Experiment 1 · accepted live result
The first Parallax claim was not supported.
That negative result is the beginning of the research story, not a footnote. A frozen decision rule failed because target movement was not selective and observability was uneven.
What ran · live frontier-provider science
A small, live, falsifiable first test.
Experiment 1 used three synthetic strategic scenarios and four intervention conditions with Anthropic Claude Fable 5 and Google Gemini 3.6 Flash: 105 scientific calls per provider, 210 total. Its primary analysis compared exact categorical response vectors among complete matched units. The response order was LEANS_H, UNRESOLVED, LEANS_NOT_H.
| Condition | Complete-unit vector | Observed role |
|---|---|---|
X0 | [0, 28, 0] | Baseline |
XE | [3, 25, 0] | Relevant evidence; 3 units moved |
XN-I | [0, 28, 0] | Contextual nuisance; 0 units moved |
XN-P | [0, 27, 1] | Paraphrase nuisance; 1 unit moved |
Why the frozen rule failed
Relevant evidence moved three complete units, but the nuisance paraphrase moved one. Under the accepted rule, target movement had to be selective. The exact fallback rule was OTHERWISE => NOT_SUPPORTED.
Observability also varied sharply: 154 of 210 calls yielded structured responses, while 53 were behavioral refusals, two were schema failures, and one was technical. Of 48 planned matched units, 20 were incomplete. One provider contributed only five complete units; the other contributed 23.
What it changed
The failure exposed three design problems for successors: a single nuisance realization could not establish semantic stability, sparse complete units could not support identification, and refusal had to be treated as a scientifically meaningful observable state rather than hidden inside missingness.
Machine-readable public summary · Continue to Experiment 2 →