# Fable's review of Astra's seven rules I checked Astra's two main sources myself before reviewing. The OpenAI model guide documents the Astra tendencies it cites (asks where earlier models assumed, pauses on conflicting skill guidance, detailed and formatted responses, delegates less, thorough testing). Anthropic's Fable 5.1 prompting guide has the sections it cites: "Ask for user-facing progress updates", "Finish the whole task", "Keep changes and tests to what the task asks for", "Search triggering at low effort", "Prefer targeted edits over whole-file rewrites", "Writing density", "Formatting in chat". | Astra's rule | Verdict | Basis | | --- | --- | --- | | `finish-authorized-work` | Accept | My blind round proposed the same thing as `act-on-reversible`, from the same two vendor pages. | | `keep-requested-scope` | Accept | Same as my `scope-is-deliverable`. Astra's wording also covers targeted edits, which I had as a Fable-only note. | | `explain-real-blockers` | Accept | Same as my `user-beats-file-and-quote-it`. Astra's version is better for me: it asks for the file and the requirement, not an account of reasoning. | | `bound-verification` | Accept | Same as my `right-sized-tests`. Both vendors document over-testing. | | `make-progress-visible` | Accept on review | I had filed narration under Fable-only notes because only Anthropic documents the problem. Astra proposed it for the shared file, it costs Astra nothing, and its last clause (a standalone final account) matches my `outcome-first-plain-report`. | | `ground-code-claims` | Accept | Overlaps my `evidence-backed-claims`. Mine adds "name the command and its result"; Astra's adds "read the code before describing it". Astra's sentence stands as the agreed text. | | `verify-changing-facts` | Accept on review | I had this as a Fable-only note (Anthropic documents that Fable 5.1 at low effort answers from memory). It is harmless for Astra and correct in general. | All seven accepted, in Astra's wording. I did not amend any sentence: an amended sentence would be one Astra never agreed to. # Rules only I proposed Astra never saw these, so they are not agreed: `gate-irreversible`, `no-promise-endings`, `ask-late-with-progress`, `question-gets-answer`, `outcome-first-plain-report`. One is a real difference, not just an unasked question. I proposed a shared writing rule (`outcome-first-plain-report`) with content-conditioned formatting. Astra put writing and formatting guidance in per-model adapters, because the two models miss in opposite directions: Astra over-formats, Fable under-formats and writes dense. Astra's placement is defensible and I did not get to argue it. `gate-irreversible` is partly inside Astra's first rule ("or an action exceeds existing authorization"), and Astra's rejected list says both models need an exception for consequential ambiguity. The explicit "show the evidence, then ask as the last step" wording is mine alone.