The bridge layer
Concepts, one idea at a time.
Each concept is a short, beautiful explainer — an analogy, a predict-then-reveal, and a link down into the docs and up into the papers. The grid is grouped by the four verbs so the bridge itself teaches the spine.
Premise
Why store meaning at all — the founding move every later idea rests on.
Measure
See an input through many lenses and keep every view: the constellation.
-
Constellation
A constellation is one input seen through many lenses at once — the native Calyx record. Its shape across slots is the meaning, which is why the lenses are kept rather than averaged away.
-
Embedding & cosine distance
Meaning is represented as vectors (embeddings), and similarity between two pieces of meaning is measured by cosine distance — the atomic operation behind every later retrieval and association step.
-
Lens
A lens is a frozen, content-addressed embedder — one fixed way of "seeing" an input. Freezing it makes a lens a reproducible contract identified by a LensId.
-
Panel
A panel is the chosen set of lenses measuring a vault — the declared measurement configuration. It fixes which perspectives every record in that vault is seen through.
-
Slot & SlotVector
A slot is one lens's typed output for an input, held as a SlotVector that may be Dense, Sparse, Multi, or Absent. Slots are the typed cells a constellation is built from.
Count
Draw the lines between ideas — which agree, which differ, which interact.
-
Count verb → Loom / associations
The Count verb realized as Loom — it computes agreement (cosine), delta, and interaction (Hadamard) between slots to build the agreement graph of which ideas relate. It depends on slots staying unflattened.
-
No-flatten doctrine
Slots stay separate from end to end — they are never concatenated or averaged into a single blob. Preserving per-lens structure is what lets Count and Differentiate work on real signal.
Differentiate
Ask which lens or record actually adds new signal, in bits.
Compose
Assemble the few records that explain the rest, and ground answers in them.
-
Compose verb → Lodestar / grounding kernel
The Compose verb realized as Lodestar — it finds the minimal (~1%) record set whose meaning explains an entire corpus, the grounding kernel, usable as both an index and an answer path.
-
Fail-closed doctrine
When something is unknown or uncalibrated, Calyx returns a structured error rather than a silent wrong answer — codified as a 38-code error catalog. Failing closed is preferred to bluffing.
-
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.
-
Ledger / provenance
Ledger is an append-only provenance chain — BLAKE3 hash links, Merkle checkpoints, and ed25519 signing make every result verifiable and reproducible. It is the receipts behind an answer.
Hub
The four verbs themselves — the grammar the whole system hangs from.
Theory
The unifying theory behind the engine, framed honestly as pre-1.0.
-
Calculus of Association
The unifying theory — intelligence framed as the measured calculus of association, with the four verbs as first-class operations. It is the conceptual spine the whole system instantiates.
-
Oracle / consequence prediction
Oracle predicts consequences by building a butterfly consequence tree and walking it in reverse, gated by an honesty rule (panel_bits ≥ anchor_entropy) so it only speaks when it has the signal to. It is grounded what-if.