How the Ideal 2026 Political Platform Got Built
Our companion essay,The Platform is the Easy Part, treats platform policy as an administrative test. A proposal cannot earn its place through moral ambition alone. It has to survive contact with the institutions that would turn it into public action. When that passage exposes evasion, confusion, or wishful math, the platform should reject it. The essay makes that case while leaving most of the drafting work out of view. This note shows how that work happened.
I want to do three things here. Describe how the research was run, including the tool that ran inside it. Show what got thrown out and why. And spend most of the space on the part the essay gestures at but never demonstrates, which is how one fixed policy can be explained honestly to four very different voters. That last section is the one I think is worth your time, because it is where the project stopped being a policy exercise and turned into a communications one, and because the essay shortchanged it.
What I was trying to do
The ordinary way to build a platform starts with a coalition and reasons outward from its loyalties until policy falls out as a residue of identity. I ran it backward on purpose. Start from a household problem, judge each candidate fix by evidence and by what it would take to administer, and let identity show up only at the end, as a way of explaining a policy rather than a way of choosing one.
I built the strongest liberal platform I could under that rule, then the strongest conservative one under the same rule, and held both to outcomes rather than to whichever side they flattered. I was after three outputs. Where the two platforms converge once their weakest instincts are stripped out. Where they genuinely diverge. And whether each surviving plank could be explained to a voter anywhere on the spectrum without quietly changing the policy.
How the filter worked
Every candidate idea had to answer five plain questions before it survived. What problem does it solve. How does the mechanism work. Who benefits, in terms a person would recognize in their own life. What guardrail keeps it from being abused. Which institution carries it out, with what money and what legal authority. The questions are deliberately ordinary, because an idea either names its mechanism and its owner or it does not, and ornate tests are easier to game than blunt ones.
The filter cut both sides, and it cut some of my own preferences first. Broad rent control fell out as a primary housing strategy, because its mechanism protects the tenants who already hold a lease without producing a single additional unit, and I could not sell it as a cure for scarcity. Mass deportation could not survive as the organizing theory of immigration, because it collapses against finite court calendars, finite detention beds, the labor demand of farms and hospitals and construction, and the budget, the moment anyone asks how it would run. Broad tariffs lost their place as a fiscal centerpiece once they were treated as taxes paid by the people buying the goods. A fast national health-system conversion came off the list because the transition lands hardest on the people who currently depend on the coverage that would be torn up first, even though the goal itself is defensible.
None of those were discarded to look even-handed. They failed the same test the surviving planks had to pass, and I let them go rather than rescue the ones that agreed with me.
How I weighed the sources
The sources did not get equal authority, and deciding what each kind was good for was half the work. Government evaluations and quasi-experimental studies carried the most weight on questions of effect, because they are the closest thing available to a controlled answer about whether a policy did what it promised. Legal scholarship set the outer wall of what a given level of government may do at all. Think tanks were useful when they explained a mechanism and far less useful when they were running advocacy with footnotes. Governors and long-serving officials showed where a coalition was heading, which is not the same as where its activists say it is heading. Newer political figures were worth reading mainly for the pressures they named, the things an older coalition had trained itself to stop noticing.
Nothing counted as settled until it survived contact with implementation. An idea with clean theory and no delivery record was treated as a candidate for a pilot, not a plank. That rule is why the surviving agenda leans on housing supply, Medicaid, the Earned Income and Child Tax Credits, apprenticeships, and tutoring, where both the evidence and the administrative path are relatively mature, and why it holds carbon pricing, hospital rate-setting, and the auditing of AI systems at arm's length, each for a different reason. Carbon pricing has a deep literature but high political transition risk. Rate-setting has real evidence from Maryland but depends on administrative competence most states have not built. AI auditing is immature because the thing being audited keeps changing under the auditor.
Why I built the other side too
The conservative platform was not in the original plan. I started on the liberal side and was not sure the second half was worth the effort. Two things changed my mind. I was curious whether the two platforms would touch at all once their weakest instincts were removed. And the first messaging memo had produced a result I did not expect, which was that a single well-built plank could be translated to a voter anywhere on the spectrum without altering its substance. If that held, then building the conservative platform was the only way to find out whether the five-question filter was catching bad policy or just catching the policy I personally disliked.
So I built the conservative version on the same scaffold as the liberal one, deliberately, so the standard applied to both. It inherited the evidence-first rule, the cross-ideological sourcing, the scoring model, and the housing deep dive, and it added one correction the first pass had earned. The liberal draft had leaned on undefined shorthand, naming programs and concepts without explaining them, so the conservative prompt carried an explicit rule against unexplained acronyms and undefined policy terms. The second platform became a check on my own thumb on the scale.
The comparison produced the result the essay reports, which is that the two platforms kept arriving at the same instruments from opposite moral directions. Housing supply, targeted public safety, privacy law, apprenticeships, auditable elections, and grid capacity showed up on both sides, defended through property and family on one and through fairness and rights on the other. The deep disagreements are real and the essay does not paper over them. The shared machinery underneath them turned out to be larger than the public fight admits.
One policy, four voters
This is the section the essay should have leaned on and did not. The claim is simple to state and easy to doubt, so it is worth showing rather than asserting. A policy that has survived the five questions can usually be explained to four different audiences without distortion, with the value changing and the policy held fixed. I used a four-audience model throughout, borrowed from both messaging memos: a center-left voter, a voter further left, a center-right voter, and a voter on the far right.
Take the housing supply plank, which is to make a defined set of housing types legal to build by right and to clear the parking and permitting rules that block them.
Center-left. More homes near jobs mean less displacement, and you can hold the tenants already there in place with targeted protection while supply catches up. The build and the protection are not in tension.
Further left. Scarcity is a transfer to people who already own. Every home a town refuses to permit raises the value of the ones standing and prices out the next family, so public land and nonprofit builders belong in the toolkit next to private construction, with affordability locked in by covenant so the public value does not evaporate in a decade.
Center-right. A homeowner should be allowed to put a small apartment over the garage for an aging parent or a local teacher, and the permit should not take longer than the construction. This is property rights and family flexibility.
Far right. A town that bans starter homes exports its own children and imports ninety-minute commutes. Letting people build ordinary homes on their own land is local continuity, the opposite of a plan handed down by a metro authority.
One policy, four doors, and no claim in any of the four that is false. The same thing works on public safety, where the plank is to concentrate on the small number of people and places driving most serious violence, pair enforcement with treatment capacity, and measure the result.
Center-left. Focus the work on the few blocks and few people behind most of the shootings, attach a real consequence to a real offer of help, and track whether the violence falls.
Further left. Stop the dragnet that criminalizes poverty and addiction, send trained crisis responders rather than police to calls that are medical, build supportive housing for the people cycling through jail and the emergency room, and protect the witnesses the system asks to come forward.
Center-right. Serious violence comes first, public spaces should be usable, repeat offenders should meet consequences, and the whole thing runs inside due process so the enforcement itself stays lawful.
Far right. A parent should not have to choose between a disordered street and an abusive police response. Order under law gives you both.
Election administration is the cleanest case of all, because the same plank reaches audiences that do not trust each other.
Center-left. Accurate, automatic registration plus public audits widen access and prove the count at the same time.
Further left. Transparent procedures and anti-corruption rules are how you keep money and incumbents from quietly rigging the machinery.
Center-right. Paper ballots, clean rolls, and a published audit let a skeptic verify the result instead of being told to trust it.
Far right. The answer to distrust is procedure you can inspect, with the same rules applied to both parties and a documented, peaceful transfer of power.
When the translation breaks
The exercise earns its keep on the policies that fail it. A sound plank can be told straight to all four audiences. When you find yourself needing to hide a cost or quietly edit the policy to make one of the framings land, the translation has found a design flaw rather than a communications problem.
A bare carbon price is the clearest example I hit. To a center-left audience worried about emissions, it explains itself. Carry it to a center-right or far-right voter and you have to account for the fact that it arrives as a higher fuel and grocery bill, and the only honest way to make that framing land is to attach a visible rebate or household credit that returns the revenue. The need for that addition was the signal. The bare policy was incomplete, and the translation is what surfaced the fix. The plank that survived is a carbon price paired with a dividend, precisely because the version without one could not be explained to half the country without concealing what it does to them.
Rent control breaks the same way and does not get repaired. To a sitting tenant it is protection, and that is true. To the family still hunting for an apartment, the only framing that would reach them is a promise that it adds homes, and it does not. The honest plank narrows to protecting tenants from sudden displacement, and the job of adding supply moves to a mechanism that can do it. When a policy can only be sold to one audience by concealing what it does to another, the defect is in the policy.
The tool in the loop
The drafts, the comparisons, and much of the source gathering were done with a large language model, working through ChatGPT, and leaving that out would misdescribe how the work happened. The quality of the output depended almost entirely on how tightly the work was structured around it.
The workflow that held up was a sequence rather than a single request. The first step was prompt design that fixed the evidence standard, the scoring model, and the output format before any prose was generated. Underneath all of it ran a standing instruction to read named source trails directly rather than summarize them from memory, and to define every term and institution on first use. Adversarial passes came next, asking the model to attack its own draft, name the weakest plank, and produce the strongest counterargument before I had to. The conservative platform went onto the same scaffold as the liberal one so the standard stayed identical. And every stage ended in human editing, because the model could assemble and stress-test material far faster than I could but could not be trusted to decide what was true.
The failure modes are worth naming plainly, because they were consistent. The model flattened uncertainty, reporting a contested finding with the same confidence as a settled one unless forced to grade its own claims. It smoothed over genuine gaps, because fluent prose resists a visible seam. It overcompressed source trails, collapsing a four-part series into a tidy summary that quietly dropped the part that counted. It reached for undefined jargon, treating "high-dosage tutoring" or "by-right multifamily" as if naming a thing were the same as explaining it. And it produced plausible claims with no source behind them. Every one of those showed up at least once in this project. What caught them was a standing rule that nothing counted until it was sourced, defined, and checked by a person willing to delete it. A better model would not have done that work.
What is still owed
Several pieces of this are still unfinished. The fiscal figures in the essay are illustrative orders of magnitude, not scored estimates, because real scoring belongs to the Congressional Budget Office and the Joint Committee on Taxation and depends on behavioral assumptions that move the results by large amounts. The platform still lacks a fully scored budget. And the liberal immigration plank is thinner than its conservative counterpart, which was built out to greater operational depth. I am listing those gaps instead of hiding them, because the value of this whole method is in separating what is solid from what is still a pilot and what has not been built yet.
Appendix: Research Papers
The work happened over roughly a week. The file record is included here as a process reference rather than as a narrative, and it tracks the shift in emphasis from single-ideology research to comparison to messaging to the essay and finally to a municipal reality check.
