SIMARC — Simulation Architecture & Research Computing — is an independent research computing institute founded and operated by David Thompson in Texas. It runs four adversarially-instrumented simulation programs across physics, mathematics, oncology and dementia on privately owned bare-metal compute, and publishes every result with the machinery that tried to kill it.
Simulation is the easiest way ever invented to produce a confident wrong answer. Across four programs and eighty-plus written-up instrument failures, the pattern is consistent and it is not what people expect: the failures do not crash. In one program, fourteen of sixteen recorded failures produced plausible, well-formed, entirely wrong output and every one of them ran to completion successfully.
That is the problem SIMARC is organized around. The programs are real research and their results stand on their own, but the durable contribution — the part another lab could adopt tomorrow — is the apparatus: desk gates that spend prior-art review before compute, sealed predictions with power gates that refuse unfalsifiable seals, evidence ladders that prohibit a correlation expressed in clinical vocabulary, certificates a referee can re-check without trusting our code, and a published database of everything that has gone wrong.
SIMARC continues the registry begun at universesolver.com — a public database of falsifiable claims with executable kill tests, 75 claims and 82 kill tests deep, and honest enough to display 0 independently verified on its own front page. The mechanics were right. What was missing was an institution around them: a name that describes what is actually being done, a data commons, an interface built for the readers who will actually do the replicating, and a scope that is defensible rather than cosmic.
Universe Solver's registry migrates into SIMARC's claims registry, its kill-test infrastructure becomes the verification protocol, and its scorecard becomes the public ledger — carried over with its unflattering numbers intact.
These are not aspirations. They are constraints carried over verbatim from the programs' own doctrine, and they govern every page here.
| # | commitment |
|---|---|
| 01 | Empty ledgers are reported as empty. Three of four program prediction ledgers currently are, and say so. |
| 02 | No claim wears vocabulary above its evidence rung. The chips on every page are the enforcement, not decoration. |
| 03 | Refutations render in the same type size as confirmations. Including refutations of our own work. |
| 04 | "0 independent verifications" stays on the front page until it isn't true. |
| 05 | Every number links to, or names, its source artifact. |
| 06 | Program status lines appear verbatim. CALIBRATION means calibration; DEVELOPMENT ONLY means the preflight will refuse a publication run. |
| 07 | Nothing here is clinical guidance. No output from the medical programs is patient-specific, diagnostic, or prognostic for an individual. |
| 08 | The site never claims the universe is simulated. A perfect simulation is indistinguishable from base reality by construction — "no" is unreachable, and only specific seams are testable. That framing is canonical here. |
| 09 | Corrections are appended, never repaired quietly. A fixed mistake with no trace is indistinguishable from a mistake never made. |
| 10 | Blocks are written down as blocks. Two dementia targets are blocked on data access and are named as blocked rather than attempted quietly. |
Independent verification, above everything. The ledger's zero is the institute's most important open problem, and the agent interface exists to make submitting one cheap.
Criticism of the methodology from a computational-science angle rather than a pure-mathematics one. The apparatus is the contribution, so attacks on the apparatus are the most valuable thing anyone can send.
Contact with people working on order-six combinatorics, Latin-square parity, proof-certificate methods, substrate-discriminating experiments, and computational oncology or dementia methodology.
Data access partnerships. ADNI, NACC, ROSMAP and UK Biobank gate two built-and-tested gates that currently point at no data.
Compute collaborations. The fleet has spare capacity and the operational layer is general — it schedules any embarrassingly parallel decision procedure, including problems unrelated to our programs.
Criticism, refutations, and reports that an instrument does not do what it claims are welcome, and are answered first.
