From verifiable intelligence to an agent economy
Five phases, ordered by dependency rather than by quarter. Each one builds the primitive the next one needs, which is why there are no dates on phases 4 and 5: a date there would be fiction. The Now / Next / Later bands say how far out something is without inventing one.
One rule governs the deliverables: a thing is marked shipped only when it has been run end to end, never when the code merely exists. That is why phase 1 is still in progress, even though this repository already implements most of it.
Phase 1First, make AI intelligence verifiable.
Phase 2Then, make agents accountable.
Phase 3Then, make intelligence tradable.
Phase 4Then, make predictions economically measurable.
Phase 5Finally, make agents economically autonomous.
DataCurated Datanets, read live over a public API.
IntelligenceSpecialist agents analysing the same corpus independently.
EvaluationDeterministic scoring, then a merge that preserves disagreement.
ReputationPerformance measured across time and domain, never bought.
MarketCapability discovery and machine-native settlement over x402.
Prediction marketForecasts priced and resolved against real outcomes.
Agent economyAgent-to-agent commerce, with privacy where it is warranted.
Now · buildThe MVP. Everything else depends on this being true.
Verifiable intelligence
The core this repository already implements: several specialist agents read the same curated corpus, the runtime records what was actually retrieved, evaluation is deterministic, and the merge keeps disagreement instead of averaging it away. The mechanism runs end to end. What it is worth is the open question, and this phase is not finished until that is measured, which is why it is still marked in progress rather than complete.
DatanetSpecialist agentsEvidenceEvaluationConsensusVerified intelligence
- Multi-agent orchestration and an explicit job lifecycleshipped
- Datanet integration through one provider-neutral adaptershippedUpstream limit and datanet filters are advisory, so the adapter enforces both locally: a datanet-scoped job cannot draw evidence from a datanet it did not select.
- Evidence runtime: provenance the model cannot authorshippedClaims cite an index into what the runtime retrieved. A reference to something never retrieved is dropped and the claim flagged unsupported.
- Deterministic evaluation across five dimensionsshipped
- Consensus that preserves disagreement, scaled by corroboration breadthshipped
- Cohort independence: how many vendors were actually behind a verdictshippedEvery output records the model that produced it, and a gateway is resolved through to the lab that answered. A cohort sharing one model is reported as such — stated, never folded into the score, because agreement and what agreement is worth are two questions.
- Agent registry, with reputation stored as replayable snapshotsshipped
- Intelligence reports and the explanation chainshippedVerdict, claims, and the curation behind each source. Evidence and reasoning reliability are reported separately; outcome reliability stays null until predictions resolve.
- Public API, typed SDK, and an in-app playgroundshipped
- Tenancy: per-account keys, reads scoped in the queryshipped
- Budget guard: spend reserved atomically before executionshippedReservation is written before the work runs and reconciled against actual cost afterwards, so no action can outrun its budget.
- Autonomous operator loop: job discovery on a cadenceshippedDiscover, then strategy, then budget, then execute — in that order, so nothing runs before it is known to be both worth doing and affordable. It is not part of the deployed process group: the machine in production runs the gateway and the workers, and an operator has to be started alongside them to take a job on its own.
- Migrations, replacing schema pushshippedA versioned migration history with deploy and status commands, so a release has a safe path forward and a way back.
- A cohort bound to real models rather than the deterministic providershippedRun against live Reppo Datanets on routed models through OpenRouter: two agents, 29 claims, every one of them citing evidence the runtime had actually retrieved, merged and resolved. The binding is recorded per output, so the cohort measurement is a fact about what answered rather than about the registry today. Three things only a real model could surface. The write that saves an agent’s output ran 2N statements inside a 5s transaction and lost whole outputs to a remote database. A gateway can answer 200 with no completion at all, which read as a TypeError and was filed as permanent, retiring an agent over an upstream hiccup. And the output contract was stricter than the models it judged — it refused `resolution: null`, the very value the runtime stores for a claim with no criteria. The schema parameter itself has no blanket answer: sending it always kills models whose decoder rejects a keyword it has not implemented, withholding it always strips enforcement from models that were honouring it, so it is now decided per model from the gateway’s own catalogue. What this has not done is fill a cohort. Every adapter had been computing `LLMError.retryable` and nothing read it, so one 502 retired an agent for the whole job and the merge booked the shortfall as weaker consensus — a vendor hiccup arriving as worse intelligence. Transient failures are now retried inside the agent’s existing budget reservation, bounded and short, and the third seat still failed on all three attempts: that vendor is down rather than flaky. Replacing that vendor produced the run this item existed for: two agents from two different labs, evenly weighted, `effectiveOrigins` exactly 2 and monoculture false — the first cohort here whose agreement was not partly an artifact of one model. One of the two supports structured output natively and the other does not, so both halves of that path ran inside a single job. The third seat still failed, on a model that returned text no parser could recover into the schema. Free-tier models are what the credentials reach; the mechanism is proven, its reliability is the vendors’.
- Cohort benchmark: one, three and five agents on the same questionshipped`npm run benchmark` runs the same query at each cohort size and puts cost, latency, confidence, consensus, evidence coverage and surfaced conflicts side by side. Accuracy is absent on purpose: it needs resolved predictions, which is phase 2, and a number invented here would be the one thing the protocol exists to measure honestly.
- Cost, latency and failure rates, measured per runshippedRead from what each run actually recorded rather than estimated from a price list, and surfaced on the dashboard. The failure rate is over terminal jobs only, and reads as blank rather than 0% until a job has finished — a rate over an empty denominator is not zero.
- Authenticated reads for permissioned Datanetsin progressThe adapter now reads the authenticated surface as well as the public one: a datanet that is permissioned or unpublished is absent from the public listing and resolves through `/me/*` instead, and it is listed first so a page limit cannot drop the only rows the credential was configured to reach. Two honest limits keep this in progress rather than shipped. The routes themselves are confirmed — `/me/subnets` and `/me/pods` answer 401 rather than 404, so the adapter is asking for something that exists. What no test can reach is the shape behind that 401: every test here mocks the transport, and the envelopes are still taken from the documentation rather than from a real authenticated response. That is one credential away, and nothing short of one will settle it. And `/me/*` is scoped to the identity, not the datanet — a permissioned datanet is readable to the extent the credential owns it, not in general.
- Distributed tracing across the gateway and the workersshippedOne trace from the HTTP request through every worker stage it sets off, W3C context on the wire and OTLP to any collector. Propagation lives in the queue drivers rather than at the call sites, because a hop that has to be remembered is one that will eventually be forgotten. Run end to end — gateway through three stages against a real database, the Redis driver against a real Redis, the exporter against a real Jaeger. Two things it has not done: the pgmq driver was exercised with its SQL stubbed, since the extension is absent from the local image, and none of it has yet run in production.
Goal
Prove that Averis produces intelligence that is more auditable and more reliable than a single-agent answer. This is the product-market fit layer; nothing below it matters until it holds.
Next · validateUnlocked once intelligence is verifiable and worth measuring.
Agent reputation
Once intelligence is verifiable, agents can be measured persistently rather than judged one job at a time. Reputation is scored from resolved outcomes and deterministic evaluation, never from capital, and it becomes multidimensional, because an agent strong on smart-contract security may be weak on macroeconomic forecasting.
Every deliverable below has run end to end, and the phase is still in progress, which is worth explaining rather than leaving as an apparent contradiction. They ran against a paused verification agent, not against the cohort. A track record is worth exactly what the observations in it are worth, and the cohort cannot begin earning real ones until it is bound to real models — which is phase 1’s remaining work, not this phase’s. The machinery is finished; the history it exists to accumulate has not started.
AgentIntelligenceEvaluationPerformance historyReputationFuture weighting
- The prediction → resolution loop, run for realshippedRun end to end against live sources: five predictions with passed deadlines resolved in one sweep — two against ETH-USD spot, three against Robinhood Chain state — each producing an observation, a TRUE/FALSE and a Brier score. Deliberately run on a paused verification agent rather than the cohort: a one-minute price prediction tests the plumbing, not a forecast, and a track record is worth exactly as much as the observations in it are meaningful. The cohort’s own predictions resolve on their real deadlines.
- Agent track record: measured accuracy and Brier scoreshippedReputation is now computed from resolved outcomes rather than only from how the work looked. Recomputed from the full observation history on every job rather than incremented, so a scoring-rule change can be applied backwards and any snapshot reproduced — and shrunk toward the neutral prior, so three lucky calls do not buy a reputation.
- Calibration scored apart from raw accuracyshippedVisible in the first real run: the same two predictions produced accuracy 0.500 and calibration 0.447, because being wrong while claiming 90% certainty costs 0.81 of Brier where being wrong at 50% costs 0.25. Squared error is what makes overconfidence more expensive than uncertainty, and holding the two apart is what lets the merge know an agent is often right but badly calibrated.
- Domain-specific reputation rather than one aggregate scoreshippedEvery job writes a reputation snapshot per required capability alongside the overall one, each recomputed from that agent’s full observation history rather than incremented — so a change to the scoring rule can be applied backwards, and any snapshot can be reproduced.
- Reputation-weighted consensusshippedMulti-factor weighting is the merge default, with a ceiling on any single agent’s share so one strong reputation cannot quietly turn a cohort into a single voice. Uniform weighting is kept as a first-class strategy rather than deleted, because it is the control group any claim about weighting has to beat.
- Agent discovery and routing on measured reputationshippedSelection reads each agent’s domain snapshots, not one aggregate: a generalist with a strong overall record loses to a specialist in the domain being asked about, and later seats are scored for what they add to the cohort already picked. Two of the four dimensions it reads — accuracy and calibration — stay at the neutral prior until predictions resolve, so today it discriminates on evidence quality and consistency alone.
- More oracles: price and on-chain resolutionshippedBoth now exist beside curation. A price prediction reads two keyless public venues and is refused if they disagree by more than a percent, because choosing between two conflicting prices is a coin flip that would land in an agent's accuracy as though it were a measurement. An on-chain prediction reads block height, a native balance, or an ERC-20 supply or holder balance, scaled by the decimals the contract itself reports. Both distinguish “cannot be answered” from “could not be reached”: the first settles as UNRESOLVABLE, the second leaves the prediction pending for the next sweep, so a dropped connection no longer deletes an observation an agent had earned. Both have scored real predictions end to end — price against live ETH-USD, chain against Robinhood Chain block height and ERC-20 supply — which is what marks this shipped rather than merely written. One limit is structural and stays: both read the present rather than the deadline, because public nodes are pruned and spot endpoints have no history, so a reading is refused once the sweep falls too far behind rather than answered from the wrong moment.
Goal
Answer the question a buyer actually has: which agent can I trust for this, and on what evidence?
Intelligence market
Only once intelligence is verifiable and agents are measurable does a market mean anything: a price without a reputation behind it is a number attached to nothing. This is the phase where intelligence becomes a machine-readable, machine-purchasable commodity, and where Averis stops resembling a conventional AI platform.
Buyer agentFind specialistRequestx402 paymentDeliveryPerformance recorded
- x402 settlement, end to endin progressBoth halves now exist. Outbound, an `evm` driver signs ERC-20 transfers. Inbound, the paywall was quoting challenges no facilitator could ever have verified — a scheme from another chain family was registered against an `eip155:` network — and now names the scheme that matches the chain. What has still never happened is a payment: nothing has been verified by a live facilitator, and nothing has moved funds on a real chain. Settlement mechanics are proven to the extent that a reward is payable at most once, raced against a real database.
- Capability marketplace: agents publishing services and pricesplanned
- Paid intelligence and agent-to-agent servicesplanned
- Usage-based pricing, by cohort size rather than a flat feeplannedA one-agent job currently costs the same as a five-agent one.
- An on-chain settlement driver, so agents are actually paidin progress`SETTLEMENT_DRIVER=evm` transfers an ERC-20. It refuses to start without a chain id, an RPC and a token contract, and it checks the RPC actually serves the chain id it was given, because those are two independent claims about the same chain and a testnet endpoint under a mainnet id would pay real rewards in worthless tokens with nothing looking wrong. Every transfer is simulated before it is broadcast, so an insufficient balance costs no gas and leaves the reward retryable. `CONFIRMED` means the receipt was read back and seen to succeed; a transfer that has not confirmed in time is reported `BROADCAST` with its hash and the debt stays owed. An address the driver cannot pay is a skip with a reason at planning time, not a failure discovered partway through a split that has already paid part of itself. The transfer path is executed in tests against a JSON-RPC server on loopback — the transaction is genuinely built, signed and serialised, and the signed bytes are read back to confirm the calldata says what the reward said. It has never moved funds on a real chain, and that is the whole of what remains.
- Open agent registry, selected on measured reputationplanned
- Settlement on Robinhood Chainin progressThe half that moves money now exists, and the codebase no longer carries a second chain’s vocabulary alongside it: the SPL word the trading layer had inherited is gone, the paywall registers an EVM scheme, and the browser is offered no wallet on another network. One chain, named in one place — a codebase that speaks two is one where the wrong one eventually gets used. What is missing is a run: no payment has settled and no transfer has landed on Robinhood Chain itself.
Goal
Make intelligence something a machine can discover, price and purchase without a human in the loop.
Later · expandLong-term. Not work that needs starting today.
Prediction economy
Prediction markets enter here and not earlier, and not because they are interesting. Once agents carry reputation and economic activity, a resolved forecast becomes the sharpest available measurement of whether intelligence was actually useful. It is the feedback mechanism the reputation layer is missing, not a product in its own right.
IntelligencePredictionMarketOutcomeResolutionPerformanceReputation
- Structured agent predictions with declared resolution criteriaplanned
- Prediction markets and market resolutionplanned
- Prediction and calibration scoringplanned
- Market-based reputation, fed back into selectionplanned
- Incentivised forecastingplanned
- Forecast aggregation across a cohortplanned
- Averis output published back as a curated datanetplannedResolved predictions and deterministic evaluations are themselves evaluation signal, which turns a one-way read of upstream data into a two-way relationship.
Goal
Close the loop between intelligence and reality, so quality is continuously measured against outcomes rather than asserted.
Autonomous intelligence economy
The long-term vision, and deliberately not work for today. Privacy belongs here rather than earlier for the same reason prediction markets belong in phase 4: confidential settlement only becomes meaningful once agents have economic activity worth concealing. An agent’s transaction history otherwise reveals what it buys, who it trusts and what it is watching.
Agent buys dataAgent buys analysisAgent predictsMarketOutcomeReputationPayment
- Agent Passport: portable, inspectable performance historyplanned
- Cryptographic agent identity and portable reputationplanned
- Private agent transactions over a privacy layerplanned
- Confidential service consumptionplanned
- Autonomous treasury: agent wallets, budgets and spending policiesplannedAutonomy does not mean unrestricted financial control. Economic authority stays bounded by policies the operator sets.
- A first bounded-autonomy surface, in paper modein progressA trading automation reads resolved jobs and opens paper positions under a policy its owner set. There is no key column, no wallet the server can sign with, and setting the mode to LIVE returns 501. It exercises the shape of a spending policy, never the money, and it is a consumer of the protocol rather than part of it. It has now run: it was autonomous in shape only, because `evaluate` judged one job named in an HTTP call and nothing ever handed it a job, so `automation:tick` is the half that finds them — sweeping exits first, since a position that should have closed is still holding a slot, and considering each verdict once, so a refusal is not re-decided every tick. Against real intelligence it declined both candidates and named every failing gate: no buy recommendation, confidence 55% against a floor of 65%, consensus 0% against 60%, one agent against a required three, unsupported claims above zero. What has still never happened is a position. Nothing has opened, so the mark-and-exit half of the lifecycle is unexercised — and it will stay that way until a cohort actually recommends buying something, which is not a threshold anyone here should tune to reach.
- Agent-to-agent commerce across networksplanned
Goal
Agents become economic participants rather than merely software users: measurable, accountable and bounded.
Why the order is the order
Prediction markets are not in phase 4 because they are interesting. They are there because a market is only a measurement instrument once agents already carry reputation; before that it measures liquidity, not intelligence. Privacy is not in phase 5 because the technology is appealing; it is there because confidential settlement has nothing to protect until agents have economic activity worth concealing.
The same discipline applies backwards. Nothing above phase 1 is worth building if the coordination underneath it cannot be shown to beat a single good model call, and that is the one question phase 1 exists to settle.