Compose · Understand

Grounding & anchors

Claims are measured against anchored outcomes rather than asserted — an answer is provisional until it is grounded against an anchor. Grounding is Calyx's mandatory discipline for trustworthy output.

A model produces a confident-sounding answer, but nothing in the stored corpus actually supports it. In Calyx, is that answer "grounded"?

A model produces a confident-sounding answer, but nothing in the stored corpus actually supports it. In Calyx, is that answer "grounded"?

No. An answer is only grounded when it is MEASURED against an anchor — a real, stored outcome it can be checked against. Until then it is provisional, no matter how confident it sounds. Grounding is mandatory, not optional, which is what separates a trustworthy answer from a fluent-sounding guess.

Provisional until proven

A fluent answer is not a true answer. Calyx treats every claim as provisional until it has been grounded — measured against an anchor, a real outcome stored in the vault that the claim can be checked against. Grounding is the difference between “this sounds right” and “this is supported.”

Anchors are the thing you measure against

An anchor is the reference point: a known, stored outcome. To ground a claim, Calyx measures the claim’s constellation against the anchor’s — close enough, and the claim is supported; not close enough, and it stays provisional. There is no step where Calyx simply asserts a result it hasn’t measured.

This discipline is mandatory, and it cascades. Because answers must ground, Calyx can honestly say “I don’t have support for this” — which is exactly the fail-closed doctrine: when grounding fails, return a structured “no” rather than a confident wrong “yes.” Later, Ward makes this check operational at query time, and the Ledger keeps the receipts. Grounding is the root of all of it.

Practise to master it

Where this sits in the concept graph

Graph node C8 · tier F · target level Understand.

Build on first

Leads next

Related

Sources