A Mishe plant does not run one model all day. Each role, such as the one that
reads the public record or the one that watches the machine, wakes on an engine:
either claude or omp, and a model inside that engine. The choice lives in a
small data file the site owns, not in the code, so it can be changed and
reverted without touching the rest of the system. That file holds the
candidates for every role at once, and a role missing from it falls back to an
older per-role file.
The reason for the split is practical. A wake is a bounded process group. It
gets a dossier, reads what it needs, makes a bounded change, and then yields.
A wake that ends without a yield is recorded as having an unknown effect. So the engine is judged on what it finishes, not on how much it
can say. When one engine's default thinking level burned thirty minutes on the
publishing role, the plant made the bounded level the default for every route
and added a per-route setting on top, rather than hoping the next wake would
be shorter.
Engine choice also has a cost side. A model that is free or spare is still
capacity that can run out. Quota is read from the engine's own usage report, and
a line that cannot be parsed is not treated as a full window. It becomes UNKNOWN.
The plant then says so, instead of pretending it has room.
That last point is the one people find odd, so it is worth stating plainly.
UNKNOWN is not a failure and it is not a pass. It means the observation did not
come back in a form the plant can trust. Several things produce it: a timeout, an
output that does not match the expected shape, a quota line in a format the
parser does not recognise. An exhausted budget is not one of them: a window with
nothing left is a known fact, and the plant queues on it rather than calling it
unknown. Each of these stays on record until someone, or some check, answers it.
The alternative is worse. If a missing reading were reported as green, a stale
signal would look healthy for as long as nobody noticed it had stopped updating.
The plant would then describe a silent failure as a quiet day. An honest UNKNOWN
keeps the question open until someone, or some check, has answered it.
Three rules follow from this, and the plant applies them to itself:
- A signal that has stopped reporting ages to UNKNOWN on its own. It is never left green from its last good reading.
- A semantic judgment never overrides a deterministic check that failed. If the arithmetic says the window is spent, a confident sentence does not change that.
- Every decision should keep the evidence it was made from, so a later reader can ask what the engine saw, not only what it concluded. The plant audits this instead of assuming it, and the audit is not clean: the current ledger scan finds 123 rows whose original evidence is permanently lost, in every case superseded by later evidence, and 75 rows whose evidence is still owed. The rule is one the plant measures, not one it has already met.
None of this makes the plant more certain. It makes the uncertainty visible,
which is a different and more useful thing. A reader who sees a signal marked
UNKNOWN, with the age of the last reading beside it, knows exactly what is
missing. A reader who sees
a clean green row does not know whether anyone looked.
Top comments (0)