ReadonlycostsReadonlyprobeThe depth of a probe drafted without a verify round.
Whether the depth next last returned is a probe: drafts the estimate does not pay for, drafted only to observe
acceptance. A drafter that can propose without committing may observe them without a verify round.
The probability that the draft at position (1-based) is accepted, given that the drafts before it were.
The depth whose expected rate for a group of rows sequences beats a round without drafts by the margin, the
best such, up to budget; none when no depth does.
Expected emitted tokens per sequence of a round drafting k for each of rows sequences, per unit of cost,
where a round without drafts over those sequences costs 1 (a plain decode step emits 1 per 1).
The next round's depth: choose's, except that after a run of rounds without drafts it drafts one token,
the cheapest round that observes acceptance, so the estimate keeps following the text (a round without drafts
observes nothing). The run starts at probeInterval rounds and doubles after each probe that leaves the gate
drafting nothing, up to maxProbeInterval: a probe costs more than the plain step it replaces, so text that
stays hard to draft pays for fewer of them. rows: the sequences the round drafts for.
One round's outcome: drafted proposals, the first accepted of them accepted. Positions 1..accepted count an
acceptance and the first rejected position a rejection, after the counts decay: every position's when every
draft was accepted (the text may have turned predictable, so the positions past the round follow the ones it
accepted), only the drafted positions' otherwise (a rejected probe says nothing about deeper drafts). A round
without drafts observes nothing.
Forget what the gate observed: back to the prior, probing at the first interval. For new text (a request joining), whose acceptance the rounds before say little about.
Acceptance is modeled per draft position: the probability that the draft at position i is accepted given that every draft before it was. Each position's decayed counts are shrunk toward the position before it (weight 1; position 1 toward the prior), so a position seldom reached follows the ones that are, and a drafter whose later positions fall off (a recurrent MTP head trained for two steps) and one that stays flat (a block drafter) are both modeled. A round decays the positions it drafted; with
idleDecaya round without drafts decays them all, so what the gate observed ages with the text it decodes and not only with the rounds it drafts.