REPRESENTATION · VINDEX3 ABI §9.1–9.2
SELECTION, NOT CONVERSION
A region set may carry multiple physically present variants. A profile selects a present variant. That is the only legal representation model — a profile saying a format cannot conjure bytes into it.
What is a VINDEX3 representation?
A representation is a physical encoding of a logical model object. One object — target.decoder_stack, an expert bank — can carry several representations with different precision, storage and execution characteristics, catalogued beside the original with fidelity recorded against the source. Its semantic identity never changes with the encoding: a profile selects among physically present variants, and can never request bytes that were not extracted.
What is mixed-precision quantization?
Mixed precision keeps different parts of a model at different numerical precisions — sensitive tensors at BF16, tolerant ones at four-bit — instead of forcing one width on everything. In VINDEX3 that decision is a compiled precision map carried by the artifact: each semantic component is bound to a physically present representation, selection fails closed if the bytes are absent, and the choices answer to recorded quality evidence.
No runtime conversion, ever. The bytes executed are the bytes stored.
WHAT THIS FIXES
Quantization used to be a fork: a second file with no lineage, whose claim to be the model was a filename. And the word exact was free — name a lossy copy the baseline, and nothing was left to contradict it. Here, every copy lives beside its original as a named variant, and what each one is worth — its fidelity — is recorded against the source at extraction: derived, never asserted. The loophole where lossy becomes exact by promotion is closed by construction.
ONE REGION SET — THREE DEPTHS
A REGION SET, IN PLAIN TERMS
A region set is one logical thing in the model — a layer's routed gate-up weights, say — and it may hold more than one physical print of itself. A profile selects a print by name; the baseline is what loads when no profile chooses. New prints arrive as incremental packs: independent, checksummed segment files beside the baseline, with untouched segments hard-linked rather than rewritten — so adding a representation costs the pack's bytes, not the container's.
REGION SET — layer.12.routed.gate_up
The identity never changes. The physical form is chosen.
exact-q6k · baseline
fidelity source-equivalent
storage routed/layer_012.q6k
exact-q6k · baseline
fidelity source-equivalent
storage routed/layer_012.q6k
native-mxfp4
fidelity source-exact
storage routed/layer_012.mxfp4
native-nvfp4
fidelity —
selection fails closed — before any byte is read
Single-copy forbids storing the same bytes twice — not deliberate alternative encodings. Multiple variants of one region set are the format working as designed.
The identity never changes. The physical form is chosen. — 2 physically present variants (exact-q6k, native-mxfp4); baseline exact-q6k. Selecting an absent variant fails closed. Ambiguity is refused, never guessed.
the ABI — §9.1 Representation variants — profiles select bytes, they don't request formats
A profile saying "format": "mxfp4" cannot turn Q6_K bytes into MXFP4 bytes by declaration. Exactly one representation model is legal: a region set may carry multiple physically present variants; a profile selects a present variant.
{
"region_set": "layer.12.routed.gate_up",
"variants": {
"exact-q6k": { "storage": "routed/layer_012.q6k", "fidelity": "source-equivalent" },
"native-mxfp4": { "storage": "routed/layer_012.mxfp4", "fidelity": "source-exact" }
},
"baseline": "exact-q6k"
}- Selecting an absent variant fails closed, naming the region set, the requested variant and the variants actually present — before any byte is read.
- No runtime conversion, ever. "No hidden decode-time repacking" (§10) holds by construction: the bytes executed are the bytes stored.
- Incremental packs. New variants are added beside the baseline as independent, checksummed segment files — the multi-terabyte baseline is never rewritten.
- Single-copy, clarified. The v1 principle forbids storing the same bytes twice; it does not forbid deliberate alternative encodings. The
baselinevariant is the canonical authority; additional variants are opt-in, per-component, and individually removable.
On the graph shape, the same model appears as multiple Representation entries on a logical object (canonical BF16, Q6_K, NVFP4, …), each with recorded fidelity, produced by REPRESENT / vindex represent (§18.3) — selection semantics are identical: present variants only, fail-closed by name.
Same concept, same URL, the depth of your choosing — and the depth you choose is remembered as you move through the rest of the specification.
Representation variants — learn: what a variant is · inspect: the object, operated · spec: the clause that governs it. One concept, one URL, the depth of your choosing.
ELIGIBILITY — SEMANTIC, NOT SHAPE-BASED
Decoder linear weights and expert weights are eligible for representation — the prize at MoE scale. Embeddings, output heads, norms, routers, small vectors, whole auxiliary components, and anything unrecognised are preserved at source precision. Fail safe, never fail small.
PRECISION MAPS — STRUCTURAL, NOT ENUMERATED
A precision map is a compiled program: a default encoding, role-based eligibility, and exceptions matched in declaration order — first match decides. It is a policy, not a transcript of a particular model. The same four-line map compiles against a 416-tensor stack and a 280-tensor one.
PROMOTION — ONLY promote() WRITES A SELECTED_REPRESENTATION EDGE
A backend supporting a format is capability, never authority. Promotion requires measured evidence against a versioned quality gate — and routing evidence is deliberately kept apart from logit evidence. Refusals on this ladder are typed, from QualityUnproven to ConflictingCandidates.
represented → available → runnable → dispatched → measured → selected
Selection, not conversion
One identity, three prints: the baseline holds, a variant is chosen, the choice returns — and then the absent print is asked for, and the refusal lands, before any byte is read.
Fidelity names what a variant is worth. What a whole profile may claim is derived from it — by a fold, not a declaration.
PUBLISHED 29 AUG 2026 · REVISED 30 AUG 2026 · VERSION 3.0 Candidate
CITEsite build 43a1b6b · built 2026-09-02
CITE THIS
Web page · 3.0 Candidate
Hay, C. (2026). Representation. In VINDEX3 Specification (Version 3.0 Candidate). VINDEX3. https://vindex3.org/representation
A chapter of the specification, citable on its own. Cite the version rather than the site — the canonical reference, and how to use it, live at /cite.
SOURCES
- vindex3-format-spec.md §9.1–9.2 (the 3.0 Candidate)
- reference implementation — represent/map.rs · policy.rs · selection.rs