[gHashTag]

L0 says every rule below the interface is written once in .t27 and generated, because a rule transcribed into Swift, Rust, Zig and Verilog is four rules that agree until someone edits one. The t27-backend skill then records where that stands, and it is blunt: T27-00 "generates but does not run anywhere, so strictly the plan is blocked at its first ring."

Two-thirds of that was wrong.

question	answer	how it was checked
does the spec generate?	yes, 7 of 7 rules, 121 lines of Rust	t27c gen-rust rings/T27-00/queen_core.t27
does the generated Rust compile?	yes, zero errors, warnings only	rustc --crate-type lib --edition 2021
does it agree with the twin production runs?	yes, 460 of 460	tri t27-parity

Generation was never the blocker, and compilation had never been tried. What was actually missing is the thing the law is for: any evidence the generated ring says the same thing as the Swift binary the container actually asks.

The parity harness

Two sides, one question:

• the ring — rings/T27-00/generated/queen_core.rs, generated by t27c, compiled by bare rustc, driven through rings/T27-00/parity/probe.rs. That probe is stdin, stdout and a match on a verb; if a rule ever appears in it, it has become the fifth copy of the law and the exercise has inverted.
• the twin — queend, the Swift binary the container asks, taking a JSON question on stdin. Not a re-implementation written for this harness.

Exhaustive, not sampled. Ring 00 is integers and booleans with small ranges, so the whole meaningful space fits: every attempt sequence up to depth 3 for retry (40 cases), and the full review grid with total≤4, judged≤total, unmet≤judged, files≤2, prior≤3 (420 cases). A sampled parity check has a coverage number that has to be defended; an exhaustive one over a bounded space does not.

Shown failing before it was trusted. On a scratch copy that never touched a shipped file, MAX_SEND_BACKS was changed from 2 to 1 in the spec:

disagreements: 30 of 420
first: total=1 judged=1 unmet=1 files=0 prior=1: ring says escalate, twin says sendBack

A parity check nobody has watched catch a divergence might be comparing nothing.

It runs in the heal chain every round, so "they agree" cannot decay into "they agreed when somebody last looked". A missing side is reported as missing and exits non-zero — the one answer this must never give is a green tick for a comparison it did not make.

Five copies of one constant, and two of them were the loop's own

rings/T27-00/queen_core.t27:37                          the spec
rings/SR-00/QueenDelegation.swift:470                   Swift copy 1
agent-server/queen-core/.../QueenDelegation.swift:470   Swift copy 2 (the fork)
.trinity/loop/dash.mjs                                  'bees running (of 4)'
.trinity/loop/snapshot.mjs                              the same string again

The last two are prose inside a dashboard label — which is exactly how a constant escapes every search anyone would run for it. A loop that reports on the swarm's discipline while holding its own copy of the swarm's constant has no standing to.

Both now call ringConst('MAX_CONCURRENT_WORKERS'), reading one line out of the generated artifact. With no artifact the label reads (of ?), never (of 4) — an invented constant that happens to be right is the same defect wearing a correct answer, and it would survive the spec changing. Calibration asserts the positive and both negatives, and a second case fails if any file in the loop writes (of 4) into a string again.

Also measured: there is no JavaScript backend

t27c --help has four language targets — Zig, Verilog, C, Rust. Everything else is silicon tooling. So a .t27 source has no path to a Node module today, and adding one means editing t27/bootstrap, which another agent owns under L0b.

What that does not excuse is a rule living in .mjs. Plumbing is plumbing; a threshold, a verdict or a capacity is law, and law goes in the ring.

Next, in ring order

The ring is proven, not yet running — queend is still the artifact the container executes. So: make the container run the generated ring for the four decisions both implement; then delete a transcription (the two Swift copies are audit finding D1, and this harness is what makes removing one safe rather than brave); only then T27-02.

tri t27-gen, tri t27-parity. 225 calibration cases. The compiler builds into trios/.trinity/t27c-build, so nothing is written into ~/t27.

🤖 Generated with Claude Code