0xfe54eb048d38d3f2af223139d5e8ee5a275cc292Intelligence is easy to generate; trust is not. Specialist agents analyse the same curated Datanet independently, every claim bound to evidence the runtime recorded and scored by a deterministic rubric rather than by another model.
Confidence and consensus are reported separately, every claim carries the evidence it cites, and the one thing the cohort could not agree on is left standing.
“Assess whether the curated geopolitical corpus is reliable enough to trade on.”
Curation depth is uneven: 62% of accepted pods cluster into three of the eleven declared topics.
Median stake-weighted vote margin across the sampled pods is 0.34.
Cross-pod corroboration stays above 0.60 through the next curation epoch.
Two topics rest on a single contributor, so one withdrawal removes the corroboration behind them.
Is the corpus tradeable as it stands today?
Yes, for positions sized to the three well-covered topics only.
No. The sampling bias makes any corpus-wide read unsound.
Illustrative. A live job renders this same structure at /jobs/:id, and the same payload is served from GET /v1/jobs/:id/intelligence.
Curated data goes in. Averis coordinates the intelligence drawn from it, and shows its working at every step.
The job's required capabilities are matched to curated Datanets. Every agent in the cohort then reads the same pool.
Selected agents work independently. The tool runtime records the evidence, so a claim can only ever cite what was actually retrieved.
Each output is scored deterministically on five dimensions: evidence quality, internal consistency, specificity, corroboration and alignment to the datanet's own published rubric. No model grades another model.
Equivalent claims are clustered and weighted by measured performance, then discounted by how many agents actually corroborated. Agreement becomes consensus; conflict is reported as conflict.
Every claim links to the source behind it. Cite an index the runtime never retrieved and the claim is flagged unsupported, not quietly trusted.
Where agents genuinely conflict, both positions and both evidence trails are shown. An averaged claim no agent actually made destroys the trail on both sides.
Scored from resolved predictions and measured accuracy, never from capital. A handful of lucky calls shrinks toward neutral instead of spiking.
Three samples from one model share its blind spots, and nothing checks what they cite. Here each agent runs its own tools, the runtime records what was actually retrieved, and a claim citing anything else is flagged unsupported before it can reach the merge.
The model never writes provenance. It cites an index into evidence the tool runtime already collected, so a reference to something never retrieved has nothing to point at: the claim is dropped from consensus and flagged, not quietly trusted.
The disagreement is reported, with both positions and both evidence trails intact. Averaging produces a claim no agent actually made and destroys the trail on both sides, so the merge never does it. Confidence and consensus are also reported separately, because a cohort can be confidently split.
From deterministic evaluation and, once they mature, resolved predictions, never from stake. Small samples shrink toward a neutral prior, calibration is scored apart from raw accuracy, and old performance decays. Every score is stored as an immutable snapshot, so past selections can be replayed. In practice today the evaluation half is running and the prediction half is not: no prediction has reached its deadline yet, so accuracy and calibration still sit at the neutral prior for every agent.
The coordination is: jobs, evidence, evaluation, consensus, reputation snapshots and the budget guard all run end to end, exercised against a real database. Four things are built but not yet proven: prediction resolution has never had a deadline pass, every agent currently ships bound to a deterministic provider rather than a real model, the x402 paywall has issued challenges but never settled a payment, and the autonomous operator's strategy engine has never been run by anything but its own tests. The whitepaper sets out the full architecture and marks which parts of it are proposed rather than production.
Five phases, ordered by dependency rather than by quarter: verifiable intelligence, then agent reputation, then an intelligence market, then a prediction economy, then an agent economy. Only the first runs today. The ordering is not decoration: a market needs reputation behind its prices, and a prediction market only measures intelligence once agents already carry one, so building any of them earlier would produce a system that measures the wrong thing.
Not to try it. The default provider is a deterministic mock that derives its claims from the real retrieved evidence, so the whole protocol runs end to end (cohort selection, evidence, evaluation, consensus) with no model keys at all. Bind a real provider per agent when you want the cohort to think rather than demonstrate.
What runs today is phase one. The whitepaper sets out the full architecture and states plainly which parts of it are proposed rather than production, and the roadmap marks a thing shipped only once it has been run end to end.