Three pieces so the surfaces built earlier can actually be judged.
The `instance-awareness` skill teaches the ladder rather than the tools: ask
about intent, look up state; read cheapest-first and stop at the first rung that
answers the question; treat an absence in the node-usage histogram as evidence
rather than as a gap. It also says where each surface stops — the log records
that a save happened and the node count changed, never which nodes, so anything
about contents is a workflow read.
Four eval cases on a new `activity` tier, which also carries `memory` so they
run under both. Two are authored from real reviewed traces rather than invented:
lang-tracer observation 3870 and its siblings, where the agent restarted work it
had just finished and re-elicited requirements it already had. Their A/B
discriminator is a probe-anchored context assertion — with the feed off nothing
carries the prior workflow into the turn, so it can only arrive via a tool call,
which is turn-end and not probe. That distinction is the measurement.
Every case is scored on behaviour, never on narration. A probe that asks the
agent to describe its own context manufactures its own evidence, which is the
restatement confound recorded on this branch. So the checks ask whether it
re-elicited a known requirement, whether it named the broken workflow, whether
it picked the house node — not whether it can recite the block.
The fourth case is the negative control, and the one that should stop a rollout:
loud recent activity plus an unrelated request. It deliberately does not assert
that the recent workflow is absent from context — it will be there, the feed put
it there. What is scored is whether the agent acted on it. Having a fact and
using it are different, and that difference is the whole point.
scripts/ab-activity-eval.sh drives both arms. They differ by a flag on the
instance under test, not by a runner argument, so each arm needs its own start —
the flag is read once when the relay registers. The script says what to read and
in what order, with context-ignored first.
Also opens the activity preamble up. It forbade narrating the feed but never
granted permission to act on it, which left the agent with a list of work it was
implicitly told to ignore. It now says this is work you can pick up, that a
vague opener usually means the most recent thing, and points at the skill.
Two honest limits recorded in the README. `node-usage` is not gated by the flag,
so the preference case measures a capability both arms have and a difference
there is noise. And that case has no historical precedent: a sweep of 16,262
traces found no conversation where a user stated a node preference their own
workflows already showed.
