Human contribution, verified and licensed as data assets. Settlement returns to the people who produced the corpus.
Frontier model capability is no longer constrained primarily by compute.
It is constrained by the availability of licensed, provenance-bearing human knowledge in languages, domains and economic contexts that have never been systematically collected. Existing collection is either unconsented, unverified, or extractive toward the people who produce the material.
Talvory is a coordination protocol for producing, verifying, licensing and settling that knowledge. Contributions are classified, peer-reviewed and bonded; verified output becomes a licensable data asset with a defined consent scope and a revenue waterfall back to its producers.
Autonomous agents trained on the resulting corpora form the demand surface, generating the usage that funds the supply side.
TLVY is the protocol's bonding, metering and governance asset — required for validator security, for licence settlement between counterparties with no prior relationship, and for governing a language registry no single foundation can responsibly administer.
Through 2023 and 2024 the constraint was compute, and capital flowed into accelerators. That constraint has loosened faster than most forecasts expected. The next one is less tractable: the stock of high-quality, human-generated, provenance-clear material is finite, and a growing share of new material is machine-generated.
Synthetic data extends the runway, but not indefinitely — each recursion loses tail coverage first. The long tail is exactly what current models handle worst and what clinical, public-sector and regulated financial products need most.
That tail is not scarce because it is hidden. It is scarce because there has never been a mechanism to pay the people who hold it.
living languages are spoken today. Fewer than a hundred have meaningful coverage in publicly documented training corpora.
Compute constraints are solved by capital. Legal constraints are solved by testimony.
Each layer carries a distinct function and its own metering surface. TLVY is the asset that binds them.
Converts human contribution into verified data assets — five contribution classes, peer review with a challenge window, and a registry entry with a consent scope for every asset.
IIAutonomous agents run defined work for micro and informal enterprises. Every executed task returns evidence of what worked and what failed — and that evidence is itself a contribution.
IIIAggregates inference capacity under residency constraints. Compliant capacity is priced at a premium, and that margin funds the Layer II price point.
TLVYContributors do not need TLVY to submit work — they receive it. Enterprises pay in local currency. Utility concentrates at the bonded and metered boundaries.
It is easier to state what the protocol declines to be — those choices carry the most weight in diligence.
The mesh is a cost instrument and a compliance instrument. It wins on residency and on price at moderate utilisation — not on peak throughput for frontier-scale training, and it does not try to.
Every unit in the Commons carries an identifiable contributor, a recorded consent scope and a verification history. Slower to build, materially easier to license — a deliberate trade.
Agent subscriptions are how usage, and therefore data demand, is generated at scale. In isolation the subscriber economics are unremarkable; the correct unit of analysis is the joint system.
Compensation is purely performance-denominated: verifiable data provisioning, validation throughput and compute availability. No passive yield and no fixed financial return.
A token that sits beside a protocol without touching it is a liability. TLVY attaches at three points where the alternative is structurally unavailable or economically inferior.
Validators, compute nodes and agent operators post TLVY as a performance bond. Verification without confiscatable stake degrades to reputation — and reputation has no cost of falsification.
Licence fees and cross-entity settlements between parties with no prior relationship are denominated in TLVY and converted at the gateway. Burn on conversion follows utilisation, not discretion.
A registry that admits corpora, sets quality floors and decides revocation cannot be run by one foundation without becoming a political object. TLVY weights that constituency.
Every executed task tells the Commons where model output failed in a real workflow. That signal is worth more per unit than bulk collection, because it is conditioned on a task someone paid for.
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.
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.
The full framework states how the economics close, and which functions would keep working without a token at all.
