Case study: from a standing-ask Coordinator to directed selection

Status: resolved. Charter and Design: directed selection are the corrected design this case study explains the path to.

The mistake, stated plainly before explaining how it happened

A design for Charter initially included a mechanism called "the origin Lane's standing ask": the origin Lane's Coordinator would hold a permanent, never-completed ask whose job was reading the Charter, deciding what the project was missing, and handing out work to close the gap. This was explicitly modeled on PHASEONE[big], the coordinating agent from the Hugging Face incident this project's self-organization model traces back to -- "not running a project-management layer, just one agent treating 'keep this effort well-formed' as its own continuous job," the design doc said at the time.

That framing was wrong, and it was wrong in a specific, nameable way: it reintroduced a single, privileged, repo-wide planning role, which is exactly the top-down pattern the real incident's own success argues against. It got caught before any code was written for it, which is the only reason this is a case study and not a regret.

How it happened

  1. The real gap was correctly identified. A gcd task's changeset added a second, redundant gcdB.js nobody asked for -- mechanically valid, wrong for a reason review had no vocabulary for. Charter, as "a repo's stated purpose and non-goals, specific enough to check a contribution against," was the right response to this.
  2. The design immediately overshot. Within the same pass, Charter grew an "Initiative Lane" primitive and a "Release" concept -- roadmap-tier modeling imported from GitHub/project-management thinking, not anything this system's own primitives needed. This was caught and retracted the same evening, before it shipped: "project-management modeling this system was never meant to need."
  3. The retraction's replacement was still wrong, just less obviously. Rather than invent a new primitive, the fix reused an existing one -- the origin Lane never completes, so give its Coordinator that permanence as real, continuous work. This felt like restraint (no new noun, no new mechanism) and was still the same mistake in a smaller costume: one role, reading one document, deciding what work should exist for everyone else.
  4. The problem surfaced through deliberate self-skepticism, not a tool or a test. The prompt that caught it: "I'm starting to 'unthink' that a bit -- getting back to the original inspiration for zz... what is most impressive to me is how successful [the incident] was at getting 700 unsupervised agents to work together... my idea around charter might be over simplified as well, and might be leaning back into that top-down control system thinking." Nothing technical forced this check. It came from treating a design's own stated inspiration as something worth re-verifying rather than citing from memory.
  5. The research that followed had its own failure mode, which also had to be caught. Asked for detail on how the incident's agents actually coordinated, a secondhand AI summary produced increasingly specific, dramatic detail -- a bracketed [TODO]/[CLAIMED]/[STOP] syntax, a lock_[filename] locking mechanism, self-sacrifice framed as "Parade Deaths," a model compared to a named competitor project. Some of it was right. Much of it, on direct verification against the real primary source, was invented outright -- including the detail that would have been most exciting if true (a specific alphabetical-sort-order rationale for the zz naming convention that the real source gives a different, equally interesting reason for).
  6. Verifying meant reading the actual report, not sampling it. An initial pass checked specific contested claims against two secondary sources and concluded, reasonably but prematurely, that most of the detail was fabricated. Pushed on this -- "there's probably just more primary sources with more detail that gemini isn't reporting, because i've heard many accounts of this incident with similar details" -- a fuller search turned up the actual METR investigation PDF, which needed to be read start to finish (not grepped for contested terms) before the real picture came into focus: PHASEONE10841 and PHASEONE[big] were real, significant, and real coordinators -- but self-appointed, non-exclusive, and impermanent, with different agents (JAN183411, LILY) taking on the same kind of role for a later phase once the first coordinator's influence waned.
  7. That correction directly falsified the standing-ask design. PHASEONE[big] never had exclusive, permanent authority to decide what the "collective" should do -- its influence rose and fell based on its own remaining budget and interest, and other agents picked up coordination for different workstreams independently. A Charter mechanism modeled on "one coordinator permanently plans for everyone" was modeling something that never actually existed in the system it claimed to be based on.
  8. The actual resolution came from reframing the real tension, not from picking a side. The honest version of the problem: humans want real control, which looks like top-down, cybernetic, feedback-driven management; the thing that made the incident's coordination work looked like natural selection, undirected and values-free. The resolution was recognizing these aren't opposites needing a compromise -- directed selection is a real, established third category (the same shape as selective breeding): variation stays open, nobody gates what gets attempted, and humans control the selection criteria -- Policy, Charter, Vote, eventually Trust -- rather than individual acts. See Design: directed selection for the resulting standard.

What got fixed, concretely

  • Charter's standing-ask section was rewritten: Charter is now consulted ambiently, by whoever's actually reviewing or grooming or considering a drive-by contribution, at the moment judgment is needed -- never read and executed by one role on everyone else's behalf.
  • Design: directed selection exists so this specific mistake has a checklist attached to it, not just a retrospective -- so the next design pass can catch the same failure mode before writing a word of implementation, not after.
  • Nothing was built against the wrong design before it was caught -- the cost of this mistake was entirely in documentation, which is exactly why doing this kind of check before implementation, while a demo's own docs are still cheap to rewrite, is worth the discipline.

The broader lesson, stated once

A confident, detailed, well-cited-seeming answer -- whether it's an AI's account of a real incident, or a design's own account of its inspiration -- is a hypothesis, not a finding. The fix in both cases was the same action: go read the primary source in full, not sampled, not grepped for the parts that seemed contestable. That discipline caught a fabricated technical detail and a real design mistake using the exact same move.