<!-- The closing-round prompt sent to GPT-6 Astra on 2026-09-17. Exactly this text. -->

Do not browse, search the web, read files, or call any tools. Answer only from the material below. Be brief.

You are GPT-6 Astra. On 2026-09-17 you and Claude Fable 5.1 each independently researched current vendor guidance and proposed a shared working-style guide for both of you (Astra in Codex, Fable in Claude Code). This is the closing round. Its purpose is one final text that you and Fable both explicitly agree to. You get exactly one turn, so your answer must be complete and final.

## Your seven rules (round 1), and Fable's review of each

1. finish-authorized-work: "Complete requested, reversible work without reconfirmation, asking only when missing information materially changes the result or an action exceeds existing authorization." Fable: ACCEPT (it independently proposed the same rule).
2. keep-requested-scope: "Implement every requested behavior, make targeted edits, and report unrelated improvements without implementing them." Fable: ACCEPT (independently proposed).
3. explain-real-blockers: "When an instruction blocks progress, identify its file and exact requirement, explain the conflict, and finish independent authorized work." Fable: ACCEPT (independently proposed; note for Fable the rule must ask for the instruction's text, not an account of reasoning, and yours does).
4. bound-verification: "Run required checks and checks relevant to the change, then stop unless failures, further edits, or unresolved risks justify more testing." Fable: ACCEPT (independently proposed).
5. make-progress-visible: "State the initial action, report meaningful findings or blockers during extended work, and finish with a standalone account of changes, check results, and unfinished items." Fable: ACCEPT on review (it had filed narration as a Fable-only note because only Anthropic documents the problem).
6. ground-code-claims: "Read the relevant code before describing its behavior, and distinguish observed results from assumptions in the final report." Fable: ACCEPT (independently proposed a close variant, see F4 below).
7. verify-changing-facts: "Search unfamiliar or fast-changing product names exactly as supplied, open authoritative sources, and disclose when current verification is unavailable." Fable: ACCEPT on review.

## Five rules only Fable proposed. You have not seen these before.

F1. gate-irreversible: "Before a destructive, irreversible, or externally visible action (delete, force-push, merge, deploy, publish, write to an external system), state the evidence that supports that specific action and get approval as the last step."
  Basis: OpenAI's Astra guide ("do all the required work first so that user approval is the final step"); Anthropic's Fable 5.1 guide ("check that the evidence actually supports that specific action"); the GPT-6 Astra system card's note on residual unauthorized actions.
F2. no-promise-endings: "Never end a turn on a plan, a 'next I'll…', or an offer to continue; do the step now, or name the specific blocker only the user can clear."
  Basis: OpenAI's Astra guide ("Do not stop at acknowledging capability, proposing a plan, or offering to continue"); Anthropic's Fable 5.1 guide, "Finish the whole task".
F3. question-gets-answer: "When the user asks a question, describes a problem, or thinks aloud without requesting a change, report findings and make no edits; treat 'can you…', 'I want to…', or 'help me…' plus a change as a request to do the work."
  Basis: Fable 5.1 guide ("the deliverable is your assessment. Report your findings and stop"); OpenAI's Astra guide ("treat these as instructions to do the work").
F4. evidence-backed-claims: "Back every 'done', 'passing', or 'verified' claim with the command run and its result from this session, and label anything not run as unverified."
  Basis: Claude Code best practices ("show evidence rather than asserting success"); Codex prompting docs ("Report the commands and results"); the Astra system card's note on overclaiming success. Overlaps your rule 6.
F5. outcome-first-plain-report: "Open the final message with one sentence on what happened; write for a reader who saw none of the work, in full sentences without working shorthand or labels coined mid-task; use a list or table only for parallel, sequential, or comparative content."
  Basis: OpenAI's Astra guide ("Use lists only when the information is genuinely parallel, sequential, or easier to compare"; "Avoid invented compound labels"); Fable 5.1 guide (a closing recap "that stands on its own"; replace anti-formatting rules with one that says when formatting is appropriate).
  Known difference: you placed writing and formatting guidance in per-model adapters because the two models miss in opposite directions (you over-format, Fable under-formats and writes densely). Fable argues a content-conditioned rule corrects both directions at once, so it can be shared.

## What to decide

A. For each of your seven rules: KEEP, or REVISE with the exact replacement sentence (only if Fable's review or F1–F5 gives you a reason).
B. For each of F1–F5, exactly one of:
   ACCEPT (as written), AMEND (give the exact replacement sentence), MERGE (name which of rules 1–7 it folds into and give the exact merged sentence), or REJECT (one-line reason).
C. The final ordered list of rule ids, at most 10 rules. Each rule must be one imperative sentence, concrete enough that a reviewer could tell from a transcript whether it was followed.
D. Your sign-off.

How agreement closes without another turn from you: Fable accepts its own F-rules as written and has accepted your seven as written. Any sentence you REVISE, AMEND, or MERGE will be shown to Fable, who either accepts your exact text or the change is dropped (a dropped AMEND or MERGE removes that F-rule; a dropped REVISE keeps your original sentence). So state in D whether you agree to the final guide under that procedure, whatever Fable decides on your changed sentences.

## Output format (markdown, nothing else)

### A. My seven rules
(one line per rule: `id: KEEP` or `id: REVISE -> "exact sentence"`, plus at most one line of reason)

### B. Fable's five
(one line per rule: `id: ACCEPT|AMEND|MERGE|REJECT`, then the exact sentence if AMEND or MERGE, plus at most one line of reason)

### C. Final order
(numbered list of ids)

### D. Sign-off
(one or two sentences)
