Collective research keeps the evidence visible
LensWave Collective organizes observation, challenge and risk review around a shared question. The target design separates market observation, strategy review, integration and risk, and operations and reflection.
Roles describe responsibilities, not an equal number of models already reasoning live. Sources, availability times and versions must remain attached to evidence; confident agreement cannot resolve missing or conflicting facts.
LensWave Collective preserves sources, versions and disagreements. Bounded persistence, shared reuse and recovery have passed local synthetic acceptance checks. An explicit research service is implemented; its latest complete native workflow, live model collaboration and production scheduling still require acceptance.
Strategy Lineage makes a thesis reproducible
Strategy Lineage preserves assumptions, suitable conditions, triggers, exits, invalidation rules and parameter versions. Replay, simulation and future execution should share strategy semantics while retaining explicit differences in venue rules, costs and data.
Input/output specifications define what each module accepts, returns and checks. These specifications and strategy reference archives have been partially validated. A catalog entry does not establish live executability or promise a return.
The trading core follows events into facts
LensWave Trading Core processes market events, strategy instructions, orders, fills and account changes. Event time and information availability time must remain distinct so research cannot see future data.
Models may propose research; execution depends on deterministic rules, authority and state. Exact sizing, atomic capital reservations and risk admission determine whether an instruction is eligible to proceed.
Each approval binds a specific version; each instance respects its account and shared capital constraints. LensWave Trading Core connects checks, reservations, orders, fills and fees. Unknown outcomes require reconciliation of the original operation. Components have local simulation evidence; continuous live execution remains unaccepted.
A partial fill is not the end of an order
Suppose an order fills only in part. The filled quantity creates exposure; the remainder may still reserve funds.
A cancellation request cannot release every reservation before confirmation and subsequent fills are reconciled. If a request times out, query the original operation using its identity. Preserve necessary reservations while the result is unknown and avoid duplicate submissions. Reconcile order events, original fill records and account records before declaring completion. This describes design and bounded simulation evidence, not accepted end-to-end live venue execution.
A partial fill is not the end of an order Filled quantity creates exposure. Remaining funds stay reserved until the result is reconciled.
The learning engine starts with sample governance
LensWave Learning Engine is an R&D direction for turning explicitly authorized records into training samples. Selection checks provenance, versions, time, corrections, privacy and allowed use, removing leakage, duplicates and unverifiable labels.
Profit does not automatically make reasoning correct; waiting, rejection and revision can also teach. A record being available for review does not grant permission to use it for training.
A smaller model with defined tasks
The plan starts from a smaller base model adapted to trading through offline fine-tuning, distillation and related methods. Initial tasks are strategy understanding, finding missing rules, explaining risk and summarizing reconciled records.
A defined scope helps specify data, runtime resources and evaluation criteria; it does not promise speed, cost or accuracy gains. Evaluate unseen cases against the unadapted base and a baseline using the same tools and information.
A trading model must earn its release
LensWave Trading Model is intended for defined research tasks; authorization and control rules continue to govern trading permissions. The plan is to train offline through fine-tuning, distillation and related methods, then evaluate on independent, time-separated samples.
In shadow mode, the candidate model is observed without control over live accounts. Human review precedes a controlled release, with earlier versions available for rollback. Updates must not change running strategies or account permissions without confirmation. A proprietary training dataset, model weights and demonstrated training gains are not yet available.
Ecosystem connections need acceptance too
Data, compute, tools and agent coordination support application development. Integrations require checks of provenance, licences, runtime environments and task delivery; they do not depend on hardware or node packages already having been delivered.
Evidence today, delivery still ahead
Foundational interfaces, bounded research persistence, grid and limited DCA simulations, and membership and Credit simulations have defined-scope evidence. Complete native research workflows, continuous live execution and independent autonomous authorization still require acceptance. Specialist datasets, model weights, evaluated releases and ecosystem feedback remain R&D goals.
Membership, Credit and trading capital
Current personal entry points are Free, Starter and Pro. Membership provides service access and capacity; Credit pays for new research work. Trading principal, venue fees and network fees are managed separately. Saved-strategy capacity and running-instance capacity are also distinct.
Research is quoted before a maximum amount is reserved; delivery is followed by settlement and release of the remainder. Refunds and unknown usage retain their own records. Viewing delivered results, deterministic checks, approval and risk exits do not incur another research Credit charge. Zero Credit does not mean zero trading fees. These flows are locally simulated; real charging is disabled, and old Team parameters do not mean a team product is available.
AI Portfolio: allocation, shares and fees
AI Portfolio explores combining screened onchain yield sources. Compare underlying assets, return sources, denomination, maturity, available liquidity, concentration and exit costs. Vaults with shared underlying exposure must be assessed together. AI supports research and allocation proposals; deterministic rules constrain capital limits, asset allowlists and execution permissions.
The proposed fee model combines an annual management fee with a high-water-mark performance fee. After management fees and relevant costs, only new net gains above the historical high-water mark, adjusted for subscriptions and redemptions, qualify. Recovery of prior losses does not incur another performance fee. Rates, crystallization and formal terms remain undecided. Current calculations and flows are simulated; neither principal nor returns are guaranteed.
Onchain Bot: from intent to authorization
Onchain Bot is planned to connect lightweight interaction with AI trading research. A reviewable plan should specify the network, asset, capital limit, expiry, fees and exit conditions. A market opinion alone does not authorize the use of funds.
Future onchain execution will require quote freshness, slippage limits, contract and function scopes, and transaction identifiers with confirmation tracking. An unknown result requires checking the original operation before another submission. No executable service has been released; wallet connections, social channels and networks are not assumed to be integrated.