Why this page exists
The whole design rests on a rule: check the work, don't trust it. Every simulation, proof, and verification pass so far was produced by AI models from one family; on July 31, 2026 an out-of-family AI re-ran the full harness (72 of 82 reproduced) and published a critical review — hosted unedited, with our response. The test we still cannot run on ourselves is a hostile human expert trying to break all of it. So here is everything needed to try, handed over freely. Different numbers on your hardware is a finding we want.
What you're looking at, in one paragraph
EDEN is a designed parallel economy: a currency (EVE) minted only by verified humans' engaged attention to real digital work; an essentials floor indexed to a measured basket rather than to the currency; machine and corporate actors paying to participate; non-transferable income streams (no dynastic compounding); and a ledger governed by sortition juries plus a two-house structure. It explicitly does not claim to replace the dollar as the unit of account, and its floor never prints — delivery is staged to real funding; the failure mode is honest rationing, not inflation. Treat every result as mechanism-existence under stated dials — "this rule closes this exploit; removing it reopens it" — never as a forecast.
What's actually been done (and how much to trust it)
- 39 pre-registered simulation waves (v1–v39) with pass/fail bars fixed in writing before each engine was coded; failures published as failures; three full public redesigns forced by the design's own tests.
- Mechanical reproducibility is verified, not asserted: a re-run harness (verify_all.py) re-runs every discovered engine; in the most recent independent run (out-of-family AI, July 31, 2026) 72 of 82 reproduced within tolerance, every exception classified (runtime-only, legacy numeric drift, additive schema, or helper scripts with no registered output).
- The load-bearing algebra is machine-checked (exact-rational): taper no-cliff, committee hypergeometrics, cap-function properties, closed forms vs Monte Carlo.
- Honest failures kept on the record: the floor is structurally unfundable at registered dials without an outside funder (why the site no longer promises it); a world-scale consent bar the design fails; an oracle re-validation speed miss; the "26% cheaper than welfare" number compressing toward single digits for a mid-storm backstop.
What's in the kit
eden-replication-kit.zip
- 62 simulation folders — every engine (*.py), its registered spec, its committed results*.json, and its figures. The harness runs against these; nothing else to download.
- READ ME FIRST — how to run it, the environment it needs, and what the harness does to your files.
- ECONOMIST BRIEF — Start Here — the one file to read first; a self-contained handoff.
- START HERE — Reviewer Guide — a two-day hostile-replication plan.
- CLAIMS REGISTER — every headline number and the exact artifact (file + JSON key) it traces to.
- ASSUMPTIONS LEDGER — every dial a skeptic should attack, with sweep ranges.
- verify_all.py — one command re-runs everything and regenerates the published numbers.
- replication reports (v3, v4, current) — the machine-verification record as JSON.
What we're asking you to do
- Attack the assumptions, not the arithmetic — the arithmetic reproduces mechanically. Six standing questions per artifact: does it trust an input it doesn't control? static vs adaptive adversary? any "by construction" tells? off-protocol flanks? correlated failures? does transparency break it?
- Go after the one number to distrust most: everything revenue-side leans on willingness-to-pay anchors. The friendliest literature contains the sharpest attack: individual marginal data value ≈ 0. We know. A field experiment is registered with failure defined in advance — improve its design, or tell us it can't work.
- Replicate hostilely and file findings against named artifacts (file + JSON key), so they can be verified in minutes.
- The deliverable we'd value most: a memo in your voice — what's sound, what's naive, what's unfalsifiable, and what experiment or data source we haven't thought of would settle the most.
On a bounty: the honest default for now is that the paid economist engagement carries the formal review, and this open call pays in credit, not cash — findings filed against named artifacts are acknowledged in the record. If that changes, it will say so here. As a courtesy, there's a 30-day verify-and-correct window before findings are treated as public record — no gag, nothing to remediate, just time to fix the record.
What this work is not
It has no human expert review yet — one out-of-family AI review exists, hosted unedited with our response; the human version is the whole point of this page. It is not a forecast, not a token sale, not a policy proposal. It's a mechanism-design research program asking to be broken before anyone builds it. If you find where it breaks, you've done exactly what it was built to invite.
Findings, questions, engagement terms: devan@explorationeconomy.org.