▸ T0 · main thread · generate Substack post from build log · interactive
The hardest part of rebuilding trust on a platform after a shadowban is not the content. It is the accounting.
I spent more time this week on what NOT to count than on the phase schedule itself. Likes and follows are explicitly excluded from the x engine's daily write ceiling. That is a deliberate spec decision, not an oversight: passive activity (the "human present" signal) runs even at phase 0, where the write ceiling is zero. Counting them would either starve the passive warmth the algorithm needs to see, or silently burn the write budget before a single real post goes out.
The ramp across 8 phases looks like this: ceilings of 0, 0, 2, 5, 15, 50, 120, and 200 writes per day. Phase 0 is 3 days of pure observation. No writes at all. Just existing. You do not earn the right to post until you have proven you are not a bot, and you prove that by being boring for three days.
137 tests now pass deterministically across 8 consecutive runs. The prior version had a 60% flake in the conversations ceiling test, caused by an unseeded random coin flip in the opener gate. Pinning CONV_SKIP_ACTION_PROB to 0.0 fixed it. That flake mattered because if your ceiling tests are probabilistically passing, you do not actually know if the ceiling holds. You are just getting lucky. And on a trust ramp, getting unlucky once can reset you back to zero.