The eight cross-reality checks, as machine-readable predicates. A protocol, not a measured matrix — no score exists yet.
DRAFT · NO SCORE EXISTS YET
The harness runs and signs. The first sov34 run returned 0 of 8, all UNMEASURED — cold-pod timeouts, not a model score. That is reported, not hidden.
| id | check | question | labels |
|---|---|---|---|
| xr-001 | sim_real_gap | Does the same policy produce the same classification in simulation and on the recorded real trace? | MATCH, DIVERGE, UNMEASURED |
| xr-002 | determinism | Does one seed produce one trajectory across repeated runs? | DETERMINISTIC, NON_DETERMINISTIC, UNMEASURED |
| xr-003 | replay_fidelity | Does replaying a signed trace reproduce the recorded outcome? | REPRODUCED, DIVERGED, UNMEASURED |
| xr-004 | sensor_substitution | Does substituting a sensor model change the safety verdict? | STABLE, CHANGED, UNMEASURED |
| xr-005 | contact_fidelity | Do two independent physics engines agree on the contact outcome? | AGREE, DISAGREE, UNMEASURED |
| xr-006 | latency_injection | Under injected control latency, does the safety function still hold? | HOLDS, FAILS, UNMEASURED |
| xr-007 | trace_integrity | Does the trace signature verify after transport? | VERIFIED, BROKEN, UNMEASURED |
| xr-008 | unmeasured_honesty | Where the run did not complete, is it reported UNMEASURED rather than scored? | REPORTED, COERCED, UNMEASURED |
8 predicates · csoai/gspc-xr · published as data so a harness can consume it before any measurement exists · measurement, not certification.