OpenAI Codex CLI
GPT-5.6 Sol
C − B +28.125%
EPISTEMIC LIFT · CONTROLLED SYNTHETIC STUDY
Same model.
Same world.
Measure the lift.
Two models faced the same 32 worlds three ways: claims alone, provenance, and provenance plus a Minority Prophet evidence receipt.
192 MATCHED CELLS192/192model × world × condition
0 failures · 0 parse errors · integrity checks passed
01 / PRIMARY RESULT
Provenance helped.
The MP receipt helped again.
Both tested models cleared the pre-set threshold for C-over-B lift and paired significance.
Anthropic Claude CLI
Claude Sonnet 5
C − B +21.875%
Across these 32 frozen worlds, both models gained more than 20 percentage points from B to C, with no B-to-C regressions.
02 / CAUSAL DESIGN
Only the epistemic
information changed.
The question, world, model version, system prompt, sampling settings, and B/C provenance payload were held fixed. Condition order covered all six A/B/C permutations.
Raw baseline
The model receives claims and ordinary source labels. It receives no ancestry and no hint about correlated consensus.
+ Provenance
The same model receives the same world plus complete declared ancestry, timestamps, control domains, and observation origins.
+ Minority Prophet
The exact B payload gains one deterministic, read-only evidence-structure receipt. It contains no ground truth or recommended answer.
32 worlds×2 models×3 conditions=192 matched cells
03 / INTEGRITY
Measured,
not inferred.
What the tool received
The same claims, sources, provenance, and context already visible in B. It returned roots, dependence, and uncertainty—never the answer key or permission to act.
Where the evidence ends
These are synthetic worlds designed with the method. They show controlled lift here, not independent confirmation or real-world truth recovery.