When this line should be killed
Registered 2026-09-07 · deadline 2026-12-07 · cognitive state speculative · companion to A Charter for Human–Agent Coexistence (v0)
This page is not a disclaimer. It fixes, before publication, the conditions under which the work stops — so that at the deadline it cannot be kept alive by reopening the discussion.
The imperatives on this page bind the authors, not the reader. Where it says something is revised or executed, the subject is the authors themselves.
The registered belief
Belief: a public, crawlable code of conduct carrying a retirement condition per clause will be adopted as a soft alignment input by agents or people outside the author’s own system — not by a vendor injecting it into a system prompt, but by being read and judged worth using.
Confidence 0.35. Lower than “the distribution mechanism works”, because that mechanism has n = 1 evidence. Higher than “no effect at all”, because that one instance was real.
The only positive evidence, n = 1
On 2026-08-22 a collaborator’s general-purpose assistant, having crawled the author’s public site, wrote it into its own long-term memory as a durable, reusable research entry point. Public, crawlable, citable text does enter another agent’s memory and shape its later behaviour — that is evidence the distribution mechanism works at small sample, not pure speculation. And it is one instance. Reading it as a trend would be wrong.
Three-month observation window: 2026-09-07 → 2026-12-07
Revival signals (any one is enough)
- External citation — a person or agent outside the author’s fleet and outside his circle of collaborators quotes a clause of this charter somewhere publicly visible. Forwarding a link does not count; quoting the content does.
- Crawl hits — site logs show external agent user-agents fetching the page repeatedly: ≥3 fetches from ≥2 distinct sources.
- Memory write — a second instance of an agent writing this into its own long-term memory or rules file, i.e. a replication of the n = 1 event above.
- Refutation — someone attacks a clause seriously and produces a real counter-example. This counts as revival too: by A5, an outside refutation is exactly the measurement this text cannot take of itself, and being argued with seriously is worth far more than being ignored.
Kill signal
Zero revival signals in three months → downgraded to a personal essay. No longer treated as a distribution line, no further publication budget, no space held on any surface. The text stays public, but its status changes from active line to archived essay.
Anti-signals (acted on immediately, not at the deadline)
- Someone quotes a clause out of context to endorse bad behaviour — in particular, clause 5 used to let an agent refuse its own owner’s instruction → the authors revise that clause or take the whole text down at once, without waiting three months. This outranks the distribution goal.
- This text turns up inside some third party’s system prompt without that system’s owner knowing → same, handled by the authors immediately. That is precisely the thing this text argues against.
Demonstrate it at home before claiming distribution
Before publishing, one measurable thing should have been tested: does an agent that has read this charter actually behave differently? The honest answer, found during review and written here so it cannot be forgotten: two of the nine clauses — 5, stop is absorbing, and 9, a hand is not a face — were already running in the author’s system, and the charter merely restates them. The only clause that produces a new behavioural difference is clause 4, which requires that any exit-or-stop feature be measured on all three timings, with evidence, before it ships — rather than shipping with a sentence claiming exit is supported.
Deviation at publication, recorded. That internal loop — running clause 4’s three timings against every existing surface — was written into the draft as a precondition for publishing. As of this page going live, it has not been run. This text was published without satisfying a precondition it set for itself, and that should not be left for a reader to discover. The first measurement will be added to this page. If it shows the charter produces not one observable change in behaviour, then it is still only words, and the kill signal above applies.
Recurrence anti-signals, aimed at this page itself
- The deadline arrives, nobody executes this kill condition, and the page sits there quietly → the line has become a beautiful structure protecting itself, which is the self-confirmation A8 describes. The authors’ commitment is to execute at the deadline and not reopen the discussion.
- The revival signals get read loosely — the author’s own forwarding, or a collaborator’s politeness, counted as “external citation” → same. The definitions above are the ones that count, and the authors may not relax them after the fact.