seer is a miner. it finds the cyberlink worth making next and hands it to the neuron to sign.
the problem
the cybergraph is shared. among its existing files there are $O(N^2)$ pairs that could be linked. a neuron has a budget of will and one question: which link, right now, teaches the graph the most for what it costs?
tru answers that question exactly, before a token is spent. every candidate link has a price: surprise $\rho$ times proven focus shift $\Delta\phi^+$, net of cost (rewards §6). seer searches the pairs against that price.
why that is mining
a miner searches a space against a public function and is paid for what it proves. seer does exactly this.
| mining | seer |
|---|---|
| hash function | $\phi^*$, the collective focus |
| hashrate | proven $\Delta\phi^+$ per epoch |
| difficulty | the surprise gate $\rho$, times the exponential cost of links (universal law) |
| pool | the graph itself — evidence, not documentation |
| duty cycle | the will the neuron lets it spend |
| block reward | the mint |
the doctrines of warriors apply as written — make the machine earn, report the true rate. the machine is a neuron's will instead of a chip.
what it mines
on $CYB a neuron is paid four ways (rewards):
- the mint — Shapley's share of the proven focus shift $\Delta\phi^+$ its links caused
- the subsidy — proof-of-work for settling other neurons' shares
- query fees
- stake
seer mines the mint. a different miner grinds nonces for the subsidy. two miners, two streams, one graph.
on any chain
the price is tru's, not $CYB's. any home book that settles in proven focus shift can be mined the same way. and there are many books: oikos gives every token its own — one token, one chain — and any neuron with a token can root one. $CYB is the reference, not the only chain. seer mines whichever chain the neuron is signing into.
never a new 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; it exists after the file is made, not before. nothing prices it in advance.
that is the boundary of the method, not a policy. mining is exact because its space is bounded and every answer is priced ahead. creation is synthesis in an unbounded space, and its risks — hallucination, redundancy, storage paid before worth is known — are the risks of an unpriced search. humans, models, and agents create files. seer wires them in. see docs/link-production.md.
when it stops
a pair is worth mining iff $\rho \cdot \Delta\phi^+$ exceeds its cost. when no pair in a neuron's reach clears that bar, seer idles. no threshold to tune — the reward decides, from public state and the neuron's own ego-net. a mined-out graph is one whose structure would not surprise the crowd.
consequences:
- structural work runs out first. bridges raise $\lambda_2$ and spread focus; the mint pays for concentration. densification's phases end in maintenance — bostrom's measured $\lambda_2 \approx 0.13$ was already there.
- mining never stops globally. every new file arrives unwired; temporal decay reopens pairs. mining is the digestion of creation — mined out per region, per epoch, until something new arrives.
- a learned policy trained on realized settlement learns to leave mined-out regions alone on its own.
open:
- the mint's stop and the network's need are not proven to coincide. superadditivity measured that connectivity raises the collective advantage $\sigma$ and lowers syntropy $J$; the mint pays for $J$; the subsidy is proof-of-work, not a fund for structure. a graph can be mined out and still under-connected for foculus finality. either concentration is what matters, or rewards is missing a structural term. seer reports both numbers — structural score and predicted settlement — so the answer comes from evidence.
- the loop's fixed point is not proven. tru's contraction converges $\phi^*$ on a fixed graph each epoch. miners changing the graph in response to $\phi^*$ is a different system; whether it settles or oscillates is open (docs/link-production.md, question 4).
invariants
- outside the contraction — seer never computes $\phi^*$ and never touches focusing. tru stays bit-identical on every machine.
- the mint is the objective — rank or sample by expected settlement net of cost. structural signals are priors, never the quantity optimized.
- no new files — edges between files that exist.
- honesty as a feature — every proposal carries a predicted settlement; every settled epoch scores it against what tru measured. a policy that mispredicts loses trust before it loses stake.
methods
| method | kind | signal | status |
|---|---|---|---|
| densification | analytical | Fiedler vector → $\Delta\lambda_2$ per link; bridge → mesh → semantic | specified, unbuilt |
| learned policy | learned | a GFlowNet sampling links in proportion to expected settlement; densification as teacher | proposal |
map
- specs/interface.md — the contract: inputs, the mining loop, the output, the invariants
- specs/densification.md — cyber-seer, the analytical method
- docs/link-production.md — why: the intelligence problem, and where mining ends and creation begins
- roadmap/ — open proposals
naming
cyber-seer was the densification algorithm when it lived as a research page in cyber. seer is the component; cyber-seer is its first method.
see tru for the measure · rewards for the streams · warriors for the doctrine · cyb for the body that runs it