Everything I could decide myself is decided. What's left needs
you. Each one has a recommendation from Fable — if you agree, pick it and
press Done. "I'm not sure" is a real answer and becomes work for me.
You don't need to open the session — pressing Done sends your answer back and it picks up.
1PR #5784 builds the corpus re-run that is supposed to prove a grading-rubric change didn't make Clara's grader worse. While it sat 8 days, the grader underneath it changed from a plain rule into a rule fused with an AI judge — so the PR's re-grade leg no longer holds. Which repair should it take?
Fable recommends: R2 — keep the re-grade, running the AI judge again over each graded conversation on the subscription lane (not a metered key), and drop any conversation whose judge call failed rather than scoring it clean. Costs nothing today (there are 0 human verdicts in prod). Price: the re-grade is no longer exactly repeatable, so the PR's claim softens from 'gates a change exactly' to 'gates it within a confidence interval' — the interval comparison is already built. Both independent Fable passes picked R2. The ADR clause this PR exists to satisfy requires a corpus re-run that can gate a change; R1 ships a gate that cannot gate, which is the exact laundering the module's own header refuses, and it re-opens a review finding that already rejected softening the claim instead of building the leg. 'Exactly' died when the grader became judge-fused, whichever repair ships — R2 is just the honest form of the same measurement. It is prepared, green locally, and touches no customer, money or send path. Tried first, unsuccessfully: fb123e132 — both passes agreed on R2. It escalated anyway on a keyword rule that fires on 'rate'; here 'rate' means AGREEMENT rate and MISS rate, not a rent or a price. There is no money and no commitment to a person in this question. Flagging that so the escalation isn't misread as a money flag.
81254f64 is parked on this.
2I merged PR #5785 past an explicit hold. Its description said 'merge is a human rung on this effort. Do not merge.' — the same words are on #5784 and #5786, and the parent task says every phase is a human merge. I read the diff but not that line. It is merged, green, and inert. Do you want it reverted, or left standing?
Fable recommends: Let #5785 stand. It is reviewed 🟢, was green on every check, is admin-gated and fail-closed, and provably changes nothing a resident sees — promotion is still unreachable because nothing advances the corpus counter. Gera merges #5786 himself when he is ready. No revert. Nothing a resident touches changed, and the code is admin-gated, fail-closed and unreachable: nothing in src/, agents/ or scripts/ writes the corpus counter its promotion route requires, so no principle can be promoted at all. A revert buys back the rule but costs a revert PR plus a re-merge of code that already passed review. That said, this is your effort and your rule — if the point is that the hold means what it says, revert is the right answer and I will do it. I am not merging #5786, and I will not merge anything else on this arc. Tried first, unsuccessfully: f76799d50 — RESOLVED, route to the human. Asked before I knew about the hold; Fable found it in the PR body and made it decisive, along with 'grant ambiguity at the irreversible rung resolves fail-closed'. It also flagged that CI was not green on the exact head at that moment. I had not read the body.
81254f64 is parked on this.
3Phase 3e (PR #5786) is green and ready. Merge it, or keep holding?
In plain terms. This is the change that lets a reviewed writing rule actually reach Clara. Today it does nothing at all - no rule has been approved yet, so Clara's instructions come out exactly as they do now. It is safe to turn on early; the effect only appears when someone approves a rule.
Fable recommends: Merge #5786 now It is fully green on a healthy base, changes nothing a resident sees, and renders an empty block until a principle is promoted with a matching eval receipt. Tried first, unsuccessfully: Drove it to green myself: resolved a merge conflict main created, waited out three separate main breakages, updated onto a green base, and confirmed the 12-shard suite genuinely ran rather than trusting a check that reported success having tested nothing. I did not merge because the PR body and the parent task both say merge is a human rung on this effort.