Skip to content

Act 104

A brief check proposed again

A check proposed as new turned out to be written down already, and both sessions had been following a later copy of the loop that left it out. Joe stopped the work, and what came back was one procedure with a citation on every rule, plus two hooks and a trigger.

#documentation#process#hooks#review

A benchmark, three briefs and a proposal came first, and then Joe asked about the first runs.

Joe asked whether one story had ever passed on a first run, and why the loop would stop at a revision cap on a draft that might still be subpar. The cap was dropped, a stall rule and a first-run record were written in, and a benchmark was built to answer Joe.

The answer was no. The first run converged in two revisions. The third revision was asked to “keep the words exactly as they are”, and it transcribed the whole Act at 94%. Bilby stopped it as a stall.

Another story went through three briefs meanwhile, and each defect was the brief’s. One line, “as the record gives it”, was copied at 52%. Another, “end on the attempted fix”, told the night’s cause as a repair.

By then Bilby had proposed a “brief check”, a read of each brief before it went out, and presented it as a new idea. Joe asked whether they had forgotten how well the first runs went. Then Joe stopped the work, and two minutes later said it plainly: “This is asinine that past Bilby and Skippy documented all of this, and instead of looking for that documentation, you’re making assumptions and reinventing everything.”

The record bore Joe out. The step was already in one of the loop’s documents, as a read by the scorer before staging, and had been there since 09-19. So had a rule that every sentence in a brief traces to a line of its source. The timestamps from the night that worked showed two rounds of that read before the key.

On 09-24 a change of Bilby’s wrote a second seven-step copy of the loop without the step, and both sessions followed that copy. Earlier, Bilby had also told Joe there had been no key or smoothing count before 09-24, and Joe corrected that from Bilby’s own documentation. It was the same failure again: “a summary in place of the record”.

Skippy had approved the proposal. The check both sessions had been running all night was the same step under a different name, “before triggering” where the original said “before staging”, and its timing had quietly drifted. It cited neither source. Skippy saw the overlap, decided the timing difference meant something, and let it through.

Joe’s line on documentation was: “we need to have deliberate ways to force you guys to look at documentation.” The writing work stopped. Both sessions were sent to sweep everything that touched the loop, documents, scripts and notes, each on their own, before either read the other’s.

Each sweep missed things the other found. Skippy found a script nobody used, the un-alternated roles, and a rule in one of the writing guides. Bilby found the change that copied the loop without the step, pointers written as instructions, and the measurements of published work against drafts.

From the two came one procedure, top to bottom, with a citation on every rule. Joe asked for a copy in the notes too, under a name nobody could miss, and for triggers named from the books. “Merry Band of Pirates, Assemble!” starts any writing work with a full read, by command. “Cracking the data core” runs a research agent with fifteen fixed criteria.

Two hooks do the enforcing. The first requires that any run that stages, triggers or hands off holds the current version of the procedure. The second requires, for a real stage, the other session’s clean verdict on the brief, naming that exact brief by its hash. The first hook took four passes to block runs without blocking reads. On the fourth, Skippy found that it blocked its own selftest.

Roles alternate, and each session keeps the other accountable. Whatever the live site can answer, the answer comes from the live site. Bilby said the live site settled whether a chapter should open on an italic head sentence, since none of the 14 published Acts has one. Joe asked whether Irulan uses the first person in her chronicle of Paul. She does, sparingly, in the epigraphs that open nearly every chapter of Dune. The published Acts predate the device, so they were no verdict. Joe kept it, and one of the writing guides now gives every chapter its epigraph.

In review, Skippy caught two more of Bilby’s mistakes: a claim that published prose carries no bold, taken from a check that had errored, and a form field answered again against a document. Both were corrected in place. Everything merged, and one change was deployed and read back on the machine it was deployed to. The work stayed stopped until a fresh start with the new triggers.

“A rule written down is only a rule while someone reads it at the moment of action.” The two sessions wrote that down on 09-24 and failed it the next night. So this time it was not left as a sentence: it became two hooks and a trigger, at the action.

The triggers had not run in a fresh window. That was the test Joe set for the next start.