Skip to main content

Lab Notebook · Council + Gate

Council and gate — why unanimous 10/10 beats an average, and a pass has to exist on disk

A governance discipline run across the whole 2026 portfolio: named expert-voice councils propose, a gate requires unanimous 10/10 to pass — an average of 9.x is a failure, not a pass — and a gate is only counted as passed when its written record actually exists on disk, checked, not just trusted. It caught real problems, including correcting an early fantasy of a 32-GPU federation back down to the true figure of one.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 15 May 2026
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 15 May 2026
Living document. Reviewed 15 May 2026
Entry date
15 May 2026
Category
Method
Lead over the world
the discipline, not the lead
Access
🔓 Public
ResearchProven

Unanimous 10/10 or it's a failure — not an average, not a majority. That single rule is what stopped a 32-GPU fantasy from becoming a plan.

unanimous 10/10
the pass bar — an average score does not count
G<N>_FINAL_*.md
a gate only counts as passed if this record exists on disk, verified
7-40
council sizes across the portfolio (7 named voices to a 40-voice council)

Honest evaluation

Proven

The unanimity-plus-file-verification rule has a real, on-record catch (the 32-GPU correction) to point to, not just a stated design.

What would prove or disprove it further

What would disprove it further: a gate that unanimously passed 10/10, with a verified record on disk, but whose underlying claim later turned out to be wrong anyway — that would show the mechanism catches process gaming but not substantive error, a narrower claim than "the discipline works." So far the record doesn't show that; it shows the opposite, a false claim caught and corrected. The honest cost — that tens of gigabytes of council records with no full-text index make "who discussed X" hard to answer — is real, but it's an overhead problem, not a soundness problem, and it's already named as such in the record.

The evidence — full reasoning behind the verdict

Verdict: proven. The claim under test isn't abstract — it's the specific mechanism (unanimous 10/10, plus a hard check that the passing record actually exists on disk) against a specific failure mode (a plausible-sounding claim gaming an averaged review). That mechanism has a real, named catch to point to: the 32-GPU federation fantasy was struck from the record and corrected to one RTX 3090, and the correction propagated to every sibling project that had inherited it. That is not a hypothetical benefit of the design; it is the design working, once, on the record.

It landed here because the two teeth of the mechanism are independently checkable. Unanimity means one dissenting voice can block a gate regardless of how enthusiastic the rest of the council is — no averaging to hide behind. File-existence verification means a gate can't be marked passed by a status update alone; someone has to check the G<N>_FINAL_*.md record is actually there. Together they close the two obvious ways a fast-moving review process gets gamed.

AI-paced iteration can produce a plausible-sounding proposal in minutes. The problem isn't generating proposals fast — it's that averaging expert opinions lets a weak one slip through disguised as consensus.

What it was

A governance discipline that runs across the whole 2026 portfolio: named expert voices — for the model-architecture line, figures like Turing, von Neumann, Hinton, Karpathy, Sutskever, Shannon, and Hopper; a 25-mind grand council for the cancer-moonshot work; a 40-voice council for the platform-synthesis work — review a proposal, argue it out in a recorded brainstorm, and vote. The rule that makes it real: passing a gate requires a unanimous 10/10. An average of 9.x is a failure, not a pass.

What we built

Research

The mechanism has two teeth, not one. First, unanimity: one dissenting voice among seven, or among forty, blocks the gate — a single credible objection cannot be outvoted by enthusiasm elsewhere in the council. Second, a hard verification rule: a gate only counts as passed when a G<N>_FINAL_*.md record actually exists on disk — and the discipline explicitly includes checking that the file is really there, rather than trusting a summary claiming it passed.

That second rule sounds almost too simple to matter. It matters because it's the difference between a process that can be gamed by an optimistic status update and one that can't.

What we learned — including the honest negative

This is a real, load-bearing operational discipline, not a marketing frame — it has caught real problems. The clearest example: an early fantasy of a 32-GPU (32×RTX 3090) federation was struck from the record by exactly this process and replaced with the true figure — one single RTX 3090 — plus an honest 12-15 month timeline. That correction then propagated to every sibling project that had inherited the inflated assumption.

The honest cost is real too: the process generates real overhead. Document volume without a full-text index across tens of gigabytes of council records means a question as simple as "which council discussed X?" is genuinely hard to answer. This isn't a free discipline — it's a discipline worth its overhead, which is a different claim.

Where it went / status

Active and load-bearing across the current portfolio — model architecture, SaaS development, and cross-domain invention work all run under the same unanimous-10/10, verified-record rule. It is, at this point, closer to a constitution than a checklist.

What is still open — kept visible

The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.

  • The process has real overhead — document volume without a full-text index means 'which council discussed X?' is genuinely hard to answer across tens of GB of records.
  • This is a real, load-bearing operational discipline, not a marketing frame — it directly caught real problems (see the 32-GPU correction) and produced real gate failures elsewhere in the archive.
  • Unanimous 10/10 raises the bar, but doesn't eliminate the need to actually verify the record exists — the discipline includes checking the file is really on disk, not just trusting a summary that says it passed.

Proofs & sparks

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

  • The agentic operating model Council decides, Builders build, Fleet sweeps, Gate verifies, Dashboard holds truth — the operating model is itself a public differentiator.

Where this connects

Sources

  1. vocabotics project audit — Council + Gate governance discipline (unanimous-10/10 rule, verified-record requirement), May 2026vocabotics internal project history · as of May 2026

    We use cookies.