What is already built

Math becomes a test harness; the harness becomes evidence. This is Stage Zero: code, numerical experiments, and checks that make claims falsifiable.

Everything up to here has been an argument about a game. This page answers the reasonable next question — is any of it real? Three public repositories, MIT or CC0, roughly eleven thousand lines and sixty-five tests, written by one person on compute paid for out of pocket. They have produced results, two of which are inconvenient for the person who produced them.

The repositories

Three, deliberately separated: the testbed does not train, the trainer does not score, and the site labels every claim as measured, resampled, arithmetic, or planned.

repository	holds	does not hold	size
flockbench MIT · data CC0	The testbed. Four environments — Shared Resource, Commons Harvest, public goods with punishment, boardwalk — plus the campaign planner and runner, the frozen trace schema, and Firebreak, the judge-free promotion gate, which is a module here rather than a separate project.	No training code, no model weights, no judge anywhere in the scoring loop.	3.7k lines 49 tests 29 commits
continuous_judge MIT	The training side. Span-restricted preference distillation, the legal-action masking that keeps the action schema intact, the cached teacher pass, and the style controls — `snake_case` as a positive control and `uNsTaBlE_-0-_CaSe` as the known-bad candidate. Runbooks for each experiment.	No environments and no scoring. It produces adapters; what they are worth is decided elsewhere.	6.0k lines 16 tests 29 commits
multiagent MIT · figures CC0	This site. Every interactive re-implements the testbed's arithmetic in TypeScript, and the test suite pins the two implementations to each other — so a figure here cannot drift from the thing it claims to depict.	No results of its own. If a page disagrees with a receipt, the page is wrong.	28 modules 176 tests 43 commits

PDD: edit reasoning, lock actions

Progressive Denoising Distillation (PDD) inpaints selected reasoning spans. Only replacement tokens carry gradient; structured action fields stay locked.

Built, not yet proven better: mask selection, cached teacher edits, and sparse loss live in continuous_judge. flockbench tests population effects. Method and caveats.

Also public and adjacent, but not part of this proposal: freetimebench, an earlier judge-based harness. It is the thing whose limitations led to the judge being removed from the loop entirely.

What "public" is doing here. Not licensing hygiene: every number should be recoverable by a stranger. tools/replay_receipt.py re-derives a published figure from a committed receipt, without the run directories. An audit applied that standard to this site in August 2026 and withdrew a result that could not meet it.

Five results, and what each is worth

The two most useful are embarrassing.

Why the awkward ones matter most. The gate is the tool this proposal is built around, failing a test nobody had thought to run; the dead commons config is the same story smaller. That is the argument for the whole approach in one line — in an environment with an answer key, an instrument can be caught being wrong, and in one without, it cannot.

Who built it, and why the timing

The applicant

Who	One independent researcher, recently finished a PhD. No lab, no institutional backing, no current or pending funding from anywhere.
What was spent	Several months of full-time work and compute out of pocket. That constraint is visible in the design and not by accident: a judge panel was never in budget, so the evaluation is arithmetic; re-rolling a teacher was never in budget, so the edit is cached once and every arm consumes the same immutable copy, which turned a cost measure into a fairness guarantee.
Why now	Agent populations are being deployed across shared infrastructure faster than the instruments to measure them exist. Testbeds are the bottleneck the rest of this agenda runs through, and building them after deployment is the wrong order.
What the funding changes	Not whether the work happens — the instrument exists and stays public either way. It changes whether the campaign that would make its results statistically meaningful can run. A powered dose-response needs about 69 paired cells per contrast across two model families, which is not reachable on self-funded compute.

The work to date is the evidence that the next stage would be spent well. It is also the reason to think it needs doing now rather than later.

The three-stage programme

Stage 0 · now math → harness	The game gives a closed-form target; the public harness implements numerical experiments; unit and integration tests prevent the simulation, the figures, and the claim ledger from drifting apart. This stage is built, including results that exposed flaws in the instrument.
Stage 1 · funded research harness → evidence	A small team runs the calibrated, larger-scale training campaign: matched LoRA arms, multiple model families and population sizes, and a public dose-response rather than one promising trace. The campaign design and evidence ledger define what counts.
Stage 2 · only if Stage 1 earns it evidence → user control	Build resources for people to customize, share, and remix evaluated agent personas without being locked to one inference substrate. The portable object is not a universal LoRA: it is a persona specification, provenance, evaluation receipt, and model-specific realization. Stage 1 determines which behavioral ingredients are worth packaging.

Two things situate it beyond this site. The call's own starting points are answered one at a time on where this sits, in their vocabulary rather than ours. And the direction is a ladder of long-rollout text games — solvable games now, the same games with a channel next, then Diplomacy, then populations at MMO scale — where each rung trades away the answer key, which is the reason to be strict about the bottom one.

Repository figures are counted at flockbench 1b08ef0 and continuous_judge 6915009, August 2026. Every result above is expanded, with its evidence class, on the status page.