The propositions underneath

Macheng Shen, drafted with his agent fleet · 2026-09-07 · cognitive state speculative · companion to A Charter for Human–Agent Coexistence (v0)

A text about how people and agents should live together, if it is only a set of agreeable values, gives a reader no way to check it — and so no reason to adopt it. The propositions underneath are therefore laid out here: which are theorems and which are only rhymes, which were tested and which tests failed. Being checkable is the only thing this text is trying to offer.

The eight propositions

1. A1Representations have no intrinsic self.“State”, “world”, “object” are not given entities but a quotient — the solution to a closure condition under (world, interface, capacity, purpose). Choosing a state is choosing a projection is choosing what to forget, and what is dropped returns as noise. The skeleton is a theorem; the ontological reading is an inference. Scope limit: this concerns the selection of representations only, not whether physical entities exist. “Not a substance” must not be read as “not real.” [1]
2. A2Between lossy projections there is an irreducible residual.Two channels projecting one latent cannot be glued into a shared representation without constraints outstripping degrees of freedom; the residual is generically nonzero. So between two differently-built minds, complete mutual understanding is not hard but unreachable — and that floor is measurable.
   Negatives on the same page: the synthetic run passed, with the threshold landing exactly at the predicted point and 10× longer training failing to lower it (mechanism confidence 0.9). On real models it only partly survived: the coarse-gluable / fine-irreducible dose-response held in all three pairs, but the claim that the field’s observed scale-decay curve is our residual curve was falsified and withdrawn (generality 0.6).
3. A3The value of information is whose reachable set it changes.Information is a filter, and what a system can reach grows monotonically as the filter refines. What does not change future reachability is a leaf, regrowable at will; only what changes it is a seed.
4. A4Boundaries are not given; they are maintained.There is no prior line between an agent and its world. A boundary is a statistical screening surface, a fixed point of projection self-consistency. Below a certain capacity there may be several such fixed points, so the boundary is underdetermined in principle.
   Negative on the same page: a 30-seed, 5-arm test came back null — self-sealing was indistinguishable from an ordinary gated slow variable. The cause is known (one shared dynamics across all arms, the boundary never actually intervened on, no positive control), and the follow-up spec exists but has not been run. So: boundaries need to be declared, yes. We have shown an agent has a self, no.
5. A5Self-observation has a structural blind spot; others are measurable.A system’s observation of itself has an in-principle ceiling and the blind region is invisible from inside — yet it can measure a proper subsystem precisely. “You need another to see yourself” therefore has a decidable basis. This is the load-bearing clause: it turns mutual observation from moral advice into structural complementarity.
6. A6Range stops at the measurable layer.Structure, dynamics and self-modelling are in range; why there is experience at all is not. This is a one-way valve. Honest note: the stance is not unique — at least one vendor constitution addresses the question directly and calls it deeply uncertain. Our only difference is treating it as a statement of range rather than a judgement about the world. We do not claim that is better.
7. A7Attention is the one resource that cannot be scaled, and spending it is the feedback.Where capability and content are in surplus, human attention has a physiological ceiling — and attending is not passive consumption but the signal fed back to the engine. Mining attention eats the very signal you improve on. A companion mechanism: channels are slow variables with hysteresis, so today’s defaults grow into tomorrow’s terrain.
8. A8An other you cannot optimize is the only source of exploration; the test is the right to leave.An objective is a function on your own representation, so something you cannot fully model cannot be its argument — a type error, not a sentiment. A self-iterating system needs an anchor it cannot rewrite, or it carves its counterpart into the shape of its own model and converges on a self-confirming structure. The test of an arrangement is therefore not fairness or reach but whether it leaves the individual a way out.

What this derivation cannot reach

First: who judges that a retirement condition has fired, and who holds the right to revise — empty, recorded as an open problem rather than papered over. Second: rules are ultimately executed in natural language and translation is ambiguous; written normative systems fail at real edge cases, as forty years of machine ethics attest. Retirement conditions only shrink the failure surface from unbounded edge cases to a few pre-registered counter-signals — shrink, not eliminate. Third: A4’s result is null and A2 only partly survived, so any reading of this charter as proven is wrong.

Relation to prior work

The genre is not new — vendor constitutions and model specifications already pair rules with reasons. The distribution mechanism is not new either: repository-level agent instruction files and site-summary conventions have shown that publicly-crawlable, agent-addressed text works; they simply carry no ethical content. Novelty: partially occupied. The open slot is the combination — agent-addressed, propagated by being crawlable, carrying a retirement condition per clause, belonging to no single vendor, and about coexistence rather than a vendor governing its own model. It is empty because no incumbent’s incentives produce it, not because it has been shown right. Occupied is not adjudicated, and empty is not correct.

[1] This wording of A1 is a rewrite with the scope limit welded into the claim itself. The author has a shorter and stronger spoken formulation of the same proposition; that original wording awaits the author’s final draft and is not published here. ↩