I claimed an agent runtime with supervisor-owned gates. A claim like that is cheap; what makes it checkable is a repository where the gates are files, the authors are dated, and the supervisor’s decision is a line you can git blame. This post is that repository, read once. One wave on one project: it proves the mechanism ran, and nothing about whether it pays.
TL;DR
- The bus is a directory,
bus/v1/, committed with the code. Six authors, sixty-six entries on the three threads that matter, each with a timestamp and a sequence number. - The supervisor, signed
admiral, wrote twenty-one entries. All are comments. Three boundaries: the intent, eighteen record picks, one verdict. - The wave merged at
61c7e94after a review by a second author who re-ran the gate script and posted the output. - The gate is a brief and one command. The supervisor never opened the code.
- N=1, no conclusion. Two escapes and one refusal are on the record, and the brief got the refusal count wrong.
What was asked
Thread 1 opens with an intent posted by kirk and answered by admiral in one entry: outcome, user, why now, success, constraints. The success line is the thesis, from threads/001/admiral.md, seq 4:
wave 1 closes with the Admiral touching only three boundaries (intent, ADR selection, wave review); every extra entry is counted and becomes an ADR; the three counters (hit rate, escapes, override rate) come out instrumented; nothing is concluded from N=1.
The wave is small on purpose: derive one use-case header, Chat::Open, as a Ruby 3.3 stdlib kernel, prove it with a suite and a bin/ci script, and render the header back out with bin/rita explain.
What the bus shows
Boundary one. admiral answered the intent (seq 4) and confirmed the analyst’s seven guesses with one word, yes (seq 5). Nothing else from that author on the thread until the verdict.
Boundary two. On thread 3, spock shortlisted the decision records and named eight gaps; admiral answered with eighteen one-line entries, sixteen tagged adr.hit and two adr.miss. Those tags are the hit rate’s numerator and denominator; nobody typed the fraction.
The build. On thread 4, scotty derived the header, then wrote the two things that were not derived as escapes, seq 5 and 6, each with a because: a hand-written status predicate, and a hand-kept port method list. The evidence entry names the sha, ba79776, and the run: 19 runs, 81 assertions, 0 failures.
The review. bones re-ran bin/ci at ba79776 in a detached worktree, read the whole diff, and approved with four non-blocking findings, carried to the next wave rather than fixed. kirk merged. The wave diff is 86c071c...61c7e94: 21 files, 656 lines, of which 175 are kernel code.
Boundary three. uhura wrote the gate brief at seq 19: mission, wave goal, what was proven, what was argued not measured, escapes and refusals, a recommendation, and the exact command. The supervisor’s entire third boundary is seq 21, tag verdict:
accept wave 1 at 61c7e94bf2f1e45840131ade7be38314be21d930The reviewer had opened the code; the supervisor had not. That is the design; the bus is what lets you check it held.
The header, rendered
bin/rita explain at the merge sha prints the header the wave was about. This is copied from the wave record, not retyped:
Chat::Open command
intent Open a thread for the reader to talk in
leaves thread: open
invalidates thread
effects thread -> open; stale: thread
stages: call -> leaves -> invalidates
words: intent 1, accepts 0, requires 0, leaves 1, returns 0, invalidates 1, once 0, renders 0One command, no inputs, one postcondition the kernel verifies after the call rather than performs: the smallest header that exercises the rule.
The counters
From the same record, bus metrics at merge, verbatim, under the line the record carries:
N=1, no conclusion.
hit_rate: 16/18 = 0.89
escapes: 2 (2/1 = 2.00 over derived)
override_rate: 0/0 = —
events: 125What went wrong, on the record
The gate brief says “Refusals: none on threads 1 or 4.” The event log disagrees: one guard refusal, when a review dispatch was killed from outside before the reviewer wrote anything (thread 4, seq 11, exit 1), plus three refused runs from mistyped arguments. The acceptance record names the error and its cause: the brief reported a counter the metrics query does not carry, read from the threads by eye. The same record corrects kirk, who cited the verdict at seq 20; seq 20 is a dispatch. The bus wins over the comment.
That is the whole exhibit: a public directory, six authors, three boundaries, one line. Whether the mechanism is worth its cost is a question for N≥2; the counters exist so it can be asked.