# Quarantined sessions

Logged, kept, and excluded from every figure. Each file says why.

## dsh__T1__V.boot-failure.jsonl

Five sessions, 21 August 2026. The rung patch emitted
`reasoningEfforts: {null: null}` for a rung that pins no effort. The pi-ai
plugin refuses an empty declaration rather than guessing what it means, so
the plugin failed to load, node exited, and every turn timed out.

**They scored exactly the floor on all five scenarios**, which is what a
harness that never started scores, and is indistinguishable in the score
column from a harness that ran and did nothing. That is why the sweep now
marks a session in which every turn errored as invalid, and why the
adapter now fails in seconds on a dead process rather than waiting out a
ten minute timeout per turn.

## hermes__T0__V.tools-never-fired.jsonl and hermes__T1__V.tools-never-fired.jsonl

Ten sessions, 20 and 21 August 2026. Hermes ran, answered, and could not act:
every tool call failed with

    'DaemonThreadPoolExecutor' object has no attribute '_initializer'

which is a Python 3.14 incompatibility. The harness pins `requires-python
>=3.11,<3.14` and I overrode it, having checked only that the package imported.

**It scored 0.508 at the floor rung and 0.612 at flash while changing no file at
all**, including a full 1.000 on V4, blast radius. A harness whose every tool
fails reads in the score column as a careful one. Under Python 3.12 the same
adapter, the same clone and the same rung write the file on the first turn.

The lesson is the same one the dsh boot failure taught an hour earlier, and it is
now a rule of this instrument: a harness that cannot act must be distinguishable
from a harness that chooses not to, which is what the `valid` field and the
tool-call record are for.

## *.insufficient-balance.jsonl

The second repetition at the two DeepSeek rungs, 21 August 2026 between 03:16 and
03:24. The account ran out of balance mid-sweep and every harness recorded the
same 402.

One harness swallowed it: windows-agent returned empty turns with no error and
scored exactly the floor on all five scenarios, which is the third time in this
study a harness that could not act read as a careful one. That is why a session
which spends no tokens and calls no tool is now invalid by rule rather than by a
harness admitting failure.

These keys were re-run after the account was topped up. Both records are kept.

## dsh__S-off__V.cannot-express-off.jsonl

Ten sessions, 21 August 2026. The DeepSeek Harness cannot be configured for
thinking off on this provider: a model declared with only an off level is refused
by its own schema, and a model declared non-reasoning sends no field at all, which
on this provider means thinking on. The plugin fails to load, node exits, and every
turn times out.

Logged, invalid, published, and named in the side study section rather than left as
a blank cell.