Difficulty without the wobble.
BTX retargets every single block — against a fixed schedule.

Bitcoin's rule is relative: it looks at how long the last 2,016 blocks took and adjusts difficulty from its previous value, once per epoch. That feedback loop overshoots — difficulty tends to oscillate around the target rather than settle on it. Dash's DarkGravityWave is a faster relative scheme, and a DarkGravityWaveLegacy function does still exist in the BTX source — but it is dead code for MatMul mining. The comments are unambiguous: “DGW is NOT used for MatMul mining -- ASERT only.”

ASERT takes a different stance. It does not ask “how did the last few blocks do relative to the previous difficulty?” It asks “where shouldthe chain be by now, against a schedule fixed at one anchor block — and how far off is it?” That single change, from relative to absolute, is what removes the accumulated error and the oscillation.

The sign of the exponent does all the work. If the chain is running behind schedule — blocks coming too slowly, so time_diff > ideal_delta — the exponent is positive, the target rises, and difficulty dropsto invite blocks back onto schedule. If it's running ahead — blocks too fast — the exponent goes negative, the target shrinks, and difficulty rises.

The one-hour half-life sets the rate of that response. For every hour the chain is behind its ideal schedule, the target doubles — difficulty halves. For every hour ahead, difficulty doubles. That is exactly the steep climb operators have been seeing: sustained hashrate well above target pulls difficulty up roughly an octave per hour, so a couple of hours of heavy mining shows up as the ~4× rise.

Crucially, because the target is always measured against the absoluteanchor schedule and not the previous block, small errors don't compound from one block to the next. There's no feedback loop chasing its own tail — and so none of the Bitcoin-style ringing where difficulty keeps overshooting the level it's trying to find.

Per-block, anchored. The anchor is the ASERT activation block — mainnet height 50,000 (chainparams.cpp:212). It is re-anchored only if an optional retune or half-life upgrade activates, and none are configured on mainnet — those heights are set to INT32_MAX.

Fixed-point 2^x.A blockchain can't do real exponentials — floating point would diverge between nodes and break consensus. So the exponent is scaled by 2^16 (ASERT_RADIX_BITS = 16) and split: the integer part becomes bit-shifts, and the 16-bit fractional part is fed through a cubic polynomial approximation of 2^frac (pow.cpp:1788–1795). The result is bit-identical on every node — which is the entire point.

Slew guard. A NORMAL_SLEW_GUARD_SHIFTclamps how far a single block's target may move from its parent (pow.cpp:850–857), bounding per-block swings from timestamp games.

Fail-closed. Any invalid parameter — a zero or negative half-life, a bad anchor — returns powLimit, the easiest difficulty, rather than crashing (pow.cpp:1763). The chain degrades safely instead of halting.

What the target gates.The resulting target is compared against the block's matmul_digest (pow.cpp:2601) — so this difficulty directly governs how hard the 512×512 M31 matrix-multiply digest has to be. Network hashrate follows from it: networkhashps = difficulty × 2³² / 90.