news/2026_09_16_seer.md

seer is a miner

a few research pages in cyber described a link predictor. today they left the wiki and became a component, seer, and on the way out the framing flipped.

a miner searches a space against a public function and is paid for what it proves. seer does exactly that. the space is pairs of existing files. the function is φ*, the collective focus that tru computes. the payment is the mint — Shapley's share of the proven focus shift Δφ⁺ a link caused, gated by surprise so a copy of what the crowd already knows earns nothing. the warriors doctrine applies as written: make the machine earn, report the true rate. the machine is a neuron's will instead of a chip.

one limitation: seer never creates a file. the price exists only for pairs of files that already exist; a new file's value is the entropy of content nobody has written, and nothing prices that in advance. humans, models and agents create files. seer wires them in.

when it stops: a pair is worth mining iff surprise times proven shift exceeds its cost. when no pair in a neuron's reach clears that bar, seer idles. no threshold to tune. mining is the digestion of creation — mined out per region, per epoch, until something new arrives.

on any chain: the price is tru's, not $CYB's. every token has its own home book, and seer mines whichever chain the neuron is signing into.

two things are honestly open, on the page rather than around it: the mint's stop and the network's need are not proven to coincide — bridges raise λ₂, the mint pays for concentration — and the loop's fixed point is not proven once miners change the graph they are measured on. on the tru side the measure got its own method page: how to observe the spectral gap, the two λ₂'s that share a symbol, and the number measured on bostrom, about 0.13.

the rewrite took six passes. the final one opens with eleven words.

sources

Graph