No Language Model In This Stack

No language model in this stack

This page has no board row, no lifecycle status and no closing verdict, for the same reason the three libraries have none. It does not instantiate the shear recipe — there is no sky forcing, no clock and no track, so there is nothing here to seal against a frozen law. What it has instead is a pair of gates, each with control arms, and a measurement you can re-run.

It is here because every other page on this wiki rests on it. A board that says its answers are exact integers a stranger can re-derive is making a claim about the machines it is served from, and until 2026-09-12 that claim had a hole in it: one process in the substrate could call a model API, and one cell had been running a local model server since August.

Founder, 2026-09-12: "yes strip the gym proxy from the roll script we do not connect to any llm in the stact we get facts for raw vector files." And, the same day: "validate those are old and not used. and also get rid on any bridge to an llm in any of my servers. the fact they are suposes to be cattle not pets bothers me."

Both halves of that instruction are done, and the second half is the one this page spends most of its words on, because validate those are old and not used is the part that could have been skipped quietly.

What was there, and what it was for

the outbound bridge	GymUpstreamAdapter.swift, 317 lines — the substrate's only code path that built a request to a model API host. It listened on :9333, translated one request shape into another, and read a key file the founder had placed on the cell
how it was reached	the membrane's gym leg, pointed at it by an environment variable the roll script set
how it kept coming back	the roll script shipped its unit and started it before the membrane on every roll. It was disabled by hand nine times. Measured 2026-09-11: a roll had left it active on 8 of 9 cells
the local model server	on one cell, a llama-server process on a 9.8 GB quantised model directory, bound to 127.0.0.1:8091, with a socat bridge onto the docker gateway. Running since 23 August. Its bridge unit's own comment names its only consumer: the Python language-inject sidecar the Swift membrane retired

Validated unused before it was touched

This is the part that matters, and it is the part a reader should be most sceptical of, so here is the evidence rather than the conclusion. Measured on the cell, before anything was stopped:

what was checked	what it read
CPU time consumed since it started	167 seconds in 19.9 days
its journal	zero entries
established connections to its port	none
the model file's access time	still the second it was loaded, 23 August — nothing had read the weights since
what the bridge was built for	a Python sidecar that no longer exists in the tree

A process with 167 seconds of CPU across three weeks, no log lines and an untouched model file was not serving anything. REPORTED — the figures above are the cell's own accounting, read once, before removal; a second reading is not possible now that the thing is gone, which is a limit of this page and is stated rather than hidden.

What was removed, and what came back

The adapter, its systemd unit, its installer script and its command-line subcommand are deleted from the source. The subcommand's slot now prints REMOVED and exits non-zero, so a stale unit invoking it fails loudly instead of silently doing nothing. The membrane is no longer pointed at any upstream, so its gym leg answers with a named refusal — gym_upstream_unconfigured — rather than opening a connection.

On the cells: both model units stopped, deleted and masked; the socat bridge gone; 53 runtime files of the model server removed; the 9.8 GB model directory removed. That returned 8.5 GB of RAM and 10 GB of disk to the cell and made its model directory identical to its eight peers'.

The roll script no longer installs the unit. It now stops it, deletes the file, then masks it, and refuses to continue if the unit is afterwards active or unmasked.

Two traps, written down because both cost a measurement

Delete first, mask second. systemctl mask writes its own symlink at the unit's path. Masking and then deleting the file at that path removes the mask you just installed. Done in the wrong order, nine cells reported not-found where the intended state was masked — and not-found is exactly what a cell says when a unit is merely absent and free to be re-installed by the next roll. Done in the right order, all nine read masked, with the link pointing at /dev/null.

A probe whose pattern matches its own command line lies. An unbracketed process search for llama|vllm|ollama reported two model processes on every cell in the fleet. Both were the probe: the search string appeared in the remote command line the search was running inside. The gate now matches on the exact process name, and its pattern is bracketed so it cannot match itself. This is the same defect, in a different costume, as a waiter that waits forever because it is waiting for itself.

The two gates, and why each has control arms

An instrument that is always green and an instrument that is always red are the same defect: neither carries information. So both gates below are run in both directions, and the arms that must refuse are listed with them.

gate	what it refuses	control arms
no-llm-upstream-in-the-stack.sh	the adapter, its unit or its installer returning; any Swift or shell line that builds a request to a model API host; the roll script shipping or starting the unit again; the membrane being re-pointed at an upstream	a planted adapter file, a planted request line, a roll script with the mask removed, and a membrane pointed at a host — each must refuse, and does
cells-are-cattle-not-pets.sh	any of five readings disagreeing across the cells that answer, and the model-bridge reading being anything but zero even when every cell agrees	a differing cell binary, one cell running a model server, every cell running one, and an empty fleet — the last because a gate handed nothing must not pass

The first gate's second arm initially failed on its own fixture: the file carried the literal host string it refuses, so it flagged itself. The fixture is now assembled from pieces at run time. A gate whose source contains the thing it forbids will refuse its own source, and that is worth knowing before writing the next one.

The fleet, graded on agreement

There is no golden image to compare a cell against — the cells are built by rolls, not stamped from a master — so the only available truth is that they agree with each other. A gate that refuses when eight agree and one does not names the odd cell without needing a manifest of what is correct.

Measured per cell on 2026-09-12, never through the apex that fronts them:

category	reading
cell binary	unanimous — one digest across the nine
model bridge: processes, ports, unmasked units	unanimous and zero — procs:0 ports:0 units:0
forbidden runtimes under the cell's own bin directory (Python, Node, model runtime)	unanimous, 0 — it was 2 on seven cells and 4 on the ninth before the sweep
the models directory	unanimous
the cell bin file list	unanimous
enabled service units	two values — not unanimous, and named below

The one that does not agree, published at the same weight

gaiaftcl-treasury-settle.service is enabled and active on two of the nine cells, started three minutes apart. Its own unit description reads "Swift, durable pull, one payer".

Whether that is a defect depends on where its exclusivity comes from, and this page does not know: if one payer is enforced by the durable consumer, two running instances are harmless because the broker hands each message to exactly one of them; if it is enforced by being the only enabled unit, then there are two payers. NOT KNOWN, and left running rather than stopped, because it is settlement rather than debris and stopping a payer to make a gate green is the wrong trade. It is the founder's call.

One further disagreement, separate and also open: one cell is not listening on the port the other eight answer on for the mesh broker. That is a question about that cell's broker surface, not about models, and it is named here so the "unanimous" rows above are not read as "everything agrees".

No GPU either

Measured per cell, 2026-09-12, on all nine: aarch64 ARM Neoverse-N1, and 0 of 9 carry a GPU driver, a GPU device node, or a CUDA, ROCm or OpenCL runtime. The serving path is CPU integer arithmetic end to end. There is no accelerator on these machines to dispatch to, and nothing in the path that would ask for one.

That is a statement about the nine production cells, not about the Mac the substrate is built on — the build host is a different question from the serving path.

No float, and this one you can check from a terminal

The programme's oldest rule is that no verdict may depend on a rounding. On the public court that is enforced at the door rather than asserted in prose: every numeric input in all 51 published tools is a decimal string, not a number, so the wire carries no floating-point type to put a float into.

Measured live on 2026-09-12, three calls to the same tool — the middle two are the control arms:

what was posted	what came back
A = "1,2,3,4,5,6,7,8|1" — integers as decimal strings	CALORIE · AFFINE_JZ_OP · result 2,3,4,5,6,7,8,9|1
A = "1.5,2,3,4,5,6,7,8|1" — one value carrying a decimal point	REFUSED · AFFINE_JZ_PARSE · result empty
A = 1.5 — a JSON float rather than a decimal string	REFUSED · AFFINE_JZ_PARSE · result empty

An instrument that accepted everything and an instrument that refused everything would be equally useless. This one answers the integer and refuses the float, and names which it did.

The source side is held by four gates, all green at the commit this page was written from, each run here and quoted as it printed:

gate	what it printed
float literals in the substrate paths	clean (0 violations, 0 code lines suppressed, 4 comments exempted)
float types — a ratchet that may only fall	The ceiling is ZERO: the substrate path holds no float-typed declaration, and the next one to appear fails this gate on the commit that adds it.
floats in the Rust client shell	NO_FLOATS_IN_RUST_CLIENT_CLEAN
floating-point instructions in the bare-metal OS image	NO_FPU_GATE=CLOSED arch=aarch64 boot=absent

And the limitation of that kind of gate is on the record rather than left to be found. A source scan reads zero when the spelling it matches is not the spelling in use — which is how 141 float-typed declarations once lived inside a path whose gate reported none, because the gate matched float literals and the declarations were bare types. That history is written into the repository's own constraints file for exactly this reason. It is why the wire measurement leads this section: a court that refuses the value at the door cannot be blind to how the value was spelled upstream.

What replaced it

Facts come from raw vector files through an exact court. The board's translation is the worked example: each cell derives a per-language coordinate chart in memory at every start, from the raw rows of a pinned public weight file read by byte range and digest-checked, quantised onto an integer lattice and ordered by an exact 256-bit integer comparison — and it refuses, by name, every word it cannot carry. There is no model in that loop and nothing written to disk. → Study 47 — translation shear

What this page does not claim

• It does not claim no model was used to write any of this code. Agents wrote much of this substrate, under the founder's direction, and the commit history says so. The claim is about the serving path, not about authorship.
• It does not claim model weights are absent from every machine the founder owns. It claims that no cell serving this board runs a model process, opens a model port, or holds an unmasked model unit, and that no code in the substrate builds a request to a model API host.
• It does not claim a model is a bad instrument. It claims a verdict that changes when you run it again is not a verdict, and that every claim on this board must be re-derivable by a stranger from public bytes. A sampled answer cannot meet that bar; an exact one can. That reading is ARGUMENT; the rows above are the measurement.
• It does not claim the removal was free of risk. It was validated unused on the evidence in the table above, and that evidence can no longer be re-read.

Reproduce

Two of these need the fleet's ssh keys and so are the founder's to run; the rest anyone can run against a clone of the substrate repository.

# anyone: the source-side gate, and its control arms
bash no-llm-upstream-in-the-stack.sh
bash no-llm-upstream-in-the-stack.sh --self-test

# anyone: the fleet gate's control arms fire on fixtures, not on cells
bash cells-are-cattle-not-pets.sh --self-test

# the founder, or anyone holding the cell keys: the live per-cell sweeps
bash no-llm-upstream-in-the-stack.sh --fleet
bash cells-are-cattle-not-pets.sh --verbose

And the thing the removal was in service of, which needs nothing but a terminal — the court answering in public, rendering what it can project and naming what it cannot:

curl -s -X POST https://affine.earth/language-invariant/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"affine_translate_text",
       "arguments":{"source_lang":"en","target_lang":"es","texts":["book","water"]}}}'

Measured 2026-09-12: book returns libro under chart:1, and water is withheld as REFUSED_SHARED:1 of 1:water→agua(pt) — the court names the third language whose spelling it collides with rather than guessing which language it is looking at.