CG Common Ground | Common Ground Systems
What we did and what it producedActive

The work, decision by decision

The station design

  1. A small, fixed set of stations, not one session per project. Each station is defined by the type of work it does, not by which project happens to be active: pure routing with no building allowed, research and reading, actual construction work (generators, trackers, code), anything that will carry a partner's name in public, and anything involving numbers, bids or reconciliation that have to be exact rather than approximately right.
  2. The station holds the configuration; the conversation does not carry the work. What survives a session being cleared is the configuration itself. What does not survive, and is not meant to, is the transcript. The actual work lives in files on disk, organized by lane, so a session with no memory of its own history can still pick the work back up cold by reading the file.
  3. A station ends by naming three things. What it produced, where that now lives on disk, and which station should pick up next. The receiving station starts from that file, never from the outgoing session's transcript, which is the rule that actually makes the handoff work: send the artifact, never the transcript.
  4. Measure the floor cost directly, do not assume it. A full accounting of what one session actually paid before any real work started, broken into its own operating context, its available tools, its accumulated memory and its base instructions, showed that a session's own tool configuration was the single largest controllable cost in that fixed overhead, larger than its entire memory load, and that it is something a firm can tune per station rather than accept as fixed.

The review chain for anything unattended

  1. An unattended, scheduled change does not ship on the strength of the session that built it. Because a scheduled task runs with nobody watching it fire, a proposed change to how the fleet of scheduled tasks is configured has to clear a defined chain before it goes live: it is built, then reviewed by an independent pass, then checked a second, separate way by a check that did not write the original change and has no reason to defend it, then the plan and any objections are shared with the team and their reply is read, then a final pass applies everything that review surfaced, and only then does the session that is about to tell a person the work is ready re-verify the actual result itself rather than trust its own earlier claim that it was done.
  2. A budget constraint is a real blocker, not something to route around. When the review chain calls for a specific, higher-tier check and that check is genuinely unavailable within a budget period, the honest answer is to wait for the budget to reset or get an explicit decision to proceed anyway, not to quietly substitute a cheaper check and call the chain satisfied.

What we kept, replaced and installed

We kept the idea that different kinds of work genuinely need different configurations; that part of the old approach was correct, it was just attached to the wrong axis (project instead of work type). We replaced one session per project with a small set of durable, type-of-work stations, and replaced assuming a session's overhead was small and not worth measuring with an actual, direct measurement of what a session costs before real work starts. We installed the handoff rule (name what was produced, where it lives, what is next), the measured floor-cost discipline, and the mandatory review chain for any unattended scheduled change.

What it costs, and what we would watch

Redesigning a running set of live sessions costs real disruption while it happens, and a review chain built specifically for unattended changes takes real time to clear on purpose, by design; that friction is the point, not a flaw to optimize away, because the entire reason the chain exists is that nobody is watching when a scheduled task actually fires. What we would still watch: whether a station's own configuration quietly drifts away from what it is labeled as over time (a documentation label is not the same thing as the actual live setting, and the two can and do diverge), and whether trimming what each station carries by default is actually followed through on rather than left as a known, unbuilt improvement.

What it produced

A small, fixed set of stations, organized by type of work rather than by project, now runs the sessions behind every other system in this room. The fixed cost every session pays before real work starts has been measured directly rather than assumed, and a mandatory, multi-step review chain, including an independent adversarial check, now stands between any proposed unattended change and it actually shipping.

A slice of the project list

A few related projects.