VINDEX3

AUTHORITY · LIVING SPEC §7–8 · ABI §9.2

WHERE TRUTH COMES FROM

Four independently-derived views of the same system that must be identical — and a fidelity level a profile can never claim above, only voluntarily below.

How does VINDEX3 know a container is complete and faithful?

Four independently derived authorities must agree — what the source declared, what detection resolved, what the graph encodes, and what the container's bytes actually hold — with both ends re-hashed at verify time. Operand closure then proves the executable program is total: every stored tensor maps to an operation, every operation carries judged semantics, and a missing operand is a named refusal. Nothing is asserted; agreement is derived.

THE FOUR-AUTHORITY INVARIANT — G4

Declared ≡ Resolved ≡ Graph ≡ Encoded

Declared

what HF said

Resolved

what the reader interpreted

Graph

the logical system at G2

Encoded

what the container contains

layer_rope_theta[3] = 0

PositionPolicy::None

attention[3].position = none

attention[3].position = none

PASS

The spec's own worked example: a judged semantic, carried intact from source declaration to encoded container.

layer_rope_theta[3] = 0

rope(theta = 0)

attention[3].position = rope

attention[3].position = rope

FAIL

The named trap, rendered: NoPE ≠ rope(theta = 0). A resolver that guesses instead of judging breaks Declared ≡ Resolved — and the comparison catches it. (Values illustrative; the distinction is the spec's, load-bearing.)

Verification re-hashes both ends. Recomputed source, recomputed encoded, recorded — all three must agree. A drifted checkpoint therefore fails differently (source ≠ recorded) from a corrupted container (encoded ≠ recorded). And the tensor table is compared entry by entry, because a relabelled dtype or shape over identical bytes is invisible to every hash in the container.

Declared ≡ Resolved ≡ Graph ≡ Encoded — 1 of 2 rows agree; any disagreement is a distinct, named failure, never a warning.

Authority is derived, never asserted.

THE FOLD — THREE DEPTHS

SAYING SO, PRECISELY

A container is Canonical or Derived. A derived image cannot recompile itself — and saying so is the difference between an artifact that is missing something and one that never promised it. structurally-approximate is honest the same way: the omission is declared, listed, and folded into what the profile may claim.

What a profile may claim is folded down a graded lattice by caps — never declared. Read it, work it, or read the clause that governs it.

Authority — learn: what a claim is worth · inspect: the fold, worked · spec: the clause that governs it. One concept, one URL, the depth of your choosing.

AND YET — CONSISTENCY IS NOT SUFFICIENCY

Four honest witnesses can still tell an incomplete story: all four authorities can faithfully agree on an under-specified system. That is why agreement is only half the proof. The other half is closure — every stored tensor accounted for by an operation, every operation by judged semantics — and it lives in the execution chapter, where the two halves finally lock.

The fold

Fidelity pushed down the lattice, one judged cap at a time, until the level a profile may claim is derived — never asserted.

Agreement proves the story is consistent. Execution proves it is complete.

FOLLOW THIS IDEA — FROM DESCRIPTION TO COMPUTATION →FOLLOW THIS IDEA — THE RECORD →

PUBLISHED 29 AUG 2026 · REVISED 30 AUG 2026 · VERSION 3.0 Candidate

CITE

site build 43a1b6b · built 2026-09-02

CITE THIS

Web page · 3.0 Candidate

Hay, C. (2026). Authority. In VINDEX3 Specification (Version 3.0 Candidate). VINDEX3. https://vindex3.org/authority

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.md §1.2, §7–8 (the living spec)
  • vindex3-format-spec.md §9.2 (the 3.0 Candidate)
  • reference implementation — capability/authority.rs