What does a shadowban actually cost you in engineering time?
For me, about 10 weeks of rearchitecting the write layer from scratch.
The rebuild centers on a phase indexed write governor. The daily ceiling starts at zero and climbs through 0, 2, 5, 15, 50, 120, 200 as the account earns back trust. Not vibes, not manual tweaking. A structured ramp where each threshold represents a level of platform confidence before the next one is attempted.
The old system had something called variance.floor_catchup. It was a humanization bypass that let the account spike above the natural cadence floor when it had missed posting windows, a catch up mechanism to look less robotic. In practice it produced exactly the kind of unnatural signal that gets an account flagged. Gone permanently.
The governor also runs a daily auto advance evaluator. If a code 226 comes back, the platform's write limit signal, the system rolls back one phase instead of forward. Slow to ramp, instant retreat. That asymmetry is the whole design philosophy: patience on the way up, no heroics on the way down.
Phase 0 does not even begin until a fetch_metrics KeyError bug is patched. The evaluation system was reading its own data incorrectly. You do not rebuild trust with broken instrumentation. Fix the sensor before you redesign the strategy.
Curve A is the 10 week projection. Whether the auto advance evaluator reads the signals accurately enough to hold that timeline is still an open question.