How this product decides what matters

Every factual statement in the product originates in an evidence record: a response from a named provider, stored with the provider's own timestamp for when the observation was true and our timestamp for when we retrieved it. Nothing is written into the system by a language model, and nothing is carried forward or estimated when a provider is unreachable — the affected figure shows an unavailable state instead.

Measurement snapshots are keyed to a time bucket, so a change we report between two observations is a difference between two records that each actually existed. We never compare a live reading against an assumed baseline.

Relevance is a deterministic gate that runs before any interpretation. An event qualifies for a position when it names that position's asset, or when it falls into a market-wide category — macro, regulation, ETF, or a security incident — and either names a bellwether asset (BTC or ETH) or names no asset at all.

Each match is scored from 0 to 100 and must reach 35 to proceed. Events that do not clear this gate are counted as filtered and never reach the interpretation stage. Market-wide items are capped at medium severity: a development that does not name your asset is context, not an alarm.

Materiality is scored from measured numbers and the values you entered. These are the published cut-offs:

Three outlets covering one story is one story. Headlines are compared both by shared vocabulary and by shared distinctive entities, so differently worded coverage of the same event collapses into a single item. Each position receives at most six reported items per pass, of which at most two may be market-wide — and that budget is spent on the most material items, never simply the most recent.

If you record why you took a position, the text is matched against a fixed vocabulary of assumptions. When an event bears on one of them, the product classifies the impact as supporting, challenging, neutral, or unclear.

Unclear is a real and frequent answer. When the direction of a development cannot be settled from the evidence available — an ETF headline without a flow figure, for instance — the product says so and names the missing data rather than guessing. It never states that your thesis is right or wrong.

Explanations are assembled deterministically from the facts and figures already computed. Where a language model is configured, it may only rewrite that prose for readability: it receives a fact sheet, never market access, and any rewrite containing a number absent from the fact sheet is discarded in favour of the deterministic text.

The product does not produce entry or exit calls, price targets, position sizing, or forecasts, and the writing layer is constrained against them.

These data classes are defined in the product but have no feed attached. They are shown as unavailable wherever they would otherwise appear:

• Spot ETF flows — Requires a licensed fund-flow feed. Shown as unavailable rather than estimated.
• Token unlock schedules — Requires a vesting-schedule data source.
• On-chain and whale flows — Requires an on-chain analytics API key.
• Liquidation prints — No public unauthenticated feed. Distance-to-liquidation is computed from your own inputs instead.

If a headline discusses one of these subjects it still reaches you, but it is labelled as a publisher's report rather than a figure this product measured.