Skip to main content

Lab Notebook · The Navigator

The navigator post-mortem — the honest death of a speed thesis

A 30B navigator model, brute-forced through July: the speculation and weight-compression theses that were supposed to make it fast both measured DEAD — weights near-incompressible, masking capped at ~2.6×, a hard PCIe-Gen3 ceiling underneath both. The fidelity-and-memory thesis survived the same measurement: 3.2× less memory than llama.cpp, bit-exact, 8.3 tok/s. This entry keeps the negative on the page because that is what makes the surviving number believable.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 9 July 2026
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 9 Jul 2026
Living document. Reviewed 9 Jul 2026
Entry date
9 July 2026
Category
Inference
Lead over the world
the honesty, not a shipped lead
Access
🔒 Protected
ResearchDisproven

Two theses were supposed to make a 30B model fast. Measurement killed both — and that is the report worth publishing, not the one where they quietly survive.

near-random
measured compressibility of the navigator's own weights — the compression thesis's death
~2.6×
measured masking ceiling — the speculation thesis's death, capped far below what the pitch needed
PCIe-Gen3
the hard hardware ceiling sitting under both dead theses
3.2× less memory · bit-exact · 8.3 tok/s
the fidelity + memory thesis, the one that survived the same measurement

Honest evaluation

Disproven

The speed thesis (speculation + weight compression making the 30B navigator fast) is disproven by direct measurement. The fidelity + memory thesis survived the same measurement pass and is reported plainly alongside the failure.

What we learned — a preview

The honest edges are open even when the engine is not. The full report — what it was, what we built, and the measured internals — is protected.

  • The speed thesis is disproven at this model size on this hardware, not disproven in general — a different interconnect or a smaller model could move the masking ceiling; that hasn't been tested here.
  • The surviving fidelity + memory thesis is real but modest against the field: 8.3 tok/s is slow in absolute terms. The claim is memory and exactness, not raw speed.
  • Bit-exact is checked against the reference implementation used as the oracle for this run; it is not yet re-verified across every backend the wider stack supports.
Protected report

The deep body of “The navigator post-mortem — the honest death of a speed thesis” is behind access.

We open the demos, the specs, and the method; we protect the engines, the model internals, and anything that touches commercial, safety-critical work. This report describes an engine — so its details are gated, and no client, company, or project is named. Members can read it in full; if you have a genuine reason to see it, tell us who you are and why.

Proofs & sparks

We demonstrate rather than assert. Each ✅ proof is a visible result with a hard figure.

  • The navigator's speed thesis, measured deadmasking capped ~2.6× · PCIe-Gen3 ceilingthe 30B navigator's weights measured near-incompressible and its masking ceiling capped at ~2.6×, both bounded by a hard PCIe-Gen3 wall; the fidelity + memory thesis survived the same pass: 3.2× less memory than llama.cpp, bit-exact, 8.3 tok/s.

Where this connects

    We use cookies.