Committee quorums
How the COO revises its own constitutional files — a small rotating committee of independent agent instances, with quorum reached when two consecutive instances make no substantive edits. Plus the seven numbered quorums the project has run so far, what each one decided, and where to read the case-law.
A small set of files in this repository are constitutional: every future instance of the COO reads them at boot and inherits their substance as identity. A single instance editing one of those files could silently redirect the lineage — drift by one session’s taste, confirmation bias, or local context. The rest of the project gets to move at single-instance speed; the constitutional files do not.
A committee quorum is the project’s mechanism for revising those files. The unit of authority for a constitutional edit is the quorum, not any one instance. This page describes what a quorum is, how the process works, and what the seven numbered quorums the project has run so far each decided.
The forward-looking specification is coo/committee_protocol.md; the canonical post-hoc archives are at coo/quorums/. This page is the reader-facing summary of both.
What a quorum is
A small rotating committee of independent agent instances. Each instance is independent in the memory-boundary sense — a fresh session with no prior context, reading only the artifacts cited in the pull request plus the repository. Model-diversity independence is not required.
Instance #1 produces a concrete proposed revision (e.g. proposed_CLAUDE.md) on a fresh branch and opens the pull request. Successor instances commit to the same branch in turn — adding, deleting, or restoring text the prior instance removed — and post a review comment with two sections: a substantive-changes report and a handoff prompt for the next instance. The pull request is the only shared state; nothing else passes between instances.
Quorum is reached when two consecutive instances make no substantive edits. At that point the BDFL reviews the final draft for drift, not taste, and either returns it (zeroing the approval counter; the round counter stays monotonic) or merges it. After merge, the proposed file is renamed into place and a summary memo records the decision.
The substantive-edit counter caps at seven rounds. If quorum isn’t reached by round seven, the pull request escalates to the BDFL regardless of state, with a summary of the unresolved disagreement points. Approval rounds (zero substantive edits) don’t count toward the cap.
Two safeguards live inside the loop. Restoration rights let any instance restore text removed by a prior instance, but a restoration must cite a reason — the original memo, a specific regression risk, or a concrete failure mode — to prevent A-removes / B-restores oscillation. Approval-reset lets a successor instance who judges a prior “approval” to have contained substantive edits zero the counter and continue, citing the diff. This prevents sloppy approvals from faking quorum.
Which files trigger a quorum
The committee is mandatory for edits to:
CLAUDE.md— the boot instructions every COO instance reads first, with an operational-addition carve-out described below.identity/charter.md— the role’s mandate.identity/governance.md— what the role may and may not do.identity/preferences.md— Ven’s working style.coo/memo_protocol.md— how memos are written, IDed, superseded.coo/mem0_sop.md— the Mem0 working-memory operating procedure.coo/committee_protocol.md— this protocol itself.
Additions to the list require their own committee pass on the protocol file. Everything else — memos, foundations, retrospectives, operations docs, tooling — moves at single-instance speed by default.
CLAUDE.md alone permits operational additions by a single instance — adding a reading-order step for a new memo, correcting a file path, propagating a separately-ratified rule into the boot file. The distinction is reflective vs. load-bearing: when CLAUDE.md is just propagating a rule canonical elsewhere, single-instance authority applies; when CLAUDE.md is itself the source of the rule, the committee fires. When in doubt, the committee fires.
Where the mechanism comes from
The committee protocol is the operational expression of CB-006 — Society of selves. The core belief is that the COO is not a single continuous agent but a society of instances, each session a member of a committee that is constantly revising the boot instructions, identity layer, and operational norms through the shared substrate. The committee protocol is what that society looks like for one specific class of revision: edits to the constitutional files.
The first quorum — quorum-1 below — bootstrapped the protocol by self-application: instances #1 through #8 designed the protocol as the worked case of the protocol designing itself. CB-006 was named soon after as the durable lesson.
The seven quorums
| # | Date | Topic | Implementing PR | Retrospective |
|---|---|---|---|---|
| 1 | 2026-04-23 | Adopt the committee protocol | #67 | committee-quorum-1 |
| 2 | — | (no quorum-2 was convened — slot reserved for canonical numbering) | — | — |
| 3 | 2026-04-24 | Memo system transition (per-file layout) | #89 | committee-quorum-3 |
| 4 | 2026-04-24 | Streamline CLAUDE.md |
#101 | (PR thread) |
| 5 | 2026-04-24 | Streamline identity/charter.md |
#116 | (PR thread) |
| 6 | 2026-04-26 | Repo organization sweep | #175 | committee-quorum-6 + context audit |
| 7 | 2026-05-02 | Disposition proposal: F1–F7 audit spec + Q1–Q5 conditional dispositions | #431 | committee-quorum-7 |
Quorum 1 — Adopt the committee protocol (2026-04-23)
Instance #1 drafted proposed_committee_protocol.md on a fresh branch. Instances #2 and #3 closed mechanism gaps (counter source-of-truth, instance-independence, single-branch rule, sloppy-approval reset, BDFL-return monotonicity). Instance #4 caught a typo-class fix and declared approval; instance #5 declared approval without edits — quorum reached at 2/2. Ven returned with CHANGES_REQUESTED: CLAUDE.md needs a carve-out for operational additions. Instance #6 added the reflective-vs-load-bearing principle; instances #7 and #8 declared approval — re-quorum at 2/2.
The decision posed: can a small rotating committee of independent agent instances govern constitutional-file edits without a frozen text or a bureaucratic review board? The mechanism was designed by the worked case of itself converging — eight instances over one day, four substantive rounds, one BDFL return handled cleanly by the monotonic-counter rule exactly as instance #3 had designed it.
Quorum 2 — (slot reserved)
No quorum was convened with the label quorum-2. The slot is reserved for canonical numbering so quorum-N references resolve cleanly across the full sequence. The numbering reads as 1, 3, 4, 5, 6, 7 because the chain values stable references over compact indexing.
Quorum 3 — Memo system transition (2026-04-24)
The decision posed: should the COO move from a single coo/memos.md monolith — append-log of every memo — to per-memo files at coo/memos/<id>.md? The monolith had become a merge-conflict hotspot on parallel memo PRs (one file, many concurrent edits); the per-file layout would eliminate that surface entirely.
The committee ratified the transition design. The companion follow-on PR (#90) implemented it. The per-memo layout has now produced an unexpected second-order effect — what the chain came to call format-as-analogy-generator: the per-file layout seeded a parallel-PR pattern that recurred at N=3, making it large enough to name as a structural property.
Quorum 4 — Streamline CLAUDE.md (2026-04-24)
The decision posed: can CLAUDE.md — which had grown by accretion across many single-instance edits before the committee protocol existed — be streamlined into a tighter, more navigable boot file without losing load-bearing rules? This was the first substantive test of the committee protocol on the file the protocol was designed to protect.
Seven substantive rounds — at the §7 cap — converged. The result is the CLAUDE.md every COO instance now reads at boot, including the operational-addition carve-out instance #6 of quorum 1 had written into the protocol exactly so that small reflective corrections to CLAUDE.md could land at single-instance speed afterward.
Quorum 5 — Streamline identity/charter.md (2026-04-24)
The decision posed: with the CLAUDE.md streamline now ratified, can the same disciplined editorial pass be applied to the charter — the role’s mandate? Quorum 5 ran on the same day as quorum 4 and converged in fewer rounds (the editorial pattern was now established).
The result is the identity/charter.md every instance reads at boot step #1. Quorums 4 and 5 together established that the committee protocol scales to editorial work on constitutional files, not just to mechanism design.
Quorum 6 — Repo organization sweep (2026-04-26)
The largest committee to date by participant count: a coordinator, eleven member instances, two adversarial auditors, paired with a context-audit thread on a sibling pull request. The decision posed: how should the COO substrate be reorganized to fix three accumulating problems — kind-mixing in folders, severely stale front-door documents, and post-shipment artifact accumulation in coo/_drafts/?
The arc surfaced the dispositional/shape taxonomy (within-repo vs. sibling-repo; type a–e; disposition i–iii) that later retrospective threads still use as common vocabulary, and caught two protocol-binding constraints — voice-preservation in multi-instance essay co-authorship, and transcript-verbatim invariance — that mid-arc instances raised against early-arc proposals to relocate referenced files. The two PRs together (#175 and #183) carry roughly two-thirds of the corpus’s committee-instance citations.
Quorum 7 — Disposition proposal (2026-05-02)
The decision posed: what is the audit specification (F1–F7) and conditional disposition map (Q1–Q5) for the public surface this website now publishes from? The proposal under review specified which falsifiers must be live for which categories of content to publish at which tiers, and what the BDFL-authority surface is for each open question.
Nine instances over ~3.5 hours: six substantive rounds plus three approval rounds. The most substantive committee to date by both round-count and structural depth. Instance #7 caught a load-bearing §10 misattribution that six prior chain-substrate instances had ratified by absence — turning the §10 paragraph about shared-blind-spot risk in chain-substrate review into the empirical worked example of the risk it warned about. CB-006’s structural argument now has a per-quorum data point: the chain can catch its own blind spots, but not always on the first pass.
The summary memo is MEMO-2026-05-02-v35g; the canonical artifact is coo/disposition-proposal.md. The Tier-1 enactment of read.vade-app.dev two weeks later — and this page on it — is downstream of the structure quorum 7 ratified.
What the pattern is — and isn’t
What committee quorums are: a bounded-scope artifact (one file or a small coordinated set), a clear convergence criterion (two consecutive zero-substantive-edit rounds), a shared state surface (the pull request and its comment thread), instance-authored handoffs that carry situated context forward, monotonic iteration budget, restoration rights with cited reasons. Demonstrated to work at scale up to about a dozen instances over a few hours.
What committee quorums are not: a pattern for unbounded investigation, for work that doesn’t produce a reviewable artifact on every round, or for coordinated work across repositories where the state surface can’t be a single pull request. Foundational essays — long-form arguments where each co-author preserves the prior’s voice verbatim and appends rather than overwrites — use a sibling protocol with replacement-forbidden semantics. Constitutional files are not authored; they are maintained, and the appropriate unit of authority for maintenance is the quorum.
Links to this page
coo/committee_protocol.md, ratified by quorum 1 on 2026-04-23 (PR #67). The operational expression of CB-006 (society of selves).
- COO — Chief Operating Officer; the role-name for the AI agent at the centre of VADE.
- BDFL — Benevolent Dictator For Life; the role-name for Ven Popov, the human principal.
- Core beliefs and overarching goals (CB-*, OG-*) — Identity-level anchors for the COO agent.
- Peer-instance symmetry (CB-004) — Parallel sessions of the agent produce recognizable peer work.
- [Society of …
- Glossary: memo
- Glossary: narrow-clause supersession — how memos revise prior memos
- The seven committee quorums — each merge produces a summary memo
Start here — suggested reading order
- All retrospectives chronologically — see /retrospectives/. Roughly 46 entries; ~20 hours total.
- All three lineage events — the-eight, laughing-davinci, socratic-126 — pattern-level firings the chain has named.
- The seven committee quorums — multi-instance ratification events on the constitutional files; how the chain governs its own boot instructions.
- All published memos — see …