From unverified work to a licensable asset
How does human work become an asset with a provenance record strong enough to survive procurement scrutiny — without an intermediary whose interests diverge from both sides of the transaction?
Every structural choice follows one question
Where a design decision looks arbitrary, it is usually the residue of a constraint stated earlier — or one the implementation phase will surface.
Verification precedes value
Nothing is paid on submission. Work enters provisional credit — an accounting entry, not a transferable balance — and settles after adjudication. The delay is the mechanism: it removes the incentive to flood the queue with near-zero-cost submissions.
Bonds secure what reputation cannot
Reputation ranks but cannot enforce. Adjudicators hold bonded positions in TLVY; reputation only decides how much bond a given adjudication volume requires.
Consent is a scope, not a copy
The chain carries a claim about what may be done with a corpus — licence classes, jurisdictions, retention limits, expiry — never the corpus itself. A scope mismatch blocks the licence rather than issuing a warning.
Demand and supply are one loop
Agents exist because a data network with no usage cannot discover which corpora matter. Failure signals from paid tasks make collection targeted rather than speculative.
Permission boundaries are visible
Some components are permissionless, some bonded, some legally gated. Presenting them as uniform would only move the discovery into diligence — a worse outcome than disclosing it in advance.
Three metering surfaces, cut by permission
Layer I meters verified contribution records, Layer II meters completed tasks, Layer III meters inference capacity. Permission classes cut across all three.
| Component | Permission class | Rationale |
|---|---|---|
| Contribution submission | Permissionless | Open entry is what makes the collection surface wide enough to matter |
| Peer review pool | Reputation-gated | Low-accuracy reviewers lose routing priority |
| Adjudicator validation | Bonded | Adjudication carries economic consequences and must be insured |
| Compute node operation | Bonded + residency | Residency claims need something forfeitable behind them |
| Agent deployment | Open, bonded at volume | Experimentation stays free; sustained commercial operation does not |
| Licence issuance | Contractual | Licence terms are legal instruments, not an on-chain function call |
| Settlement | Legally gated | Executed by regulated intermediaries; the protocol instructs and never holds funds |
Utility concentrates at the bonded and metered boundaries. Contributors receive TLVY rather than needing it; agent subscribers pay in local currency, and the protocol's fee accrues in TLVY at the conversion step.
Six classes, three trust edges
Contributors
Submit work and hold no bonded position. Their capital at risk is their time.
Adjudicators
Carry forfeitable bonds and governance weight over quality floors — a structural tension mitigated by seat-term limits and an independent technical chamber, but not engineered away entirely.
Compute nodes
Bond sized against declared throughput, since overclaiming is the failure that matters.
Agent operators
Pay per-task fees that flow into the contributor pool rather than being retained.
Data licensees
Hold no on-chain position; their relationship with the protocol is legal, not cryptographic.
Settlement gateways
Regulated intermediaries under their own licences — a counterparty risk recorded in the risk register.
Trust-minimised
On-chain verification covers contribution provenance, adjudication records and supply accounting.
Bonded
Covers every place a party could profit from misreporting its own conduct.
Contractual
Covers everything the protocol cannot see.
Usage generates the training signal
A misparsed request, a wrongly filled field, an unrecognised local term — each failure record enters the Commons as evaluation-class material.
Execute
An agent completes a task in a live workflow.
Record
The divergence between plan and result is logged as a failure record.
Verify
Adjudicators confirm it and the registry admits it.
Fine-tune
A fine-tuning run consumes the verified record.
Improve
Accuracy rises on the exact workflow that failed.
Verification latency
If time from submission to settlement exceeds contributor tolerance, volume falls regardless of reward. Phase I treats median time-to-settlement as a gate criterion.
Depth threshold
Accuracy rises with Commons depth only above a coverage level. Below it, more collection moves nothing — so capability is sequenced by domain, not by language count.