Small autonomous pods, coordinated through protocol — not management.

What a cell is

A cell is the smallest research-producing unit at alphabell. The size is bounded for a reason: at three contributors there is enough peer review to catch basic mistakes; at seven contributors there is still enough coherence to share a model of what the cell is doing. The constraint is not strictly enforced — cells temporarily slip outside the range during fissions and mergers — but it is one of the few hard guidelines.

A cell exists when (a) its formation proposal has been approved by a long-tenured-contributor quorum, (b) it has at least three signed-in contributors, and (c) it has an explicit problem statement registered in the internal index. A cell stops existing when it dissolves itself, when it fails to publish for four consecutive quarters, or when it is removed by quorum vote (rare — six times in eight years).

What a cell decides for itself

• Its problem statement, refinement of, and any pivots away from the original formation statement.
• Its working cadence — sprint length, meeting cadence, internal review rhythm, async-vs-synchronous balance. Some cells operate fully async; some prefer dense weekly working sessions.
• Its tooling choices, internal coding style, and language preferences. Cells share the lab's research infrastructure but are not required to share each other's incidental tooling.
• Who joins. Cells onboard their own new contributors, with the long-tenured-contributor quorum acting only as a backstop.
• How it publishes — including style, ordering, and whether to release work as a single bundle or as a sequence of smaller index entries.
• When to dissolve. Cells decide when their problem is exhausted, when the contributors want to move on, or when a fission is in order.

What a cell does not decide for itself

• Whether its work is dual-use, and therefore whether it needs an interpretability pairing. The lab governance protocol decides that on the basis of the charter's dual-use definition; cells cannot opt out of pairings the protocol assigns them.
• The capability-evaluation methodology for any RSI-axis work. That is governed at the axis level.
• The publication delay window for any release that crosses the dual-use threshold. The publication policy governs that, and the long-tenured-contributor quorum can extend the window if the paired interpretability cell requests it.
• Compute commitments. Cells request compute; the scheduler allocates compute. Cells cannot directly trade compute among themselves.

The protocol layer

Cells coordinate through a lightweight protocol layer. Three things matter most:

• Signed proposals. Any non-trivial cross-cell coordination — formation, fission, merger, pairing, publication-policy override, axis-stewardship rotation — happens through a signed proposal. Proposals are written in a templated format, posted to the internal proposal queue, reviewed by a quorum, and ratified or rejected.
• Weighted consensus. Voting on proposals uses tenure-weighted ballots. Long-tenured contributors carry more weight than recent joiners; the exact weighting curve is documented in /governance. The weighting is intentional — it slows down newcomers' ability to flip protocol decisions in their first months — but it is also bounded so that no single contributor or small group can unilaterally decide anything.
• Async by default. No required synchronous meeting. Time-zone overlap is treated as a constraint, not a coordination strategy.

How cells form, fission, and dissolve

Cells form when a problem statement attracts at least three signed-in contributors and survives the formation-proposal review. Most cells form out of conversations among existing contributors; some form out of contributor-cohort onboarding work; a few form out of axis-steward suggestions when the steward identifies an under-resourced sub-area.

Cells fission when their effective research scope outgrows the seven-contributor ceiling. Fission proposals describe the partition of contributors, the partition of tooling, and any shared dependencies that will continue across the new cells. Fissions are the single most common reason for cell-count growth in the lab.

Mergers — the formation of a joint cell from two existing cells — are uncommon. Seven joint cells have been formed in eight years. They usually consolidate work that has been spread across cells for tactical reasons and is now ready for a single research thread. See news/fourier-67-team-up for the most recent example.

Cells dissolve when they decide they are done, when they fail to publish for four quarters, or — rarely — when the long-tenured-contributor quorum removes them. Three cells dissolved in 2025: one was finished; one merged into kalman-04; one was removed after a sustained inability to converge on a problem statement.

Where to read more

• Active cells — the current state of the federation.
• Governance — the proposal and quorum mechanics, in detail.
• Charter — what every contributor signs.
• Contribute — how to be paired with a cell.