feat(prose-tests): walk on Sonnet, judge on Opus, escalate a failure

[leeovery]

Summary

• Walker moves to Sonnet. Both agents went to Opus back when a walk was judged by the single message an agent returns — a compressed closing summary read as a skipped step, the walker looked unreliable, and the answer was to spend more. With the walk captured turn by turn (feat(prose-tests): judge the walk as told, not the summary returned #561), that reason is gone.
• Measured, not assumed. A Sonnet walk of implementation-picks-first-task passes every path step with quoted evidence, identical world, no markers. An Opus walk of the same case made the same calls in the same order. Across both cases and both models the code-checked layer came out identical.
• Sonnet was better on one axis: it declined a system-reminder pushing auto-mode bias against stopping, citing the prose line naming harness auto mode as exactly the thing to ignore. That's the STOP-gate discipline the pipeline depends on.
• Asserter stays on Opus — the cheaper half (one prompt in, a verdict out) and the half the result rests on.
• A FAIL is rerun once on Opus by the orchestrator. A failure is where spending more is worth it, and a defect a stronger walker also hits is a defect. Escalation happens in the moment, never by editing the definition, so a routine run is always the model the definition names.
• Both models are reported on a confirmed or flaky failure, so an escalated rerun is never read as like-for-like.

Test plan

• Prose suites green (39/39 on corpus + invariants; full prose gate unchanged)
• The Sonnet verdict this rests on: 5/5 path steps, further claim PASS, world identical: true, all three deterministic checks PASS, zero markers
• design/prose-tests.md P6 rewritten to match

🤖 Generated with Claude Code

Stack

1. docs(design): prose-tests programme design log #544
2. feat(prose-tests): the framework — cases, worlds, runner, skill #545
3. test(prose): feature happy-path corpus — five worlds, seven cases #546
4. test(prose): bugfix corpus — the investigation-centric surfaces #548
5. test: retry recursive teardown removals — kill a class of phantom failures #549
6. fix(entry-skills): close the handoff fences — six files render their arms wrong #550
7. docs: a contributing page for working on the system #551
8. fix(entry-skills): every handoff arm says to invoke the skill #552
9. fix(implementation): environment setup belongs to the setup reference alone #553
10. fix(prose-tests): the asserter is told which substitutions were armed #554
11. feat(prose-tests): the mid-flow substitution, and a world only prose can describe #555
12. test(prose): claims assert consequences, not what was displayed #556
13. feat(prose-tests): record everything the agents do, results included #557
14. fix(discussion-entry): the handoff reports the source it actually had #558
15. fix(prose-tests): the stop hook records, and names the model that walked #559
16. fix(prose-tests): command output was never actually recorded #560
17. feat(prose-tests): judge the walk as told, not the summary returned #561
18. feat(prose-tests): decide in code what an agent should not be deciding #562
19. test(prose): a case starts where a session starts #563
20. feat(prose-tests): walk on Sonnet, judge on Opus, escalate a failure #564 👈 current
21. test(prose): give the eight read-only cases something that can fail #565
22. test(prose): only walks that can be observed, and checks that survive the trip #566
23. fix(prose-tests): the verdict names only the model the record names #567
24. test(prose): discovery, walked to the point where work first exists #568
25. fix(prose-tests): the asserter judges which of prose or walker was at fault #569
26. docs(conventions): a step whose reference routes every exit still signposts #570
27. test(prose): discovery's epic arm, to the same durability boundary #571