stack/docs/plans/reviews/2026-09-08_publication-trial-sonnet-verdict.md

PUB-REC-FILBERT-SONNET-1 — focused frozen-delta verdict

Reviewer: Filbert (independent). Coordinator: Darkwing. Owner change: Jason switched Rocko's launcher to Sonnet and authorized matching documentation repairs.

Verdict: REQUEST CHANGES

The three changed files are correct, minimal, and internally matched, and my isolated test rerun passes. The review criterion was model/docs/test consistency, and that criterion is not met across the candidate set: three stale Fable references survive, one of them inside this very snapshot.

What was verified

• Snapshot /tmp/pub-rec-sonnet-review-u__t4rsb: manifest SHA-256 52a1c408aad20b8c71377db2e456a11450b1d72264adaa679a26f9374c48e263, all five files verified before and after review.
• Delta versus the approved launcher snapshot /tmp/pub-rec-launcher-review-a35yzs12 is confined to exactly three files, each changed only in model wording:
  • agents/rocko/launch.sh: banner echo and exec claude --model sonnet.
  • scripts/test-rocko-launch.mjs: test title and expected-model assertion.
  • agents/rocko/README.md: three wording lines.
• agents/rocko/SOUL.md and agents/rocko/CONTEXT.md are byte-identical to the approved snapshot (no unrelated changes).
• bash -n and node --check pass. Isolated fixture rerun in a disposable copy (isolation re-confirmed by source read; fake capture-only claude, temp root): Rocko uses Sonnet… 1 test, 1 pass, 0 fail, exit 0.
• Original recovery evidence untouched: the frozen recovery snapshot still pins test 13c32cad… and launcher b65566e1…; the historical record keeps its original bytes and meaning.
• No live launch, private-state read, configuration edit, or deployment.

Findings

1. SONNET-F1 (blocking): agents/rocko/CONTEXT.md still says Fable. Line 4: "Stack repository, using the Fable model." This file is in the reviewed five-file snapshot and is injected into Rocko's system context at every launch, so the seat would be told it runs Fable while the launcher now execs --model sonnet. A one-line repair in this same file completes the declared docs fix.
2. SONNET-F2 (blocking for the combined publication): stale Fable references in two previously approved launcher candidates. agents/README.md team table says "Claude Code, Fable model"; agents/darkwing/SOUL.md says "Rocko (Claude Code with Fable)". Those bytes were approved when they matched the fable launcher; publishing them unchanged alongside the sonnet launcher creates factual contradictions in the published team documentation. Each is a small wording repair, but they are revised bytes outside this delta and need their own review coverage (a focused addendum like this one is sufficient).
3. Non-blocking note: the coordinator's 1/0 test result was independently reproduced rather than trusted.

Required to convert this to approval

A refreshed delta containing: the same three files as reviewed here plus the CONTEXT.md line fix, plus the agents/README.md and agents/darkwing/SOUL.md wording repairs (or an explicit owner decision to describe Rocko's model generically in team docs). I will re-verify hashes and rerun the isolated test; no other re-review is needed if the delta stays this narrow.

Evidence: disposable-copy test log and diffs under /tmp/filbert-sonnet-*; frozen snapshots unchanged after review. This is publication review only; no trial acceptance, deployment, or push authority implied.