on wages set each year to cover its cost, and meets the same work response. The basic income, negative income tax and asset endowment are matched to Compassionism's gross cost per adult-year; the public grocery network and X-Cents run at their own cost.
• The shaded row. Wherever Compassionism appears, the row beneath it switches off its two mechanisms that are theoretical, yet to be empirically tested, and cost nothing in the model (both the model's stand-ins, not the design; session 11): the octave wage raise (d19) and the PTF/PTH inflation damping (d33). Across the round these carried most of Compassionism's edge (see round-close-out-and-next-round-plan.md in the project files).
• The headline measure is poverty severity (FGT₂ on the living-wage basket) averaged over all 20 years, with year 20 beside it (d37); year 20 alone overstated Compassionism's gains.
• Pre-computed, on the testbed's profile. The table is data written by node harness.js frontdoor from testbed … a5, not a live run (d36). The testbed's profile differs from the live scenarios (θ on realised PTF density, PTH cuts housing only, discounts skip the basket's tax share, wages indexed to prices, the work response, the design-neutral BLEI gate, and N7_BLEI), so the two sets of figures differ. The table's caption says so.
• Not modelled, for any design: public costs of poverty (d29), a production side for conversion rewards (H1), households and children.

Session 11: what changed (Duke's review of the prepared page)

• The opening line (d41, Duke's option 5). "Compassionism aims to end extreme poverty while incentivizing participation and contribution: test that claim against five other designs on the same simulated people." It is a statement, so Phase 11's first check now accepts one sentence ending in a full stop or a question mark. The subtitle adds that poverty is measured against a living-wage basket, a higher bar than extreme poverty, with the deepest shortfalls weighted most (FGT₂), so the claim is tested on a stricter line than the one it names. README.md opens with the same sentence.
• Wording. Everything a reader sees now calls the octave wage raise and the PTF/PTH inflation damping "theoretical, yet to be empirically tested", and says both are the model's stand-ins, not the design (d46). Earlier session notes below keep the word they used ("unsourced").
• The design panel. Compassionism's entry is rewritten from Duke's notes: BU as a restricted currency that expires; conversion at elevated rates set by market demand and by community-validated quality (the multiplier rate); the octave as conversion capacity, a safeguard against exploitation and an open ceiling for creators with demand; wage work still paid in dollars, with creative projects able to hire at elevated rates paid in expired BU; an allowance that can vary by region, age and labor-market needs (the model holds it flat). A second paragraph says what sets the design apart: it creates both a supply of a new restricted currency and a demand for it, and aims to hold basic living costs steady through community-owned supply before letting prices fall, where current policy manages inflation through the money supply and interest rates (d47: stated by mechanism). The negative income tax entry no longer says the allowance is the same for every participant by design. The grocery and X-Cents entries are rewritten (below).
• A needs-based top-up (d39, d44). PR.ccoTop(G, s, t): Compassionism with its flat BU cut by a share s, and the savings paid as extra BU to participants only, max(0, G − t × wage earnings), t = 0.5 (nit.part, read by tbGate, tbLab and tbCashFlow; unset everywhere else, so every other row is bit-identical). tbMatch('ccoTop', …) sets G so the total cost equals Compassionism's. A 30-seed sweep (testbed 30 topup ref,adv,st) tried s = 0.1, 0.25 and 0.5 and a 0.3 phase-out: only s = 0.1 deepened the 20-year FGT₂ cut in all four reference and Adverse cells (the Stress Test was added afterwards), so a5 and the page use it. At full seeds (a5):

Environment, model	Compassionism: FGT₂ 20-yr change / share in poverty / hours	With the top-up (same cost)
Reference, engine	−3.21 / 45.4% / −4.9%	−3.38 / 47.4% / −7.2%
Reference, Hub	−1.76 / 53.1% / −14.5%	−1.87 / 60.4% / −21.7%
Adverse, engine	−7.67 / 74.4% / −7.5%	−7.76 / 75.5% / −10.2%
Adverse, Hub	−4.18 / 83.9% / −25.8%	−5.30 / 87.3% / −30.6%
Stress Test, engine	−3.71 / 76.1% / −3.5%	−3.70 / 76.3% / −4.5%
Stress Test, Hub	−3.58 / 76.5% / −5.5%	−3.53 / 78.1% / −8.6%

The top-up deepens the cut in four of six cells and leaves it about the same (within 0.05) in the Stress Test; it raises the share in poverty and lowers hours in all six, because a benefit that shrinks as earnings rise works like a tax on work. It is shown as its own row beneath Compassionism's mechanisms-off row, with its own mechanisms-off figure in the note, rather than replacing the main row (d44).

• Grocery at its real scale. The Mayor puts New York's existing grocery stores at more than 1,000 (nyc.gov transcript, Jul 27, 2026), so five stores of the same size would serve about 0.5% of shoppers (TB_GROC_PILOT, derived; the 9,000 sq ft La Marqueta store makes it if anything high). The page shows the plan as announced (five stores: about $4 per adult-year, no measurable change) and the network at full scale (every adult 15% off food: $807, −0.04 at reference). Neither models the effect on other grocers' prices.
• X-Cents from its paper (d48). The paper was read in full, and the page links https://bettertobest.github.io/x-cents/. PR.xcFull(f, h) adds the Power of 1 to the adult exchange: one designated day a week, coins count as $1 for food, so food bought with pennies costs 1 cent per dollar (TB_XC_P1 0.99), with the week's food shopping moved to that day (f = 1; f = 1/7 if no shopping moves). The main row covers food (reference: $8,017 per adult-year, −1.17; exchange alone $3,613, −0.56; f = 1/7 $4,243, −0.65). Food and housing (groc.housing) is a sensitivity row, since paying rent and utilities in coins needs landlords and utilities to join: $21,514, −2.00 at reference, and +0.24 in Adverse, where its contribution cuts hours by 37%.
• Each design at its proposers' size (d45). The page gains a second view. Basic income $12,000 a year (TB_UBI_PROP, the $1,000 a month paid in Vivalt et al.'s study); a negative income tax guaranteeing the 2026 HHS poverty guideline for one person, $15,960, with a 50% phase-out (TB_NIT_PROP); the stakeholder grant, $80,000 of 1999 × 321.943 / 166.6 (CPI-U annual averages) = $154,594, once, to adults under $25,000 (TB_ENDOW_PROP; no return in the model). Reference: $11,999, −1.67; $1,607, −0.09; $2,922, −0.44. Because designs of different sizes are not a fair contest by themselves, every costed row in both views now shows its change in 20-year FGT₂ per $1,000 (reference, engine: Compassionism −0.28, with the top-up −0.30, without the two theoretical mechanisms −0.16; matched basic income −0.14, matched NIT −0.20; X-Cents −0.15; basic income at $12,000 −0.14; stakeholder grant −0.15).
• Link and layout review (Duke's screenshots). The row labels, the "What each design is" link and the caveats box's "Assumptions and Known Limitations" link were same-page fragment links (href="#…"). They work on GitHub Pages, but claude.ai's file preview routes every link click as a URL on claude.ai, so they opened a new chat. They are now buttons (.fd-go, with data-go naming the target), which no host can send off the page; the page has no fragment links left, and Phase 11 checks that. All 49 external links on the page, the replication page and the README were requested: each resolves (the six that answer 403 are publishers and sites that refuse scripted requests; the two DOIs redirect to the right papers). Below 1,180 px the caveats box stood taller than its text: the base .fd-side rule came after the narrow-screen rule and overrode its direction, so each box took its 300 px flex-basis as a height; fixed. The replication page's section links stay fragment links (d49): correct on the live site, but they misroute in claude.ai's preview.
• Worse off, and participation as a choice (d50). The column compares each group with itself under no program, and its losses come from the contribution that pays for each design. Under Compassionism they fall mainly on adults who do not take part, who pay the contribution and receive no BU. The column now names them "adults who chose not to take part", its tooltip and the table's caption say so, and the design panel states that participation is open to every adult. It also states the model's limit: each adult's choice is set at the start and kept for 20 years, so an adult who would gain by joining cannot.
• Checks. unit: two new testbed tests (the top-up at G = 0 equals Compassionism at the smaller BU, and pays participants only; Power of 1 at f = 0 equals the flat exchange, housing raises the cut, five stores cost a small fraction of the full network). domtest Phase 11: the statement check, both views in the data and in rendering (24 views), and one new check (the view switch moves the comparators and not Compassionism, every costed row states its change per $1,000, X-Cents links its site, and no visible text calls the mechanisms unsourced): 95 checks. The engine in index.html is unchanged; validate passes.
• Reproduce. node harness.js testbed 500 a5 ref --json=a5.json, node harness.js testbed 200 a5 adv,st --json=a5.json (about 9 minutes each), node harness.js frontdoor a5.json; the sweep, node harness.js testbed 30 topup ref,adv,st --pilot=20 (about 4 minutes).

Session 10: what changed

• Tooltips (dashboard i4, and Duke's review). Every tooltip was a CSS ::after box drawn inside the element it explains. Inside the front door's table, which scrolls sideways on narrow screens, the scroll box clipped the top of every column tooltip, so the first lines were hidden behind the column labels. The same boxes, though invisible, widened the page to 421 px on a 390 px phone. One floating box placed by script (initTips, position:fixed) now serves every tip: it cannot be clipped, it opens below its host when there is no room above, it stays 8 px inside the window, it opens on hover, tap and keyboard focus (not on a mouse-click focus, as v4.20 intended; a tapped tip stays open until the next tap, since phones send mouse events around a tap), and it closes on leave, blur, scroll, resize and Escape. Checked on a touch-phone viewport as well. The CSS boxes remain as a fallback when scripts do not run. Checked in a headless Chromium: the column tooltips show in full, and the page is 390 px wide at 390 px. domtest checks the behaviour (it has no layout engine); decision d42.
• What each design is. A panel under the comparison with one entry per design: the proposal, how the model simplifies it, and its source, linked. Every row label links to its entry, and the "Compassionism as" switch has its own ⓘ explaining the engine and Hub versions. Sources: Vivalt et al., NBER Working Paper 32719 (basic income; also the source of ρ); Moffitt (2003), Journal of Economic Perspectives 17(3) (negative income tax); Ackerman and Alstott, The Stakeholder Society (1999), and Hamilton and Darity (2010), Review of Black Political Economy 37 (asset endowment); NYC Mayor's Office, July 27, 2026 (public grocery; the 30% and 15% are the city's projections, not results); the CurrentSea X-Change paper on Academia.edu (X-Cents). domtest Phase 11 checks every entry, every source link and every row link.
• The front door's question (d41, following d35): "Dollar for dollar, which does more against poverty: Compassionism, a basic income, or a negative income tax?" README.md opens with the same sentence, as Phase 11 requires. The subtitle now says which designs are matched to Compassionism's cost and which run at their own.
• N7_BLEI in the testbed profile (d40). TB_PROFILE_G turns it on for every testbed row that does not set it (the n7 section still sets it both ways); the global default stays false, so the checks still test index.html. The front door was regenerated (testbed 500 a5 ref, testbed 200 a5 adv,st, then frontdoor). Against session 9's figures, engine-model rows move by 0.06 point of 20-year FGT₂ or less (0.16 at year 20) and hours by 0.2 point or less; the largest move is the framework-scale basic income at reference, 0.15 point over 20 years and 1.18 at year 20, in the fiscal-breakdown regime (an 89% contribution). No conclusion changes. At reference, per $1,000 of cost, 20-year FGT₂ cut: Compassionism (engine) 0.28, without its two unsourced mechanisms 0.16, matched basic income 0.14, matched negative income tax 0.20.
• Metadata (d43). The page's description had last been updated at v4.15; it and the sharing descriptions now say what the page compares. The replication page's JSON-LD cited the simulation as v4.17; now v4.22.
• Version labels. META.VERSION 4.22; the page's static labels, title and JSON-LD; the replication page's eight labels (Phase 10), its highlights, and a v4.22 panel in each Version History list; the "Unreleased" headings below renamed for this release.