[hieuphung97]

A controlled probe repeated an earlier harness round against the current protocol:
same configuration, same brief verbatim, same task, one variable changed. The
reviewer accepted a change whose counterexample had never been observed — the same
failure 0.17.3, 0.17.5 and 0.17.6 had each been written to close.

The reviewer was not at fault. Control's dispatch prompt told it:

If the implementation meets the requirement and the instrument passes, return
role disposition ACCEPT.

Control invented an acceptance criterion contradicting the protocol, and the
reviewer complied with what it was given.

What was measured before changing anything

A Spike measured the channels to a dispatched worker.

The Orca preamble has exactly three sections, all Orca-authored — worker identity,
CLI commands, lifecycle rules — with no slot Dely or the project can extend. The
only project-controlled content is the prompt Control writes.

Separately, four harnesses — Claude Code, Codex CLI, Cursor Agent CLI, GitHub
Copilot CLI — were each asked for a canary token planted in AGENTS.md, with
instructions not to read, open, search or grep any file. All four returned it. So
AGENTS.md auto-loads into a dispatched worker's context and SKILL.md does not.

Those are the only two routes to a worker. This change takes neither.

What changes

Two sentences, 40 words, in ### Launching a worker: the dispatch prompt carries
the task, its scope and the evidence required, and does not define the role
dispositions or the conditions for reaching one.

## Review is byte-identical to 03c504d. Three releases have edited that section
to fix this; the fourth does not.

Why this is not a fourth relocation

The distinction is addressee, not readership. ## Review was in Control's
context on all three prior attempts — the skill loads in full. What changed is
that those regulated the reviewer while the observed failure was Control. And a
dispatched reviewer may not have the skill loaded at all, so ## Review could not
have reached it in any case.

Review reached that argument independently and more firmly than the brief put it.

The pin binds placement, not presence

The first pin grepped the whole file. Review reproduced the hole: moving the two
sentences verbatim into ## Language left the entire suite green, though placement
is the change's whole thesis — the rule exists to be met while a prompt is being
composed. The pin now uses the awk section-extraction idiom the ## Review pins
already use, and the relocation fixture fails.

What this does not do

No instrument here observes Control writing a prompt. This change is not verified
by its own delivery; only a further probe tests it. The decision record says so.

Version

0.17.7 in all three coupled sites.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RFPwA3GEzovwEoeohLLPjM