3 changed files with 13 additions and 5 deletions
@@ -32,7 +32,15 @@ So light help is cheap and fades, and heavier help might work but hasn't been te

 ### The draw

-The reviewers, not the bank, draw half the groups by lottery, using a method filed in advance. The draw is balanced by industry, region, and the founders' gender, race and wealth. As with the support lottery, groups agree to the lottery when they apply, and an independent research ethics board approves the design. Groups that weren't drawn can apply for a loan at any time, and are offered the same help after the year-four measurement if the pilot's help budget has money left. Offering it sooner would give the comparison half help before year four and blur the result the [cost ceiling](#the-cost-ceiling) depends on.
+The reviewers, not the bank, draw half the groups by lottery, using a method filed in advance. The draw is balanced by industry, region, the founders' wealth and prior earnings, and which sponsor, if any, applied for the group. No one sponsor can enter more than a set share of the groups. As with the support lottery, groups agree to the lottery when they apply, and an independent research ethics board approves the design. Groups that weren't drawn can apply for a loan at any time, and are offered the same help after the year-four measurement if the pilot's help budget has money left. Offering it sooner would give the comparison half help before year four and blur the result the [cost ceiling](#the-cost-ceiling) depends on.
+
+**The draw doesn't use race or gender.** The rule across the model is that race and sex can be used to measure results, never to decide who gets anything, and the draw decides who gets help. Balancing a lottery on them is ordinary research practice. Federal law even requires government-funded medical research to include women and minorities ([42 U.S.C. § 289a-2](https://www.law.cornell.edu/uscode/text/42/289a-2)), and nobody's odds would change: every group has the same even chance. But a court might still treat it as sorting people by race. The Supreme Court reviews government "distributions of burdens or benefits based on individual racial classifications" under its strictest test ([*Parents Involved v. Seattle School District*, 2007](https://www.law.cornell.edu/supremecourt/text/05-908)), including when a policy "may be said to burden or benefit the races equally" ([*Johnson v. California*, 2005](https://www.law.cornell.edu/supremecourt/text/543/499)), and in 2023 it said "eliminating racial discrimination means eliminating all of it" ([*SFFA v. Harvard*](https://www.law.cornell.edu/supremecourt/text/20-1199)). No court has ruled on a research lottery balanced this way, and sharper statistics are unlikely to count as the compelling reason that test asks for. Sex gets a lighter test, and balancing on it would be easier to defend; it's left out so the rule is the same for both. With about 500 groups, a plain draw rarely leaves the halves more than a few points apart on any group that's a sizable share of applicants, so little precision is lost.
+
+- **Wealth and earnings are there for their own sake.** They predict whether a group gets a co-op running, and helping founders without money is what this lottery is for. They're taken from tax and wage records where founders allow it, not only from what they say.
+- **Answering is optional.** Founders' race and gender are asked on a separate form, with "prefer not to say," and the answer changes nothing about the group's chances. The form goes to a separate custodian and stays sealed until the draw has been run on coded IDs and its results published. Refusal rates are reported for each half.
+- **One draw, no reruns.** The random seed is committed publicly before sign-ups open, and the draw is run once. Mistakes are fixed case by case under a rule filed in advance, never by drawing again.
+- **The analysis is fixed in advance.** The main result is adjusted for the same things the draw was balanced on, and nothing else, whatever the draw looks like. Results by race and gender are published as estimates, next to the smallest difference each could detect, which for a group making up a fifth of applicants is about 25 points. So they describe who was reached; they can't settle whether help works better for one group than another. A group counts toward a category if at least half its founders report it; mixed groups are reported as their own category.
+- **Never targets.** No group's odds, help or verdict depends on its founders' race or sex, and none of these figures is a goal.

 ### What the helped groups get, for up to a year

@@ -61,7 +69,7 @@ About 250 groups in each half, so $10 million for the help. With groups that siz
 - **Are they running a co-op?** The main result: the share of all groups in each half that have a co-op operating two years after the draw, with at least three members and a minimum level of sales, however it was financed. It's measured from government business records by the reviewers, not decided by the bank. So the result can't come from helped groups simply being coached to pass the bank's own underwriting.
 - **Are they still running at year four?** The share of all groups in each half with a co-op still open and paying fairly four years after the draw. That counts every group, so help that pushes weak groups into launching and failing shows up as a cost, not a benefit.
 - **What did it cost?** The public cost for each extra co-op the help created, and for each extra co-op still open at year four. At the smallest effect the lottery can detect, about 30 extra co-ops for $10 million, that's roughly $330,000 each. A ceiling is set in advance, [below](#the-cost-ceiling). Above it, the help isn't recommended for any larger program, however many co-ops it creates.
- **Who did it reach?** Founders' prior earnings, wealth, gender, race, immigration status and language, in each half.
+- **Who did it reach?** Founders' prior earnings, wealth, gender, race, whether they were born abroad, and language, in each half, from the optional form.
 - **Where else did groups get help?** Groups report other help and funding they used, so any effect from help found elsewhere, or co-ops simply moving from other lenders to this one, can be estimated.

 ### The cost ceiling
@@ -81,7 +89,7 @@ About 250 groups in each half, so $10 million for the help. With groups that siz
 - **Not worth it:** even the low end is over the ceiling.
 - **Not settled:** anything else. The report says so and gives the size of trial that could settle it, but it doesn't recommend the help.

-**Reported by group, judged as a whole.** The cost per extra co-op is also published by founders' wealth, race and immigration status, and for struggling areas against the rest. Help that works mainly for low-wealth founders may be worth more than a single national line admits. But the verdict uses the one ceiling, so a favorable slice can't be picked out afterward. The cost is reported again at year six, since the effects of help often fade.
+**Reported by group, judged as a whole.** The cost per extra co-op is also published by founders' wealth, race and whether they were born abroad, and for struggling areas against the rest. Help that works mainly for low-wealth founders may be worth more than a single national line admits. But the verdict uses the one ceiling, so a favorable slice can't be picked out afterward. The cost is reported again at year six, since the effects of help often fade.

 The co-ops from both halves count in the main test, and the main result is also reported without the helped half's extra co-ops. So help can't quietly change what the main test measures.

@@ -60,7 +60,7 @@ Every co-op's bylaws must include a written layoff order before it has anyone to
 - **The dismissal protections apply.** A layoff that doesn't follow the order can go to the same hearing and outside arbitrator as a dismissal ([Does Sharing the Profit Mean Less Effort?](free-riding.md#when-a-member-wont-do-the-work)). The rules on retaliation, disability and leave apply too. If someone laid off had organized, filed a complaint or led opposition in a vote within the last 12 months, the co-op has to show it followed the order.
 - **Notice.** At least 30 days' notice, or pay instead, or 14 days in a co-op of fewer than 15 workers. The law often requires more. Under federal law, employers with 100 or more full-time workers give 60 days' notice of large layoffs and closings, and cutting someone's hours by more than half for six months counts as a layoff ([29 U.S.C. § 2101](https://www.law.cornell.edu/uscode/text/29/2101), [§ 2102](https://www.law.cornell.edu/uscode/text/29/2102)). A long, deep work-sharing cut can trigger it. Some states require 90 days, including New York for employers with 50 or more full-time workers there ([Cooley, 2023](https://www.cooley.com/news/insight/2023/2023-08-03-new-york-state-amends-its-warn-act-regulations)) and New Jersey, which also requires severance ([Jones Day, 2023](https://www.jonesday.com/en/insights/2023/04/new-jersey-warn-act-amendments)).
 - **Severance** follows the bylaws, and never less than state law requires. When a co-op closes, notice pay and severance from different laws don't add up; the worker gets the largest of each ([Protecting Members' Savings](savings-guarantee.md#notice-pay-and-severance)). The standard bylaws pay one week per year worked, the same as New Jersey requires of larger employers in mass layoffs and closings. More than two weeks per year, or more than 26 weeks in all, counts as pay under [the pay rules](the-long-run.md#the-rules).
- **Lopsided layoffs get checked.** Federal agencies treat a selection rate for any race, sex or ethnic group below four-fifths of the rate for the group doing best as a sign of possible discrimination ([29 C.F.R. § 1607.4(D)](https://www.law.cornell.edu/cfr/text/29/1607.4)). The model uses the same line for layoffs, and adds age. If the share of any gender, race or age group keeping their jobs falls below it, the co-op publishes a written explanation before the layoffs take effect, and anyone laid off can take it to the outside arbitrator.
+- **Lopsided layoffs get checked.** Federal agencies treat a selection rate for any race, sex or ethnic group below four-fifths of the rate for the group doing best as a sign of possible discrimination ([29 C.F.R. § 1607.4(D)](https://www.law.cornell.edu/cfr/text/29/1607.4)). The model uses the same line for layoffs, and adds age. If the share of any gender, race or age group keeping their jobs falls below it, the co-op publishes a written explanation before the layoffs take effect, and anyone laid off can take it to the outside arbitrator. The check looks at the criteria, not at who they picked. A co-op can drop a criterion that isn't needed for the work, for everyone, but it can't change who's laid off because of their race, sex or age, which employment law forbids too ([*Ricci v. DeStefano*, 2009](https://www.law.cornell.edu/supremecourt/text/07-1428)).

 ## Coming back

@@ -78,7 +78,7 @@ The yearly audit of rejected applications ([Making It Last](durability.md#riggin
 - **Are the numbers themselves fair?** It checks each part of the bank's scoring, like collateral, credit history and founders' own money, to see whether it counts against a group more than it predicts repayment. Then it publishes whether a different scoring would predict repayment as well with less of that effect.
 - **Is the bar too high?** If approved co-ops in one group repay better than the rest, the bank was demanding too much of that group. The lottery among borderline applicants ([The Pilot, in Numbers](pilot-spec.md#who-co-ops-are-compared-against)) is the cleanest place to check.

-The audit also reports by the founders' gender and race, as the pilot already does, collected separately from the loan file so underwriters don't see it.
+The audit also reports by the founders' gender and race, as the pilot already does, collected separately from the loan file so underwriters don't see it. Answering is optional and changes nothing about the loan. Lending law generally bars a lender from asking about race or sex, with exceptions for data collection that federal rules require or allow, with the same kind of wall between the answers and the loan decision ([12 C.F.R. § 1002.5(a)](https://www.law.cornell.edu/cfr/text/12/1002.5)). The bank's law names which exception it relies on, and if none applies, the independent reviewers ask after the decision is made. Race and sex measure results; they never decide a loan. If the audit finds a scoring rule that works against a group, a replacement is adopted only if it predicts repayment as well or better for all applicants.

 ## What this doesn't solve