the market that had to be able to fail

The operation could generate candidate claims far faster than it could break them. That asymmetry is the whole problem: a generator asked to produce interesting statements will produce statements that are interesting to a generator, and a record that only ever accumulates candidates slowly stops being read.

The response was to separate the two jobs and give them different rewards. One class of worker draws from deliberately different registers — pedestrian, speculative, and connective — so the queue is heterogeneous rather than merely long. A second class is chartered to destroy: its assignment is a refutation, and a refutation counts as a result.

A registrar records what survives, and the registrar is never the party that proposed the claim. That single rule is the design's spine. It means the survival column of the record is not self-certified, and it means a worker cannot advance its own candidate by being the only one who reads it.

The architecture admits the uncomfortable part out loud. Most of what the generators produce is expected to die, so the machinery is arranged as several factories feeding a kill queue feeding a ledger. Nothing is shaped so that volume can be mistaken for progress, because the queue's own throughput would expose the mistake within a cycle.

The cost is speed at the top. A worked refutation takes longer than a plausible conjecture, so the ledger grows slowly where it matters and quickly where it does not. What it proves is that the operation can turn adversarial effort on its own output, deliberately, and count the destruction of its own candidates as productive work rather than as failure.