6 · Intelligence and data

This is the layer that holds an opinion. Everything else moves data or executes instructions.

Mindshare

Mindshare measures where collective attention is concentrating and how fast. It is the foundational signal, because in this market attention reliably leads liquidity.

Three inputs combine:

• Social velocity: rate of change in mentions, weighted by each source's historical accuracy rather than its audience size.
• On-chain adoption: unique holders, retention across epochs, direction of net flow.
• Behavioural analytics: what users do after encountering a token, not what they say about it.

From question to signature

1. Parse: resolve to intent plus parameters. "Buy 5 SOL if it drops under $120" yields intent conditional_order, asset SOL, size 5, trigger < 120 USD.
2. Ground: check against live state: does the asset exist on a supported network, is there routable liquidity, is the balance sufficient.
3. Reason: form a judgement and assemble the evidence for it.
4. Explain: render at the user's demonstrated level.
5. Offer: present the action as one confirmable control.
6. Sign: the user approves. Only then does anything execute.

Grounding, and what it rules out

The platform does not answer from model memory. Every factual claim in an answer traces to a record in the intelligence core with lineage attached. Where data is missing or stale the answer says so rather than interpolating.

This constrains the product in a useful way: the assistant cannot speculate about an asset it has no data on, and it cannot produce a confident-sounding answer from nothing. The failure mode of a grounded system is "I don't have that", which is recoverable. The failure mode of an ungrounded one is a plausible fabrication, which is not.

Risk interpretation

Risk is assessed continuously per token across liquidity depth and stability, holder concentration, contract-level flags such as mint authority and transfer restrictions, deployer behaviour, and flow anomalies.

The result is a judgement with its components exposed, so a user can disagree with the conclusion while still seeing the inputs. The platform deliberately does not publish a single opaque safety score: a number that cannot be interrogated is not interpretation, it is just another datum to be taken on trust.

Agents

Agents are scheduled or triggered workflows that chain the above. A daily brief runs discovery, sentiment and smart-money passes and delivers a summary to Telegram at a set time. A monitoring agent watches tracked wallets against size and market-cap rules and streams matches.

Every agent run is logged against the user's account: what ran, what it found, what it pre-staged and what the user signed. See section 9.

Data governance

User data handling, retention and rights are set out in the Privacy Policy and Sub-processors list. The transparency obligations that attach to operating AI in a financial context for EU users are addressed in AI Transparency.