Grounding

An LLM asked "why did BTC dump?" with no data will produce a fluent, confident, and possibly invented answer. Wickra Copilot removes the invention from the part that matters: the facts the model reasons over are computed, not generated. This page explains why that grounding is trustworthy.

The context is a pure function

build_context(spec, feeds) is deterministic. Given the same spec and the same feeds it returns the same MarketContext, byte for byte, forever. There is no clock, no randomness, no network and no hidden state in the core — the fact list is a mathematical function of its inputs, defined by the six reductions in FACTS.md.

Concretely, the core guarantees:

• Fixed formulas. Each fact kind is a closed-form reduction with a fixed significance threshold. A fact is emitted only when the data actually clears that threshold, and the number in the fact is the number from the feed.
• Rounded, total-ordered output. Every magnitude is rounded to 1e-8 and the facts are sorted by a total order (f64::total_cmp on magnitude, then kind, symbol, ts). No float comparison is left partial, so the serialization is stable.
• Parallel ≡ sequential. The default rayon build and the sequential (--no-default-features, WASM) build produce the same bytes. A dedicated test (parallel_eq_sequential) runs the golden corpus under both and asserts byte equality; CI runs the whole suite under both feature sets.
• Cross-language identity. Because every binding returns the core's compact JSON verbatim, the MarketContext is identical in Rust, Python, Node.js, WASM, C, C++, C#, Go, Java and R. The cross-language golden tests pin this.

Honest facts, not flattering ones

The core never manufactures a signal to look useful. Two examples:

• The vol_spike golden context is empty. Realised volatility over a window divided by volatility over a window that contains it is capped near √2, so it rarely reaches the 1.5x emission threshold. Rather than lower the bar, the golden vol_spike spec produces no facts — an honest golden.
• Below-threshold moves vanish. A 0.2% price move is not a fact. The context contains only what is significant; the absence of a fact is itself information.

Where the model can still be wrong

Grounding pins the facts, not the interpretation. The LLM adapter (LLM_ADAPTER.md) takes the deterministic context and writes prose about it. That prose can still be wrong: the model may misweigh a fact, draw a bad causal link, or editorialize. What it cannot do is invent a liquidation that did not happen or a funding flip that did not occur — those numbers come from the feed, through a pinned formula, before the model sees them.

This is the whole design: make the ground truth deterministic and testable, keep the fallible language layer separate and swappable, and never let the second rewrite the first. The copilot's answer is not financial advice and may be wrong in its reasoning — but the facts under it are real.

See also

Facts · LLM adapter · Architecture.