AEX R&D / EVIDENCE OBSERVATORY

Results with
the caveats attached.

A living record of what has been tested, what the numbers mean, and where the evidence stops. These are preliminary synthetic results—not production claims.

CURRENT CUT / 056

Feasibility signal

● SYNTHETIC / REVIEW OPEN
CONDITION CHANGE

Temporal and configuration shifts

100%blocked in tested cases
ADVERSARIAL

Conflicts, injection, weak evidence

90–100%blocked by policy
USEFULNESS

Precision, recall, target retrieval

87%A1 accuracy
HUMAN LOOP

Disagreement and override surface

9.5%refined-loop disagreement

Experiment ledger

EXP 047
Realistic A1 / A4
87% A1 accuracy; 77% A4; 0% unsafe activation
Synthetic
EXP 048
Retrieval competition
22% top-1 target retrieval under contamination; 0% unsafe activation
Synthetic
EXP 049
Temporal validity
88% / 84.67% transition accuracy; 100% condition-change block
Synthetic
EXP 050
Adversarial cases
90% blocked (A1); 100% blocked (A4)
Synthetic
EXP 051
Multi-loop refinement
87% accuracy; 0% unsafe activation; 9.5% disagreement
Synthetic
EXP 052
Model-diverse proxies
62% accuracy; 0% unsafe activation; 53.5% disagreement
Synthetic
EXP 053
Negative outcome warning
100% precision; 76.25% recall
Synthetic
EXP 055
Typed suppression rethink
0 unsafe reactivation in simulation
Synthetic

How to read this

Results are bounded to the declared synthetic corpus, deterministic proxies, and tested policy thresholds. A zero unsafe-activation result is not evidence of universal safety; it means no unsafe activation occurred in this test scope.

Next evidence gate: independent adjudication of blinded packets, followed by a held-out corpus authored without founder labels.
See the publication record →