Store meaning, not tokens

Most retrieval systems start by throwing away the thing you care about. A sentence arrives, an embedding model flattens it into a single point, and from then on the system reasons about distance between points. The structure that made the sentence mean something — which idea modified which, what was asserted versus assumed — is gone before the first query runs.

Calyx makes a different first move. We keep the associations.

The cost of flattening

A vector is a lossy summary. Two sentences can land close together while disagreeing, because cosine distance measures surface resemblance, not what was claimed. When you retrieve on distance alone, you get neighbors that look right and answer wrong. The failure is quiet: nothing errors, the model just grounds itself on the nearest plausible text and states it with confidence.

We measured this on our own corpus. For a set of 200 grounded questions, retrieval by raw nearest-neighbor returned a passage that contradicted the question’s premise 18% of the time. The passages were close in vector space. They were also wrong.

Keeping the edges

The four-verb method names the discipline: Measure, Count, Differentiate, Compose. We measure each unit of meaning, count how often associations recur, differentiate the ones that survive contradiction from the ones that do not, and compose answers from the associations that held.

Concretely, a stored claim carries its edges with it: what it depends on, what it contradicts, and how strongly each link was supported by the source. Retrieval then walks structure, not just proximity. A passage that contradicts the question’s premise gets down-weighted because the contradiction is recorded, not because a second model guessed at it after the fact.

What this buys you

• Answers cite the association that grounded them, so a wrong answer is traceable to a wrong link rather than to an opaque distance.
• Contradictions are first-class. A claim that fails against new evidence loses weight instead of lingering as a confident neighbor.
• The store gets sharper as it grows, because more text means more chances to confirm or break an existing association — not just more points crowding the same space.

We do not have this fully solved. Walking structure is more expensive than a single nearest-neighbor lookup, and we are still measuring where the extra cost pays for itself and where a plain vector index is enough. That trade-off is the next build-log.

For now the claim is narrow and testable: when answers are cheap, the systems that win are the ones that remember what you meant — not the ones that remember where your tokens landed.