Long-context · quantized · any-chip · gate-checked — we hold all four at once. We have not benchmarked every other stack, so we do not claim to be the only one that does.
- native ≡ interp
- on every compile, never disabled
- 10 / 10
- verifiability self-score — our own rubric, not an external audit
- type error
- what injection becomes
- 4 of 4
- compounding axes, at once
Most labs treat verification as a test: run it at the end, hope it passes, ship. We treat it as the foundation — the thing everything else is built on and can never be switched off. That inversion is the moat.
Native equals interpreter, always
Shipped & measuredThe rule, stated plainly: on every compile, the native output must equal the
interpreter's output, and the check is never disabled. 🟢 Combined with
diverse double-compiling (build with gcc, build with clang, compare bytes),
this narrows the trusting-trust gap: the artefact is checked against an
independent oracle, not merely trusted because a tool produced it. It narrows the
gap; it does not close it, because the oracle itself is unproven.
Two consequences fall out:
- Verifiability self-scores 10/10. 🔬 Against our own rubric — a self-score, not an external audit — after the gate was hardened and shown to catch fakes. A ten out of ten we awarded ourselves is worth exactly what that sounds like; we publish it labelled, next to the gap.
- Prompt injection becomes a type error. 🟢 When capability and provenance are carried in the type system, an instruction that should not be trusted is not a policy violation to be detected later; it fails to type-check. The attack does not get blocked at runtime — it does not compile.
The compounding axes
Shipped & measuredHere is the precise "best" claim, and the reason it is defensible: there are four properties everyone wants at once, and this stack holds all four — each with the measurement beside it.
| Axis | What it means | The proof |
|---|---|---|
| Long-context | Flat cost as context grows | 2× llama.cpp at 128k, flat line |
| Quantized | Runs small, on cheap silicon | ternary weights, int4 fused decode |
| Any-chip | Not tied to one vendor | 4 backends gate-clean |
| Proven | Differentially tested on every compile, never disabled — not a soundness proof | native ≡ interpreter |
Holding all four at once is the claim we make, and it is the only "best" the numbers support. Note what we are not claiming: we have not benchmarked the field, so we say nothing about how many other stacks hold all four. The claim is about our four measurements, not about anyone else's.
The honest bound
ResearchThere are two bounds, and we publish both.
The soundness bound. The gate is tested, not proven sound. 🔬 It is
differentially checked on every compile and never switched off, which is the
strongest continuous test we know how to run — but nobody has proved it cannot
pass a wrong program. That is the named core gap on this lane, it is recorded in
our own data file (content/data/trees.json, 2026-07-11), and it is why the lane
is tiered research. Everything else in the technical story rests on the gate, so
this is the sentence to read twice.
The coverage bound. A gate is also only as strong as the coverage of what it checks. The honesty census put a number on that: 570 chapters independently audited, ~59% confirmed genuinely real after 95 repairs, and 60 non-functional stubs quarantined. 🔬 The census is still widening. We would rather own both bounds than imply the gate already reaches everywhere, or that it is airtight where it does.
Why it ladders back
Trust, certified — at the language, at the mind, and at the way we work — is what lets AI be good in the strong sense: not merely useful, but safe to rely on. The verification is not paperwork bolted to the product; it is the product. And the honest form of that sentence is the one we ship: the gate is tested on every compile, never disabled, and not yet proven sound.
The open edges
Kept visible on purpose. These are what turn 🔬 into 🟢 — honestly.
- The gate is TESTED, not proven sound — that is the named core gap. It is checked on every compile against an independent oracle; it has not been given a soundness proof. Source: content/data/trees.json (metal tree, `honest`), 2026-07-11.
- The 10/10 is a self-score against our own verifiability rubric, not an external audit — we label it as such.
- The gate's strength is bounded by the coverage of what it checks; the honesty census (570 chapters audited, ~59% confirmed genuinely real, 60 stubs quarantined) is the current bound.
Sources
- vocabotics Dashboard — Why best-in-world: trust certified at three layers, 2026-07-02vocabotics internal record · as of July 2026