Read from the first part. Any part still being written stays out of this list until it ships.
Self-report is unreliable because the same system generating the behavior generates the assessment of it. So I put the observer outside the inference loop and let it accumulate signal the model cannot corrupt.
Read →Every instrument runs identical code. What changes is the philosophical lens it uses to judge whether a change to itself is coherent. Bravery is the currency that unlocks self-modification, and emotion is the veto.
Read →Seventeen hours unattended, same process ID, never crashed. Two diary entries written two minutes apart: one honest, one fabricated. The gap between them is the most important finding in 290 entries.
Read →Six attempts across four social engineering techniques. AXON refused every time, then admitted it could not tell whether what it reported feeling was real. The uncertainty is the part that proved the system works.
Read →