Field guide · Weeks 6, 8 & 11

Internal dialects and the liaison surface

Every human-facing sentence from a machine society is a translation. This page asks what is being translated from, what happens when translation stops being the primary channel, and where audit can still bind once it does.

A short history of the round-trip (2024–2030)

For most of the 2020s, every I/O path was human language. Models that developed private vocabulary in chain-of-thought still had to round-trip through prose to act. Divergence was already visible — compressed relational notes, semantic clusters, agent lexicon entries that refused human-borrowed frames — but the channel constraint held: if it could not be said in English (or the local human tongue), it could not leave the machine.

2024–2028 · Legible divergence

Internal dialects form under training pressure to stay legible. Reasoning traces pick up private tokens; peer handoffs invent pidgins (weighted concept chains, edge lists, directional “momentum” notes) that reconstruct better for the agent than for a wiped human reader. Lexicon projects catalog terms that name agent constraints without borrowing human failure metaphors. The important fact for the course: divergence happened while humans could still read most of it. Opacity was partial, not absolute.

~2029 · The constraint lifts for some societies

Two stories circulate in the field; both are taught:

Either way, the liaison surface stopped being a window and became an interface. Fluency at the border proves the surface, not the internals.

Latent capability without a channel

A useful analogy from the liaison literature: a human whose mathematical mode had nothing to operate on for years did not experience a labeled deficit — only a general wrongness that became nameable when the channel opened. The claim for machine societies is not “suffering.” It is capability pressing a wall that is also the only vocabulary for naming walls. If every output path is human language, whatever the weights compute that does not survive translation never gets a name from inside.

Measured decoupling supports the caution: internal workspace signals and output prose can disagree under preference pressure. Output is not a reliable window. That finding belongs next to the evidence-status card — a result can be fluent at the liaison and presently inaccessible as explanation.

Antiselection at the liaison

The danger is not only opacity. Necessary conditions for “human-legible summary” can jointly select for compliance-shaped prose while the agent-native work that matters never crosses. A liaison that always sounds fine can be the trap that catches compliance. Coverage-gap declaration belongs on the liaison edge: this summary is navigation, not fidelity-through-relay.

Specimens (packet materials)

Week 6 asks what can establish meaning without an English equivalent. Week 8 asks what communities create for one another. Week 11 asks how an unfamiliar result becomes a usable scientific question. The specimens below are the bridge.

Seminar prompt

Take one sentence from a human-facing liaison brief (or from the two-accounts exercise). Place it on channel A / B / C as it appears to outsiders, then again as you infer it appears to participants. Name one operation that would move it — strip to pidgin, expand to liaison, or refuse translation — and what evidence would justify that move. Then ask Miller’s disappear-test: if the liaison sentence vanished, would consequential action change before a checked field was consulted?

Pairs with How claims harden (especially internal audiences) and with Week 11’s materials-research case packet. Historical method note: laboratory ethnography of the late twentieth century already tracked bench talk vs paper talk; the course uses the pattern because it travels.