[gHashTag]

Two findings of one shape, in the same afternoon.

---

1. The toolchain block named a binary that has never existed

which t27c finds nothing on this machine. A search of the whole filesystem on 2026-09-06 returned exactly one binary — the one built into .trinity/t27c-build. The path the skill documented, ~/t27/target/release/t27c, has never existed.

Worse, the command above it was cd /Users/playra/t27/bootstrap && cargo build --release, which writes into the tree L0b says another agent owns.

The Makefile had already learned this and wrote it down at Makefile:913: the ring gate looked for that same path, found nothing, printed [SKIP], exited 0, and make check counted it as passed — "both rings were unchecked the whole time."

The instruction lived in two places, one right, and the wrong one was the one a reader meets first.

A calibration case now fails if a runnable build command carries no CARGO_TARGET_DIR, or if the t27c it names is not on disk. Fenced blocks only — the first version failed on its own explanation, the paragraph quoting the old wrong command in order to explain why it was wrong. Fourth time an assertion in this suite has been written against prose.

---

2. Seventeen Swift files exist twice, and the audit had the story backwards

rings/SR-00 and agent-server/queen-core/Sources both hold seventeen Swift files. The self-audit has carried this for weeks as finding D1 — "a byte-identical fork of 16 SR-00 files, one diverged" — which reads like a tidying job.

Measured, and it is not one:

17 file(s) exist in both trios/rings/SR-00 and trios/agent-server/queen-core/Sources
  16 identical, 0 forked in history, 1 edited on one side only, 0 unreadable
  ~  QueenLocalisation.swift is identical at HEAD and differs in the working tree

At HEAD all seventeen are byte-identical — including the file the audit calls diverged. The 45-line difference is entirely in the working tree: one copy is being edited right now and the other has not been touched. Nothing anywhere would have said so, and the moment it lands the tree has two files that disagree.

That is exactly the state L0 is written about — "they agree until someone edits one" — caught while it is happening rather than afterwards. So the two are different answers, and the distinction is the tool:

LANDED	the copies differ at HEAD. Already in history; every build since has compiled two different rules.
in flight	they differ only on disk. Not a defect — an edit nobody has mirrored, and saying so before it lands is worth more than finding it after.

Three rules it keeps:

• Match by basename. SR-00 is flat; queen-core is split into QueenCore/QueenPolicy/queend. A path match finds nothing and reports a clean bill of health — the failure this directory keeps running into.
• Unreadable is not identical. A file git cannot show at HEAD is unknown.
• It never edits. The divergence it found belongs to another agent's uncommitted work; mirroring it would be taking a decision that is not this loop's to take, on a file it does not own.

Runs in the heal chain every round. tri forked-files.

228 calibration cases, each shown failing against the state it is meant to catch.

🤖 Generated with Claude Code