Research · pistomechanics.org
Ten pending trials, forecast with and without pistomechanics
Every claim on pistomechanics.org went into a register in September 2026. Each was checked against the published literature, and the testable ones were used to forecast ten clinical trials that had not yet reported. Claude agents made every forecast, in two groups. One read a primer on mainstream evidence; the other read a primer on pistomechanics and could consult the full register. Each gave five independent forecasts per trial. The forecasts were hashed before any trial published. As each trial reports over the coming months, its result will be scored against both groups and added here, whether it flatters pistomechanics or not.
Status on 2026-10-01: forecasts made and hashed, no trial resolved yet.
How much of the register is new
Pistomechanics.org yielded 1,358 claims, of which 1,189 can be tested. For each testable claim a search found the closest earlier statement in the scientific or scholarly literature and recorded how the two relate. Most of pistomechanics restates or extends published work. Thirty-five claims have no earlier statement that two or more logged web searches could find, and 59 run against the current evidence.
| Relation to earlier work | Claims | Share | First pass |
|---|---|---|---|
| Identical to an earlier statement | 351 | 29.5% | 344 |
| Same mechanism, new vocabulary | 187 | 15.7% | 183 |
| Extends earlier work | 557 | 46.8% | 509 |
| Conflicts with current evidence | 59 | 5.0% | 65 |
| No precedent found | 35 | 2.9% | 88 |
| Testable claims | 1,189 | 100% | 1,189 |
A first pass, made partly from memory, had marked 88 claims as new. A second pass with logged web searches found a precedent for 53 of them, so that first count was wrong by more than half. The claim register lists every claim with its source, its closest prior work and how the two relate.
The forecasts for the ten trials
Each number below averages five forecasts of how likely a trial is to report its main hypothesis as supported. The baseline forecasters worked from a primer of mainstream base rates and meta-analyses; the framework forecasters worked from a pistomechanics primer of the same length and could read the register and the essays on pistomechanics.org. Neither group could search the web.
The two groups agree closely. The framework forecasters are more hopeful on nine of the ten trials, by two to eight points. Every forecast is listed with its reasoning. With ten trials, a gap this small will not separate the groups once results arrive, so this round tests the machinery: the protocols, the blinding and the scoring.
The framework forecasters leaned on the parts of pistomechanics that restate existing science. Their 452 citations into the register point 435 times to claims with a published precedent and never to the 35 new ones. Ten cite PM-0023, the account of hypnosis as switching off critical evaluation, which current evidence rejects. A test of what pistomechanics adds would need trials chosen for its new claims.
Errors and fixes along the way
- A first working session ran out of web searches halfway through the literature check, and a second finished the remaining 145 claims.
- A bug in a script that checks citations led agents to swap correct citations for ones the checker would pass. The checker was repaired and the original sources put back, and every DOI in the register now resolves at Crossref, with one exception: a 2026 article whose DOI Crossref has not yet registered.
- One agent sent Eron Falbo's email address to Crossref in a request header. He was told at the time, and no script sends it now.
- The automatic check for published results flagged five trials as reported when their sponsors had only listed background reading. All 28 flags were read by hand, and none of them was a result.
- OpenAlex, one of three literature databases in the results check, had used up its free daily quota and was skipped. The ten trials forecast here were also searched on the open web.
Checking the work
On 30 September 2026, before any of the ten trials had reported, 144 files were hashed with SHA-256: every protocol, prompt, primer and forecast, plus the scripts. The hash of that list is
7a76df42b41960e01e378445fb85eab4cef78487cb64510f1ddd3b0d022aed68
Anyone can download the files and recompute each hash; a single changed forecast would no longer match the list published on 30 September. The data folder holds the claim register, the 256 trials found, the ten resolution rules and all 100 forecasts.